Pedestrian Behaviour Monitoring: Methods and ... - Semantic Scholar

3 downloads 3909 Views 1MB Size Report
and tourism research, social sciences, artificial intelligence, geoinformation and .... Millonig and Gartner [30] employ a combination of shadowing (see Section 2.5) ... Figure 1 (a) shows a realistic view on the GSM/UMTS cell network structure, ...
Pedestrian Behaviour Monitoring: Methods and Experiences b ¨ Alexandra MILLONIG a,b,1 , Norbert BRANDLE , Markus RAY b , Dietmar BAUER b , and Stefan VAN DER SPEK c a Department of Geoinformation and Cartography, Vienna University of Technology, Austria b Dynamic Transportation Systems, arsenal research, Austria c Faculty of Architecture, Urbanism and Building Sciences, Delft University of Technology, The Netherlands

Abstract. The investigation of pedestrian spatio-temporal behaviour is of particular interest in many different research fields. Disciplines like travel behaviour research and tourism research, social sciences, artificial intelligence, geoinformation and many others have approached this subject from different perspectives. Depending on the particular research questions, various methods of data collection and analysis have been developed and applied in order to gain insight into specific aspects of human motion behaviour and the determinants influencing spatial activities. In this contribution, we provide a general overview about most commonly used methods for monitoring and analysing human spatio-temporal behaviour. After discussing frequently used empirical methods of data collection and emphasising related advantages and limitations, we present seven case studies concerning the collection and analysis of human motion behaviour following different purposes. Keywords. Spatio-temporal behaviour, pedestrian monitoring, dataset generation, data analysis

1. Introduction Modern societies are characterised by a clear tendency towards individualism and independence. Strategies for promoting independence and quality of life for people of every age are especially important in the field of mobility. Mobility allows people to perform essential functions, including engaging in social and recreational activities when desired and reaching business and social services when needed. Especially individuals who are restricted in physical mobility due to physical-neuromuscular handicaps or handicaps caused by limited or missing sensory perception need special support. Pedestrians without physical constraints can also benefit from navigational and environmental information services when walking through unfamiliar environments. In this respect applied research has produced a number of emerging technologies and technological services such as navigation aids implemented on mobile devices that respect individual needs, in order 1 Corresponding Author: Alexandra Millonig, Department of Geoinformation and Cartography, Vienna University of Technology, Gusshausstr.20, 1040 Vienna, Austria; E-mail: [email protected].

to support self-determined mobility for completing basic daily tasks without personal assistance. Advances in this field are strongly dependent on broad knowledge about people’s behaviour with regard to motion, decisive decision processes and related influencing factors. Efficient assistance and technological services can only be developed on the basis of comprehensive investigation of spatio-temporal behaviour and underlying determinants. Researchers of different disciplines, e.g. sociology, tourism and travel behaviour research, artificial intelligence, or ubiquitous geotechnology and geoinformation, are therefore applying various methods in order to examine, analyse and interpret pedestrian behaviour. The topics of this chapter include several key terms and expressions that are defined as follows: a trajectory or a track is the path a moving object follows through space; Position Determination Technologies (PDT) comprise technologies providing the location of an object or a person by using a wireless device; video-based data collection is defined as the process of capturing visual data with one or multiple cameras and analysing the captured video contents in an automatic manner with methods from the domain of computer vision; in the context of pedestrian monitoring the use of observations means watching and registering pedestrian spatial activities; and survey techniques are self-report instruments like questionnaires or interviews, where information concerning spatio-temporal behaviour is based on the participants’ self-assessments of habits or preferences. The chapter provides an overview about common methods for monitoring and analysing human spatio-temporal behaviour. It comprises two main sections. The first part focuses on dataset generation (Section 2): commonly used methods for data collection are presented and discussed with respect to specific strengths and limitations. The second part focuses on data analysis (Section 3): several case studies are presented, describing the methods which have been used for analysing specific datasets. The chapter concludes with a comparison of the presented empirical methods with respect to several crucial criteria (e.g. positioning accuracy, covered region, main cost factor), which provides a concise overview about the applicability of specific methods for different research foci on human spatio-temporal behaviour (Section 4).

2. Dataset Generation This section focuses on empirical methods of motion data collection. The first examples present common methods to track individuals by the use of mobile devices. It starts with satellite-based localisation (GPS) which is described in Section 2.1. Section 2.2 describes the potentials of positioning based on mobile phone cells. Data collection with Bluetooth is described in Section 2.3. The following section focuses on video-based data collection (Section 2.4). Observational research and survey techniques are described in Section 2.5 and 2.6, respectively. At the end of the first part of this chapter, a brief overview about additional data collection methods (laser scanning, sensor mats, RFID, WLAN) is provided in Section 2.7. 2.1. GPS Data Collection Global Navigation Satellite System (GNSS) is the general term for satellite based Position Determination Technologies (PDT). Controlled by the USA, in the mid 90s GPS

(Global Positioning System) became operative as the first worldwide available satellite based positioning technology. GPS is based on 24 non-geostationary satellites which are geographically distributed in such a way that at any spot on earth one can receive signals from at least four satellites. This technology is widely spread due to low prices of commercial GPS devices and free system usage [37]. GPS devices receive emitted satellite signals and calculate the current position measuring propagation time. GPS receivers can be used for real time positioning applications providing three dimensional locations together with accurate timing information. Other GNSS which are currently under development include GALILEO in Europe and GLONASS in Russia. 2.1.1. Advantages and Limitations of GPS Tracking The accuracy of GPS can reach up to three meters, and positioning is feasible with high frequencies, e.g. every second. Standardised interfaces and communication protocols (serial communication interface and NMEA-0183 format) allow easy storing of positioning data. Powerful GPS chipsets are increasingly integrated in low-cost mobile phones. Accurate positioning requires unobstructed satellite signals. Consequently, GPS is not suited for indoor environments. Furthermore, in urban regions GPS performance vastly decreases due to shadowing and multipath effects. Depending on the GPS device the first position after starting the device is received after a few minutes. This so-called time-to-first-fix (TTFF) describes the maximum time which is required for determining the first position given by cold, warm and hot starts [37,54]. Assisted GPS (A-GPS) is often used to reduce TTFF by providing GPS chipsets pre-knowledge about the satellites’ current positions. This service is mainly available for mobile phones, where the relevant information is transmitted over the telecommunication network. Note that A-GPS does not improve position accuracy. Accuracy can be improved by Differential GPS (D-GPS), which considers additional information about orbit errors provided by terrestrial communication networks. The accuracy of GNSS can also be improved by increasing the number of satellites, as planned for GALILEO [26]. New generations of GPS devices use improved chipsets with highly sensitive reR and Sirfstar III . R These GPS devices benefit in position quality, ceivers such as MTK size, weight and power consumption.

2.1.2. Potential Fields of Application GPS enables collecting long-term geo-referenced trajectory data of individuals equipped with a GPS device. GPS tracking can replace the post-hoc travel diary [37] and supports travel choice behaviour and activity pattern research. For example, Janssens et al. [22] use GPS in combination with a handheld device for collecting information about travel choice behaviour in Flanders, Belgium. Bohte et al. [8] used GPS devices for tracking families in three Dutch cities for one week. De Bois [12] equipped 15 families in three neighbourhoods in Almere with GPS units in order to evaluate spatial usage (see also Section 3.2). Shoval [46,45] used GPS to track tourists. Hovgesen [21] and Nielsen [33] use GPS to track pedestrians in parks in Aalborg and in schools in Copenhagen, respectively. Millonig and Gartner [30] employ a combination of shadowing (see Section 2.5) and GPS for investigating pedestrian motion behaviour. Van der Spek [55] used GPS devices for tracking visitors of three historic city centres in the Spatial Metro project (see

Section 3.1). Apart from fields of urban design, spatial planning and human geography, the collected data is also useful for social sciences, simulation purposes and prediction models, such as Space Syntax [59]. 2.2. Cell-based Positioning Cell-based positioning is a PDT relying on mobile telecommunication technology (mainly ’Global System of Mobile Communication’ (GSM) and ’Universal Mobile Telecommunications System’ (UMTS)). Limited localisation capabilities of GPS (see Section 2.1.1) have led to the idea of using GSM/UMTS as an alternative positioning technology for Location Based Services (LBS). A comprehensive study in Italy and the USA showed that the assumption that phones would be connected to the closest antenna was true only in 57% of the experiments [53]. This makes localisation and tracking with mobile phones a non-trivial task. This section provides an overview of the network architecture (Section 2.2.1), describes passive and active data collection methods (Section 2.2.2), and discusses potential applications (Section 2.2.3). 2.2.1. Network Architecture and Location Techniques Figure 1 (a) shows a realistic view on the GSM/UMTS cell network structure, especially for urban areas. Cell coverage of base transceiver stations (BTS) is simplified using ellipse models and sharp bounds, and is classified into three different types: (1) cells covering a large area with low connection capacity to supply a basic level of service due to possible technical breaks, (2) strategically distributed cells to cover the whole area adjusted to the required capacity, and (3) additional cells to cover shadowed regions (e.g. caused by high-rise buildings). This results in a complex network structure where one place might be covered by multiple cells. Cell sizes also vary strongly depending on the environment. For densely populated regions, cell coverage decreases down to approx. 50m (high network complexity). For rural areas, cell size may increase up to 3-30km (low network complexity). This coherency between cell-sizes and population density mostly allows good approximations, but may differ in some cases. This raises the question of which cell will be selected for a place covered by more than one cell. In many cases the strongest cell in the vicinity is selected, which in general does not need to be the one associated with the closest cell center. Cell selection depends on many – partly still unknown – factors, including Signal to Noise Ratio (SNR), network load balance, mobile phones cell selection algorithm and previously selected cells. Figure 1 (b) shows the results of an investigation in the region of Vienna (Austria). Four separated places (’work’, ’home 1’, ’home 2’ and ’weekend’) with corresponding cell positions (Centre Of Cell-coverage) are shown. This result is based on a half year permanent observation of one volunteer and confirms that the probability that one place is composed by multiple cells is higher in urban areas compared to rural areas (for details see [38]). The presented results are based on Cell Of Origin (COO) positioning technique. All cells in the telecommunication network are identified by a unique ID (Location Area code and Cell Identification Number). Hence, geographic coordinates of the BTS or of the theoretical centres of cell coverage – depending on the network provider – are assigned to the unique cell-IDs. COO is the most commonly provided and used localisation technique, because it does not require additional costly network equipment.

1 BTS

Basic coverage of an area

2 BTS

4 BTS 5 BTS

3 BTS

Strategical distribution in an area

Reduce shadowing effects

(a) 48.28 48.26

work (urban area)

48.24

latitude

48.22 48.2

home 1 (suburban area)

48.18 48.16

home 2 (suburban area)

48.14

weekend (rural area)

48.12 48.1 16.4

16.45

16.5

16.55 longitude

16.6

16.65

(b) Figure 1. (a) Cell-network view. (b) (Multiple) Cells dedicated to places in urban, suburban and rural areas of one person (see [38])

Alternative techniques like ’Enhanced Observed Time Difference (E-OTD)’, ’Time Of Arrival (TOA)’ and ’Time Advance (TA)’ are based on signal propagation delay measurements. This is challenging with respect to clock synchronisation. ’Angle Of Arrival (AOA)’ techniques use the knowledge about antenna orientation in the mobile phones’ vicinity to limit the possible positioning area. The major drawback of the alternatives to GSM/UMTS is the required costly additional provider network equipment, limiting their availability in practice. An overview about alternative GSM/UMTS localisation techniques can be found in [62].

2.2.2. Data Collection Methods The mobile telecommunication device receives information about the selected cell from the network. The unique cell identification number and geographical position (if supported) can be retrieved from the device (client-based) or the cell-network (server-based). Client-based collection Client based data collection has to cope with the fact that cell information is mainly processed in the GSM/UMTS modem unit of a mobile device and often not directly accessible. Due to the growing LBS-market and mandatory emergency services E911 and E112, mobile phones and cell networks increasingly provide interfaces to position information. The Java 2 Mobile Edition Programming Interface JSR179, for example, provides an abstract interface to location information, independent of the underlying positioning technology. Hence, a location request using JSR179 automatically checks locally available positioning modules like GPS and GSM. The provided data – position and quality information – can be stored locally or transmitted with GPRS/UMTS. Server-based collection Cell network providers do not store all position logs of mobile phones [10]. However, many activities in the cell network, such as opening a data connection, phoning someone or changing the location area, implicitly cause a network location update. Location updates are stored by the network providers. Hence, location information is available, but the level of completeness depends on the users’ activities. Mobile office systems like BlackBerry are permanently online, hence their location is updated at any cell change (due to uninterrupted handover processes). The market penetration of mobile online systems might increase in the future, and such location update log files could contain high quality position information. Privacy issues are important in this respect, and may restrict collection/access to such information. Alternatively, some network providers give access to a location interface. This interface allows performing an explicit location update and retrieving the position of a given mobile subscriber number, hence allowing people tracking. Due to privacy issues, the located person must agree to such a service and usually has the opportunity to switch off the service at any time. Each location query causes process time within the network system, limiting the number of requests and sample rate. In the course of the investigations described in [38], ten persons have been observed using a location interface for six month with a sample rate of five minutes. 2.2.3. Potential Field of Application Cell-based positioning primarily benefits from the indoor availability and the high market penetration of mobile phones. Taking into account the restricted position quality (at least in rural areas) and some network-based phenomena (e.g. multiple cells using COO), this technology provides a reliable framework for Location Based Services. 2.3. Data Collection with Bluetooth Bluetooth is a short distance radio communication technology, which is used for data exchange in e.g. digital cameras, printers, handhelds, laptops and mobile phones. Blue-

tooth uses a broadcasted signal for device identifications which can be obtained by any Bluetooth device in the near environment (up to 100 meters). Each Bluetooth device carries a unique identifier number (MAC address). This identification process can be used for (1) passive tracking and (2) active tracking. In ’passive tracking’ a network of Bluetooth devices is distributed in a given environment. Each network device periodically scans for Bluetooth devices in the vicinity using the mentioned broadcast signal. Hence, other Bluetooth devices within the observation area can be traced from the network using timestamps, unique identification numbers and the related network nodes. As a precondition, passive tracking is only possible if Bluetooth visibility is activated on the tracking device. Passive tracking has been applied in city centres (e.g. Norwich), at campus areas (e.g. Koblenz Landau) or in shopping malls [15]. In ’active tracking’ the tracked device itself periodically scans for other Bluetooth devices in the vicinity. In this case ’unintelligent’ and low-cost Bluetooth devices (Bluetooth beacons) are distributed in a given environment. Hence, special software including pre-knowledge of the network infrastructure is required on the tracked device for tracing purposes. Here, the tracing information is stored locally on the device. An example of this type of research was carried out by Millonig and Gartner [30] in a shopping mall: programmed cell phones/handhelds with Bluetooth were distributed at the entrance collecting the ID’s of 50 Bluetooth beacons. In Aalborg zoo the same Bluetooth beacons were provided to children to determine their location in cases they got lost (see in [48]). 2.3.1. Advantages and Limitations of Bluetooth Tracking Approaches Using Bluetooth for monitoring peoples’ behaviour has two major limitations. Firstly, the number of people with activated Bluetooth is rather limited (experiences indicate a percentage of less than 5%) and therefore the covered sample often doesn’t represent the target group. Secondly, the accuracy is depending on the distribution density of the beacons. There are two main advantages applying Bluetooth technologies for tracking studies. Bluetooth is one of the technologies allowing position determination in indoor environments. And concerning privacy issues, participants in Bluetooth tracking studies can fully control the possibility of being tracked. 2.3.2. Potential Fields of Application In general, Bluetooth tracking applications require a network composed of a number of Bluetooth hardware. Hence, hardware and installation costs vastly increase depending on the required tracking accuracy and the covered infrastructure. For this reason, this technology is mainly applied on small to medium scales (e.g. buildings or blocks of buildings). In many cases Bluetooth devices are just deployed at entrances of infrastructures to gather information about the number of people within an infrastructure (carrying visible Bluetooth devices). However, in principle Bluetooh can extend GPS abilities with respect to indoor tracking. Here, customised software with pre-knowledge of the observed environment is required. 2.4. Video-based Data Collection Research in the field of automatic video surveillance with computer vision has matured considerably over the past 10 years, with technical publications reporting significant

progress. At the same time, abilities of computer vision are often hyped and exaggerated by industry and media, benefits are glamorised and dangers dramatised in movies and politics [17]. Digital video footage is composed of temporal sequences of two-dimensional pixel arrays having a fixed number of rows and columns, called video frames. There are three key steps in automated analysis of digital video data. 1. Detection of interesting objects in individual video frames. The objective of this step is to identify pixel sets belonging to object classes like pedestrians and cars. Large variations in human pose and clothing, as well as varying backgrounds and environmental conditions (lighting conditions, shadows) make the problem of pedestrian detection particularly challenging from a computer vision perspective. Many interesting pedestrian classification approaches have been proposed in the literature; an overview is given in [32]. An overview about human pose estimation can be found in [36]. 2. Tracking of objects. The objective of this step is to associate detected objects between video frames (and between multiple cameras) in order to obtain trajectories. Difficulties in tracking can arise due to abrupt object motion, changing appearance patterns, occlusions (people-to-objects, people-to-people, people-toscene) and camera motion. The survey in [61] categorises and provides detailed description of tracking methods and examines their advantages and disadvantages. 3. Analysis of people tracks to recognise their behaviour. The objective of this step is to automatically extract semantics out of spatio-temporal trajectory data. Trajectory dynamics analysis provides a medium between tracking and highlevel analysis. Typical motion, for example, is repetitive, while interesting events rarely occur. This repetition enables event analysis in the context of learned motion. The survey in [31] provides an overview of activity analysis in surveillance video based on object tracking. Note that many trajectory-based analysis methods could also be applied to GPS or related data. 2.4.1. Advantages and Limitations of Video-based Data Collection Video-based data collection enables obtaining pedestrian trajectories with high spatial and temporal resolution both in outdoor and indoor environments covered by cameras. The biggest advantage of sensed visual data is that the raw data basically contains richer information than mere spatio-temporal data. Automatically extracting such information enables monitoring of the environment as well as everything within the scene, such as pedestrians, animals, vehicles. Such a high-level analysis remains, however, an extremely challenging problem. The most complex behaviours can only be understood in the correct context, and it is difficult to imagine general procedures capable of working over a wide range of scenarios [31]. The performance of automatic monitoring is heavily influenced by the imaging setup. For example, one way to avoid severe occlusions and the resultant problems with pedestrian detection and tracking is to capture the scene from a bird eye’s view. Most of today’s commercially available video-based pedestrian counting sensors rely on such configurations. A top view configuration, however, severely limits the area which can be captured by a camera. Wide angle lenses or fish-eye lenses can increase the covered area, but increasing viewing angles again introduces occlusions [29].

Figure 2. Transformation between video frame and infrastructure ground plane

An important issue is whether cameras are calibrated, i.e. the transformation between coordinates in the three-dimensional world and the pixel coordinates of the video frames has been identified. Combining calibration information with a site map or site model (even a simple ground plane model) enables the system to use the absolute size and speed of the detected objects. Figure 2 illustrates the transformation of pixel coordinates of a video frame (upper right part) into two-dimensional metric coordinates of a ground plane (lower left part). While such a plane-to-plane calibration technique enables synthetic top views of an environment, one must keep in mind that the accuracy of world coordinates decreases with increasing distances of world points from the camera. This is illustrated in Figure 2 by the increasingly blurred regions towards the left hand side of the transformed image. Many built environments have an architecture which does not necessarily allow full video coverage. Too many cameras might be required to cover an entire building. Additionally, the ceilings are often low (for example in subway environments), which will lead to shallow viewing angles with the related severe mutual occlusions. A large number of cameras might demand an impractical hardware setup. Smart cameras combine video sensing, processing and communication in a single embedded device and can therefore avoid video transmission and hardware resources for video processing. An overview of distributed smart cameras can be found in [40]. The majority of the existing automatic video surveillance techniques can only claim robustness and reliability for limited scenarios – with limited sensor networks, limited video footage, limited fault tolerance and small variability of scenes [39]. Furthermore, it is widely agreed that the object-based trajectory approach only works up to a certain complexity and density of people. For crowded scenes with many individuals, the mutual occlusions become so severe that currently no tracking algorithms can handle them effectively, even with a multi camera approach. The sub-domain of crowd analysis deals with information extraction without relying on individual object tracking [63]. 2.4.2. Fields of Application Video-based pedestrian tracking currently works best for scenarios with isolated individuals and loose groups of people, see [43] for a real system and an overview of other systems. Good results can be expected in terms of tracking quality if a steep camera angle can be assured. Outdoor applications will clearly only work when the environmental conditions allow a good imaging quality – darkness or fog for example will likely not produce interpretable video contents.

2.5. Observation-based Data Collection One of the most conventional ways to explore human motion behaviour is to observe human activities and the physical settings in which such activities take place. Observation techniques belong to the most fundamental methods in social and behavioural sciences and have a long tradition. They focus upon the investigation and interpretation of visible motion behaviour. Usually they are of an explorative or descriptive character and hence commonly used in research fields aiming at the generation of hypotheses rather than testing or confirming a hypothesis. First attempts to investigate human spatio-temporal behaviour therefore mainly employed observation methods (sometimes in combination with questionnaires or interviews, see Section 2.6) in order to describe existing phenomena and look for patterns and potential hypotheses. Observations of pedestrian motion behaviour, also known as behavioural mapping or “tracking”, have primarily been used for collecting data concerning the movements of visitors at museums and exhibitions [57,60,5]. Subsequent research has been focusing on several different fields of interest, e.g. routes and activities of pedestrians in urban environments [20] or tourism research [19,25]. 2.5.1. Types of Observations Observing human motion behaviour involves following the subject at a distance and recording the movements by drawing a line corresponding to the subject’s activities on a map of the investigation area. This can be done in a very simple way by just using a paper map and a pencil. Recently, also technology-enhanced approaches have been applied, e.g. by using a digital map (see Section 3.3). There are several types of observation techniques which can be applied: • Direct (Reactive) Observation: The researchers identify themselves as researchers and explain the purpose of their observations. • Unobtrusive Observation: The researchers do not identify themselves. Either they mix in with the subjects undetected, or they observe from a distance. • Participatory Observation: The researcher participates in what is being observed so as to get a finer appreciation of the phenomena. The decision of which type of observation technique to use is mainly determined by the purpose a study aims to follow. Each technique bears its specific advantages and drawbacks that have strong influence on the quality of collected data. For further information about observational research and related studies see [13,1]. 2.5.2. Advantages and Limitations in Observational Research Observations are frequently used in explorative and descriptive research, as they are usually flexible and do not necessarily need to be structured around a hypothesis. Another positive aspect is that observational research findings are considered to be strong in validity, because the observer is able to collect extensive information about a particular (though mainly just visible) behaviour. However, in terms of reliability and generalisability there are some negative effects. Replicating or generalising observed behaviour patterns may not be easily achieved, especially when behaviour patterns have been observed in the “natural” loci of activities, and not in laboratories under controlled con-

ditions. Moreover, the sample of observed individuals may not be representative of the population, or the observed behaviours may not be representative of the individual. Unobtrusive observations are useful if a particular field of interest has not been studied in depth so far. Another advantage of observing individuals without their knowledge is the fact that so-called “observer-effects” can be avoided. It has been shown that people who know that they are participating in a study tend to adapt their behaviour – consciously or subconsciously – to what they expect to be socially desired behaviour [34]. Hence, unobtrusive observation techniques are the only way to gain insight into “natural” behaviour patterns. Major drawbacks in this respect, however, are ethical concerns, as people cannot freely choose to participate in such a study. In non-disguised or participatory studies with subjects who agree to take part, researchers have to be aware of the fact that observer-effects may occur, and especially in participatory studies researchers may lose their objectivity. 2.5.3. Fields of Application As stated above, observation techniques are mainly useful in explorative and descriptive research. They are particularly helpful for identifying behavioural patterns and determining basic research hypotheses. Though, comprehensive insight into human motion behaviour and underlying cause-and-effect relations cannot be achieved by solely applying observation techniques. Therefore, a combination of observations with other methods (e.g. interviews) is often beneficial. Section 3.3 gives an example of such a combination of several complementary methods for analysing pedestrian behaviour styles. 2.6. Data Collection based on Survey Techniques Along with observation methods (see Section 2.5), interviews and questionnaires belong to the first methods of data collection that have been applied in human spatiotemporal behaviour research [20,25]. Regarding the exploration of influence factors determining human route decision processes and related preferences, survey studies still represent one of the most important data collection techniques in transportation studies. They are relatively cheap and allow the collection and analysis of data taken from comparatively large samples. Inquiries are commonly used to gather information concerning route choice preferences, individual habits, motives, and intentions [7]. However, as spatio-temporal behaviour is mainly based on subliminal decisions, responses may be incorrect and constructed ex post. Human behaviour is never fully determined by verbalised structures [34], and people tend to modify their behaviour when they know they are being watched: they portray their “ideal self” rather than their true self [14], and the accuracy of the results gathered from questionnaires may suffer. Consequently, studies relying solely on results based on questionnaire data will have to accept a certain degree of inaccuracy [20]. 2.6.1. Common Survey Techniques Questioning participants about their habits and preferences can be done in various ways. Questionnaires, interviews and trip diaries are among the most commonly used survey techniques in spatio-temporal behaviour research. In the following, a rough overview of these techniques and related advantages and limitations is provided.

Questionnaires The use of questionnaires is very popular, as they can be easily distributed (among pedestrians, via mail or the web) and therefore allow the collection and analysis of data taken from comparatively large samples at low costs. Standardised questionnaires are selfreport instruments which provide a written text comprising the exact wording of the questions and listing the possible answers. This results in the positive effect that standardised questionnaires provide valid and reliable data which can then be analysed and generalised (depending on the chosen sample). However, standardisation also includes two disadvantages: firstly, questions that are unclear to the participant may be answered incorrectly, and secondly, as participants are forced to choose among sets of prefabricated answers, the provided options may be incomplete and certain explanations may be missing. Furthermore, the provided set of answers can influence the participants’ responses subconsciously. Interviews Personal interviews (either face-to-face or via telephone) can be structured in different ways. In a structured type of interview, only the questions are standardised, the answers are expressed freely. An non-structured interview technique is applied when neither questions nor answers are standardised and the interviewer just follows a predefined field manual. The analysis of data collected in personal interviews is more complex than data collected by standardised questionnaires, as answers have to be categorised. However, this technique is especially useful if a survey aims at including persons who belong to so-called “hard-to-reach” groups, such as individuals who either usually refuse to participate, do not understand all the questions, or cannot fit into categories of answers designed for the average citizen [11]. Trip Diaries Another frequently used method is the time-space budgets technique, including recall diaries and self-administered diaries [52,46]. Recall diaries and interviews are strongly dependant on the participant’s memory, which will result in a lesser degree of accuracy. Self-administered diaries are written in real-time and can therefore provide very detailed information. However, they demand considerable effort on part of the subjects; consequently, only few people are willing to participate in these kinds of studies, and significant variation in the quality of the information must be expected. 2.6.2. Fields of Application Due to the limitations mentioned above, exclusively employing survey techniques may not be sufficient for the detection of human spatio-temporal behaviour patterns. However, as questionnaires and interviews offer the only chance to reveal relevant decision processes, certain research question require the use of these methods for understanding the influence factors determining spatio-temporal behaviour. A combination of different data collection methods including survey techniques is often recommended. Further details concerning such techniques, sampling, and interview types (e.g. structured/semistructured/unstructured) can be found in [1,11,13].

2.7. Other Methods In addition to the data collection methods described above, a number of alternative options for pedestrian tracking have been explored in the past. These measurement techniques are seen to be in their early stages of application to pedestrian tracking and have not yet matured. We will therefore provide only a brief overview hinting at their potential and their disadvantages. The discussion below focuses on two alternatives for pedestrian tracking mostly in indoor settings (i.e. laser range scanners and sensor mats), and two alternatives to localise pedestrians (i.e. RFID tags and WLAN). A laser range scanner measures the location of surrounding objects by inferring the distance of the closest objects in a high-resolvent angular grid from the travel time of the reflected laser beam. This results in a point cloud in the local coordinate system of the scanner. Using a number of laser range scanners simultaneously and converting the raw measurements into a joint coordinate system results in a combined point cloud. Laser range scanners are typically able to perform a 360 ◦ scan rapidly (e.g. with frequency of 10 Hertz). Figure 3(a) shows an experimental environment using laser range scanners, the PAMELA walking platform in London 2. In 2007 this platform was equipped with two laser range scanners located on two opposing corners of the platform approximately at hip height. Figure 3(b) shows a snapshot of one raw data set where the handrail is clearly visible as well as a number of pedestrians walking on the platform. The second 14 12 10 8 6 4 2 0 −2 −4 −5

(a)

0

5

10

15

20

25

(b)

Figure 3. (a) The platform on the PAMELA facility in London. (b) Snapshot of the raw data of one walking experiment on the platform (units: meters). Each dot represents one raw measured obstacle location.

step is to detect objects in the raw data. Subsequently, the objects can be tracked over time, for example with a Hidden Markov Model tracking, as described in [3]). Laser range scanners can provide reliable pedestrian trajectories with a precision of a few centimetres ([3] find a mean deviation from the true path of approx. 3 cm). However, they share the same limitations as video image based methods in terms of occlusion problems. Hence tracking of a large amount of persons requires a large amount of laser range scanners, which is costly. In order to reduce occlusion problems, laser range scanners have been mounted close to the floor for detecting legs. This incurs the problem of matching pedestrians and their corresponding pair of legs. Nevertheless, a number of publications list the successful application of the technology also in dense situations, see e.g. [64,44]. 2 We

thank Kei Kitazawa from University College London for making the datasets available to us.

Sensor mats are a promising alternative to computer vision based methods and laser range scanners to avoid occlusion problems. The two main problems in this respect are the detection of footprints on the mat (which is related to the spatial resolution of the sensors in the mat) and the tracking (complicated by the fact that while stepping the footprint disappears). First results and some references in this respect can be found in [51, 50]. No large scale applications are known. Laser scanning and sensor mats do not require the tracked subject to carry measurement devices and deliver tracking information. If it is possible to equip subjects with devices or existing devices are used for localisation purposes additional possibilities to localise pedestrians are possible. Radio frequency identification (RFID) tags have been used for this purpose. RFID tags are popular in the area of logistics in order to identify objects. Two different approaches can be distinguished: Active tags require power supply increasing the costs significantly while passive tags obtain power supply from signals sent by the RFID reader posing strong restrictions on the capabilities of the tags and limiting the distance from which they can be read. On the other hand passive tags do not require special maintenance operations. The signal strengths of one active RFID tag received at a number of RFID readers (three are necessary) are compared to tabulated ’fingerprints’ in order to estimate the position of the tag (see e.g. [41]). To the best of our knowledge the achievable accuracy using such a system in real world applications is insufficiently known, whereas using very dense RFID-reader distributions an accuracy of a few metres is achievable [23]. For passive tags on the other hand the tag is registered if it is sufficiently close to the card reader (typically in the magnitude of a few centimetres). This allows to track pedestrians in the vicinity of the readers but not in between. Beside access controls commonly used at offices, museums and the like, where typically the timing is the main point of interest, also tracking information is already used in large scale real life applications (see e.g. [49] which logs the ski lifts used by a person over the course of a day including the timing and the elevation skied). Publications on RFID technology are legion. Recent introductory surveys (not limited to RFID) can be found in [28,16]. Finally an emerging technology which poses similar challenges as RFID and Bluetooth (see Section 2.3) is wireless local area network (WLAN). See e.g. [56] comparing RFID, Bluetooth and WLAN technology. The same principles are used for WLANlocalising: Either only the specific reader which is closest to the WLAN device records a position or fingerprints are compared to signal strength patterns in order to estimate the location of the WLAN device. This technology is promising since an ever increasing amount of mobile devices (such as computers and smart phones) are equipped with WLAN capabilities.

3. Data Analysis/Experiences This part introduces selected case studies in order to illustrate applicable analysis methods for specific datasets. Section 3.1 describes a project focussing on using GPS for mapping pedestrian activities in historic city centres. Section 3.2 presents the results of tracking families in a suburbian residential area. Section 3.3 describes a multi-method approach for categorising pedestrian spatio-temporal behaviour in shopping environments. Section 3.4 describes a method to extract habits from motion history based on GPS mea-

surements. Section 3.5 presents an approach for identifying prominent places using cellbased positioning. Section 3.6 describes the development of a location aware pilot system for city tourists employing camera phones and GPS positioning. Section 3.7 presents a study using vision-based analysis finding stopping pedestrians in a railway station. 3.1. Spatial Metro: GPS-Tracks of Pedestrians in Historic City Centres Urban development requires careful reconciliation of different demands, including protecting historical and cultural heritage, creating attractive retail areas and future retail strategies, and sustaining and improving the quality of life in the central area. The Spatial Metro project aimed at investigating public space and subsequently improving city centres for pedestrians in Norwich (UK), Rouen (France) and Koblenz (Germany), which are cities with approximately 100.000 inhabitants and a historical centre. The TU Delft has developed a method using GPS for monitoring pedestrian movement to measure the effects of city investments like city beautification, street furniture, lighting and information systems [55]. In each of the three above cities GPS tracking devices were distributed to volunteers at two parking facilities for one week. GPS positions have been stored at a frequency of five seconds from 10am to 6pm. After the observation phase, questionnaires were filled by the volunteers. Only trip-related and general non-sensitive demographic information have been asked. In total 1300 pedestrians have been tracked and interviewed. On average only 60% of the spatio-temporal files were valid due to issues with GPS performance in urban areas, GPS fix times, power consumption, and outliers. GPS performance decreases vastly in urban areas (see Section 2.1.1), and people tend to go into buildings where GPS does not deliver position information. High time-to-first-fix rates also result in localisation gaps [37]. The collected data has been preprocessed with respect to cleaning and validation. Subsequently, valid data has been visualised in a Geographic Information System (GIS) by (1) drawing collected tracks on different GIS layers (e.g. on map layer, see Figure 4 (a)) and (2) plotting a density analysis of questionnaire data. The first type of visualisation combines spatio-temporal data with e.g. aerial images, access routes and arrival points, commercial activities, points of interest and investments. This enables analysing spatial conditions in relation to actual behaviour of tracked samples. The second type

(a)

(b)

Figure 4. GPS tracking results (a) Norwich: pedestrians tracked from two access points for one week. (b) Koblenz: layering of commercial activities and trajectories (all track points at 5 seconds frequency).

of visualisation delivers a set of specific spatial patterns based on the aspects of origin, familiarity, purpose and duration and age, gender and group type. Hence spatial patterns for specific groups of participants could be compared. Both ways of visualisation offer tremendous insight in pedestrian behaviour, leading to conclusions and opportunities for application in practice (see also Figures 4 (b) and 5).

Figure 5. GPS tracking results in Rouen: pedestrians tracked from Haute Vieille Tour (3D density analysis).

3.2. Tracking Families in Almere Almere is a poly nuclear Newtown in the “Flevopolder” and is composed of many suburbs. The last decade, the TU Delft has developed and applied several methods to analyse the available network between the city and related suburbs. GPS based behaviour analysis offers new perspectives for contradicting or confirming the developed theory of city planning for that region. Within three suburbs families with one or two children (aged between 16 and 18) were approached by the municipality to participate in the investigation. Fifteen families agreed, leading to approx. 50 participants. Questionnaires were filled and instructions were given at the participants’ home. Each participant was provided one GPS tracking device and a battery charging unit for one week. The spatial-temporal data has been collected at an interval of 2 seconds. Several hundred trips have been collected [12]. After the observation phase the volunteers were interviewed again. The tracks have been preprocessed (e.g. outlier removal) and have been split into trips based on activities (e.g. stays or entering a building). Incomplete tracks due to low GPS performance in urban areas (see Sections 2.1.1 and 3.1) resulted in some incomplete tracks which have been removed. Subsequently, interrelations between trips and questionnaires have been sought for advanced behaviour analysis. A GIS tool was used to visualise these query based results (see Figure 6). Either individual or combined activities of volunteers have been analysed. The data provided information about network usage by different modes of transportation and the areas of activities.

(a)

(b) Figure 6. GPS tracking results in Almere-Haven: mapping of activity patterns of 13 males (a) and 13 females (b) for one week resulting in a map based on actual use of the urban tissue.

3.3. Multi-method Approach to the Interpretation of Pedestrian Motion Behaviour In this subsection the currently ongoing project UCPNavi, aiming at the classification of pedestrian walking behaviour and related influence factors, is described. In this study we apply a multi-methods-approach comprising several complementary data collection techniques. For more details about the project, methodology and preliminary outcomes see [30]. 3.3.1. Methodology The selection of appropriate methods had to fulfil several conditions: Firstly, we aimed at collecting data of sufficient quality and accuracy in larger environments (indoor and

outdoor). Secondly, as it is assumed that people might change their behaviour if they know that they are being observed, an unobtrusive form of monitoring was to be included. Thirdly, visible behaviour patterns were to be combined with interview data in order to allow the identification of relevant underlying intentions, preferences, and lifestylerelated factors. The following three methods have been chosen in order to achieve an optimal combination of empirical data collection techniques in consideration of these preconditions: 1. Unobtrusive Observation (Shadowing): Observation of the “natural”, uninfluenced spatio-temporal behaviour of pedestrians; only visible behaviour, no insight to intentions and motives. 2. Non-disguised Observation (Tracking): Continuous observation over a long period in combination with standardised interviews (data from both the structural and the agent-centred perspective); observer effects possible. 3. Inquiry (Interviews): Motivations, self-assessments of individual motion patterns; responses can be incorrect and constructed ex post. The execution of unobtrusive and non-disguised forms of observation was not feasible in parallel. Hence, a two-step approach was designed, which also offers the chance of using preliminary results for ameliorating the methods and selected features of investigation in the second empirical phase. The unobtrusive observation method applied in the first phase of the study was used to collect anonymous data of people walking in public areas, who did not know that they were being observed. The process consisted of random selection of an unaccompanied walking person and following the individual as long as possible while mapping her path on a digital map. In total trajectories of 111 individuals with a balanced gender and age ratio have been collected (57 observations on a shopping street, 54 in a shopping mall). The collected datasets have been analysed according to the velocity computed between each marked point in the observed path, additionally locations and durations of stops within the trajectory have been detected. Subsequently, speed histograms of each trajectory have been compiled, showing the proportional amount of time an individual walked at a velocity within a specific speed interval. Figure 7 shows diagrams consisting of all histograms compiled from indoor and outdoor observations (speed intervals: 0.1m/s steps, 30 intervals). In order to compile initial classes of behaviour, the histograms of each investigation area have subsequently been classified using clustering algorithms (hierarchical clustering and k-means algorithm). For each class, characteristic attributes have been identified to describe preliminary types of spatio-temporal behaviour. 3.3.2. Preliminary Results The classification process resulted in three homogeneous clusters for the indoor datasets (containing 10, 14, and 30 individual observations per cluster). The outdoor dataset analyses produced eight clusters, with a vast majority (86%) of observations belonging to the first four classes. As an example the results of the indoor analysis are now explained in more detail. In total datasets of 54 observations have been classified. The three clusters of motion behaviour can be interpreted as “swift shoppers”, “convenient shoppers”, and “passionate shoppers” (see Figure 8). Four of the eight clusters resulting from the analysis of outdoor datasets solely comprise a number of one to three subjects. Among the other four clusters, all three indoor

Outdoor observations

Cases

Cases

Indoor observations

Speed Intervals

Speed Intervals t

Figure 7. Histograms of all indoor (left) and outdoor (right) observations. Rows present individual observations, columns present speed intervals of 0.1 m/s ranging from 0 to 3 m/s. Higher intensities represent higher histogram bin values.

clusters can be identified to a certain extend. Additionally, a cluster of specific behaviour patterns was identified: these “discerning shoppers” were mainly female, walked at comparatively high speed, stopped rather often but shortly, and showed – other than individuals belonging to the other clusters – a slight tendency towards specialised and exclusive shops. Cluster No. of subjects

Swift Shoppers

Convenient Shoppers

Passionate Shoppers

10

14

30

Gender: female male Age:

< 30 30 - 60 > 60 Average age Average speed

25-30

30-35

35-40

1.19 m/s

0.61 m/s

0.24 m/s 3.57 (13)

Av. no. of stops (max.)

0.3 (3)

1.36 (2)

Av. duration of stops

6.97 s

2.58 min

4.66 min

casual or convenient

casual or conservative

casual/trendy or elegant

food store

no main focus

fashion, specialities, drugstore, bookshop

Fashion style Visited shops/facilities

Figure 8. Behaviour clusters in the indoor environment.

3.3.3. Conclusion Preliminary results of the first empirical phase indicate that a number of homogenous behaviour patterns can be observed, especially in consistent context situations. Currently ongoing investigations using a non-disguised form of observation combined with detailed interviews include and test basic findings of the first analysis. The combination of several complementary empirical techniques is a very promising approach to gain

comprehensive insight to human spatio-temporal behaviour patterns, even though some limitations have to be accepted. 3.4. Extracting Habits from Motion History based on GPS-measurements This section describes the experiences from the analysis of a long term motion history measured using GPS technology. The emphasis here lies on presenting the main experiences and in particular the pitfalls in the analysis phase, full details can be found in [4]. The aim of the study was to investigate the possibilities to extract commuting habits from the motion history of one individual alone without any prior knowledge. To this end a commuter was equipped with a GPS-tracking device which was operated on the commute to and from work on a total of 27 days within the timespan of approximately seven weeks in 2006. GPS tracking was not always successful due to sensor failures of various causes, including the failure to turn on the tracking device (the quality of GPSsensors has increased in the meantime making other causes of sensor failure less of an issue). The GPS-tracker recorded a (location, timestamp) pair every 2 seconds providing a detailed motion history. On total 70000 (location, timestamp) pairs have been collected. The intention of the data analysis is to obtain knowledge on the commuting habits covering information on: • Points of interest (POIs) somewhat sloppily defined as points which are visited frequently, where the commuter either pauses or changes mode of transport. • The routes most frequently used. • The mode choice on these routes. • Temporal information such as the probability of route and mode choice for any given day of the week, the distribution of the starting times of trips and the distribution of travel times. Corresponding to these goals the analysis can be partitioned into the a number of phases, which are detailed as follows: Data preprocessing aims at increasing data quality. This is done by the detection of outlying observations, signal losses, as well as the application of smoothing operations to reduce the noise level. While this operation is performed as a first step in the data analysis, data quality can be also increased at later stages: After mode detection has been performed the knowledge of the transportation mode can be used in order to apply more efficient outlier detection and noise reduction methods. To give just one example unreliable observations can be detected and omitted for routes walked by monitoring travel speed implied by the measurements. Inaccurate measurements tend to imply inprobably high speeds, a phenomenon frequently encountered in urban canyons where low signal quality leads to the measurements slowly drifting away from the true locations which is corrected rapidly if subsequently strong signals are received. For more details on GPS reliabilty in weak signal environments see [27] and [24]. The preprocessing results in a number of trajectories of uninterrupted measurements, that is each trajectory is a sequence of (location, timestamp) pairs such that the distance of two consecutive timestamps is less than a prespecified margin. This set of trajectories is subject to the stop detection according to the algorithm proposed in [18]. The thus found stops are combined with the places of signal losses (potentially the signal is lost due to entering a building) in order to search for POIs. This set of locations is clustered

and the cluster centres are taken to be the POIs. For this procedure the starting points of the trajectories are not included, since due to relatively long time-to-first-fix (TTFF) of the GPS-tracker used of up to five minutes the first observed location does not closely match the trip start location. This problem occurs, since the GPS-tracker is turned on at the trip starting time without delaying the trip until the location is found for the first time. For TTFF of less than ten seconds (which is achieved by state-of-the-art GPS trackers) this is less of an issue. With POIs being identified the starts of the trajectories are matched to the nearest POI and the trajectories are extrapolated backwards in time to start at a POI. This results in a set of trajectories, which all start and end at POIs. Next trajectories are dissected into shorter segments by breaking up trajectories passing the vicinity of POIs. This leads to a new set of trajectories starting and ending at POIs and not coming close to other POIs. This is necessary since some mode changes do not result in stops in one direction but do so in the opposite direction. One example in this respect is walking away from a subway station: while exiting one typically does not stop; whereas when entering the subway many times at least a short time of waiting is necessary. The output of this step is a collection of trajectories which connect the POIs and where ideally each trajectory only corresponds to one mode. The next step is the detection of the main routes used by the commuter. To this end the algorithm of [2] is used resulting in a number of main routes. The main function of this step is to eliminate infrequently used routes for which it is not possible to detect clear usage patterns due to too small sample sizes. Next for each trajectory indicators are calculated which are used in order to detect the transportation mode used for this trajectory. The indicators contain a number of factors mainly built using speed and speed variation. The classification uses a tree based classifier built based on manually labelled example trajectories. This simple approach works surprisingly well in this case. In a final step temporal information is collected separately for each distinct route/transportation mode configuration. The case study showed that it is possible using only the motion history to correctly identify the main routes used by the commuter. Furthermore mode detection was able to identify the transportation mode used with a high precision. The only errors made were misclassifications of walking and biking (due to high measurement noise and inappropriate outlier detection for the routes walked) and confusion of public transport and inner city car usage. In particular the second point might be due to a lack of a significant number of trajectories documenting inner city car usage. This is a topic for further research. Additionally the indications of the case study need to be validated using a larger set of users which is the topic of an ongoing research project. 3.5. Finding Prominent Places Most of today’s Location Based Services (LBS) provide information based solely on a user’s location, not taking into account context knowledge about the user’s current situation and needs. This often results in low-quality and inappropriate information to the user. Hence, in order to provide user-oriented services, an improvement of the responsequality of information requests is required. This section outlines a methodology for finding and classifying places where the user regularly stays in her life, in the following denoted as ’prominent places’. A detailed description of this work has been published in [38].

3.5.1. Collecting cell-data In order to draw meaningful conclusions about the motion behaviour of individuals, a sufficiently large amount of localisation data is required. 250 000 cell-based position measurements from ten volunteers obtained during half a year of permanent observation using a constant sample rate of five minutes have been collected. On average, 25 000 positions have been obtained from each volunteer in cooperation with the biggest Austrian mobile phone provider. 3.5.2. Analysing cell-data The analysis is split into two steps. First the collected cell-data are taken to find places where the volunteers spend most of their time. The found places are subsequently automatically annotated with semantics by labelling them with e.g. ’home’ or ’work’ (see Figure 9).

finding potential cell candidates

grouping cells by linkages

definfing model sequence

grouping cells based on visiting frequencies

computing individual cell-network

comparing and classifying

computing place sequence

Figure 9. Workflow of analysing cell-data for (a) finding and (b) classifying prominent places.

a) Finding prominent places Prominent places are defined as places where the user spends most of her time. In general, such places will be mainly ’home’ and ’work’ locations. Hence, cells where one volunteer has been located more often than in others (using a constant sample rate) must correspond to her prominent places. Cell-candidates are therefore first identified by filtering out cells exceeding a high dwell time. In some cases there is a one-to-one relation between a cell candidate and a prominent place. However, it often happens that one prominent place is assigned to multiple cells (see Section 2.2 for details on Cell Of Origin data collection). We have therefore developed an approach of an individual cell-network graph. Nodes of the cell-network graph represent cells and links represent cell changes. Cell-candidates are grouped if the topological distance between them is lower or equal than a predefined number of links. Due to unknown network characteristics, it might happen that not all expected cellcandidates re-presenting one prominent place are linked and therefore correct grouping will fail. To overcome this case, a further approach is used to add missing cells to related prominent places by comparing time series of visiting frequencies.

b) Classifying prominent places After grouping is finished we can compute a sequence of prominent places ordered by visits through a work day based on the visiting frequencies. At the same time we can manually define a daily routine for such work days (here [’home’ ’work’ ’spare time’ ’home’]). By comparing these two sequences we can label the computed prominent places for finally giving them semantics. 3.5.3. Experimental results The presented methodology has been applied to a 250 000 cell-based positioning data set collected by ten volunteers during a half year of permanent observation. Eleven of twelve home locations (92%) and nine of ten work locations (90%) have been found and correctly classified. (Two volunteers moved their home during observation phase.) Each volunteer has validated the result based on her provided cell-based positioning data with respect to the correctness of found and classified prominent places. All found prominent places are close to the real location. Geographical accuracy of the found places mainly depends on the cell-network distribution in the surrounding area and cannot be influenced by the method used. Hence, no quantitative validation about the localisation quality was performed. In Figure 10 is an example for visualising the results in Google Earth. This visualisation was used to validate the results together with the volunteers. The demonstrated grouping and classification results are promising and can be used as basis for improved LBS.

Figure 10. Example for visualising prominent places in GoogleEarth. Highlighted rectangles indicate the composition of cell-based positions for prominent places.

3.6. Mobile City Explorer Mobile City Explorer (MCE) is a project implementing the concept of an innovative mobile guide for personalised city tours. The concept was derived from the user perspective, integrating features for guiding the tourist according to her preferences and actual behaviour, identifying objects using object recognition, and collecting pictures and route information in an automated travel diary. A location aware pilot system for city tourists employing camera phones and GPS positioning technology has been developed in 2006. The mobile travel assistant guides tourists through the city, suggests Points Of Interests (POI) matching the tourists’ interests and recommends appropriate routes. If the tourist deviates from the recommended route, the tour is recalculated according to her personal interests and time constraints. At the same time the personal multimedia travel diary collects pictures, videos, text comments and acoustic impressions captured by the user along the route. The overall concept of the MCE-project has been published in [58]. This section introduces the results of the GPS-based motion behaviour analysis component of the MCE [42]. 3.6.1. Motion behaviour analysis Here, the sequence of visited POIs constitutes the central source of information to learn interests of city tourists. Motion behaviour analysis intends to find out in real time the POIs a user is visiting. Here a visit is characterised by spending some time at one place. An online stop detection algorithm is used to find stays in the motion data of the tourist. Stays detected by this step are classified into ’short stays’ and ’long stays’ depending on the dwell time. Here, short stays close to a POI (at least two minutes) may indicate user interest and long stays (at least five minutes) definitely indicate user interest to that certain POI. The extracted information is taken for learning the users’ interests (based on POI pre-knowledge). Hence, automatic POI suggestions and on-trip adaptations based on pre-selected interests and individual tourist motion behaviour is provided by the system. 3.6.2. Experimental results The evaluation took place in Vienna with seven city tourists (unfamiliar to that city) exploring the first district using the MCE system. A total of 21846 GPS data points were collected. At the city tour start, the tourists were prompted to select initial interests for the user profile initialisation. Then interesting POIs were assembled to a city tour and the tour map was provided to the tourist. Each time the tourist came close to a suggested POI, a notification containing a short description and a link to detailed information was sent to the tourist. Each stop at a POI resulted in an adaption of the personal interest profile and subsequently led to a reassembled city tour. Figure 11 shows strong deviations between the suggested tour after visiting 6 POIs (a) and after visiting 12 POIs (b). Based on five initial selected interests (History and sub categories) 14 new interests were learned on tour and the new main interest was Art & Culture and not anymore History. Figure 12 shows the adapted tour maps (a) after visiting 6 POIs and (b) after visiting 12 POIs. In both maps the path and the suggested POI sequence is shown. The results of the Mobile City Explorer project are promising and some concepts might in future be used in tourist applications.

User 153 Interest Profile [12]

0.12 0.1

0.08

0.08

0.06

0.06

0.04

0.04

0.02

0.02

0

0

Art

&c ult

ure

Arti Art& Cu sts & M lture Cu us lt ic ex ural F ans hib a it ion cilitie Art s m &c ult useu ure m ev en Arc ts B u hit ec ildin tur gs e Mid style s d Mo le ag Ne wC de es r n on str Build time uc ted ing ty s p Mo nu es me nts His H tor His ic e istor tor ic xh y a His ibitio l eve tor n m nts ica us l e His Monu um tor me ica l p nts ers o Re ns ligio Sc Sc C n ien ien ce hurch ce & ex es hib Tech it n Fa cto ion m ique rie u s in seu stitu m tio ns Ec on om y

0.1

(a)

ure

0.14

0.12

&c ult

0.14

Art

0.16

Arti Art& Cu sts & M lture Cu us lt ic ex ural F ans hib ac ilit Art ition mu ies &c s ult ure eum ev en Arc ts hit Build ec tur ings es ty Mid les dle Mo Ne wC de ages rn on str Build time uc ted ing ty s p Mo nu es me nts His H tor His ic e istor ica tory xh His ibitio l eve tor n m nts ica us l e His Monu um tor me ica l p nts ers o Re ns ligio S Sc C n ien cien ce hurch ce & ex es hib Tech it n Fa cto ion m ique rie s in useu stitu m tio ns Ec on om y

User 153 Interest Profile [06] 0.16

(b)

Figure 11. Example for learned user interests (a) after visting 6 POIs and (b) after visiting 12 POIs.

(a)

(b)

Figure 12. The city tour map (a) after visiting 6 POIs and (b) after visiting 12 POIs.

3.7. Video Analysis of a Train Station Hall This section describes some results of a case study for obtaining and analysing long-term video-based pedestrian track data within a large hall of an Austrian railway station[9]. One major objective was to identify places within the railway hall where pedestrians frequently stop or walk slowly. Identifying such stops can provide insights into possible congestion areas, waiting areas and areas showing an aggregation of stops. Furthermore, places where pedestrians stop potentially coincide with areas where they are reorientating in order to find specific targets. Automatic video-based analysis of an infrastructure can provide rich information about the spatio-temporal motion behaviour of pedestrians. Being aware of the limits of current vision-based people tracking systems (see Section 2.4), we wanted to get an idea of what information could be achieved with a set of temporarily digital surveillance cameras and a video-based tracking software.

(a)

(b)

(c)

Figure 13. Column (a): Simulated camera views via CAD modelling; Column (b): Real camera views with intermediate tracking result. Row 2 shows a gray information booth in front of the information display with circular layout. This information booth was built days after CAD-modelling and is therefore not represented in column (a); Column (c): Spatial distribution of detected stops with R = 1 meters and T = 3 seconds; coordinates refer to the ground plane – the dashed polygon indicates the camera’s region of interest.

3.7.1. Camera Positioning A strategic placement of cameras is essential to guarantee a good coverage of two levels of the hall with a limited number of cameras. The goal was to clearly determine the fields of view of the cameras in advance in order to select the necessary camera positions and lenses. This prevented ad-hoc and inadequate settings during the limited installation time window. The imaging system was not intended to be permanently installed. This transient nature of data acquisition prohibited arbitrary installation of power supply for sensor locations or repeater locations. The equipment for the video recording had to be placed in lockable rooms not accessible for the public. We have decided to simulate the field of view of each camera with 3D modelling in the spirit of [35]. A total of seven sensors has been defined, and two virtual camera views are provided in column (a) of Figure 13. 3.7.2. Recording and Tracking The infrastructure has been imaged during 13 days, and produced 15-minutes video samples at selected times of the day. Total recording time has been 100 hours per camera. A pedestrian tracking software described in [6] was applied to the digital video streams. The pedestrian tracks in the video frame have been transformed to world coordinates of the respective ground floor, as illustrated in Figure 2 of Section 2.4.1. The tracking algorithm delivers a sequence of timestamp/location states which have been permanently stored. A snapshot of the tracking results is shown in column (b) of Figure 13. Note that due to algorithmic reasons the tracking algorithm interprets any person standing still for more than 100 video frames (4-5 seconds) as part of the background and hence loses track.

3.7.3. Preprocessing and Preliminary Analysis We have analysed 18 video samples of one day obtained from five cameras. This amounts to a total of 22.5 hours of video material. The output of the pedestrian tracking software is spatially and temporally fragmented due to the issues described above and in Section 2.4. In a first step, obvious outliers have been removed: trajectories with fewer than six states and all points whose world coordinates lie outside of the building due to tracking errors have been discarded. The latter phenomenon results mainly from tracking errors in the vicinity of the camera’s vanishing line. The remaining tracks have been smoothed in order to reduce the noise level. 3.7.4. Results for Stopping Pedestrian Detection In the algorithm of [18], a trajectory is said to contain a stop at position (x, y) if the trajectory enters a circle of radius R meters centred in (x, y) and leaves this circle after spending more than T seconds in the circle. This does not define a unique stopping point – we have therefore averaged all positions within the circle in order to be unique. We have set R = 1 meter, and for the duration T we have alternatively used two and three seconds. Here two seconds implies that any person moving with a speed of less than a meter per second will be classified as ’stopping’. This speed corresponds to slow walking as well as stopping, allowing identifying places where people stop or are slowed down. Column (c) of Figure 13 illustrates the results of the algorithm applied to the track data obtained during one day, using R = 1 meter and T = 3 seconds. The plots have been produced using a standard kernel density estimator [47], with brighter areas corresponding to regions with frequent stops. The analysis is based on the output of the tracking software, hence tracking failures are not visible in the data. The plots should therefore always be interpreted with care. The results for the first camera (covering a part of the lower level of the hall) show a concentration of stopping places at the areas around the ticket machine, at the exit (where newspapers are sold), in the area in front of the escalators and in front of a shop entrance. The largest concentration occurs where the main pedestrian streams intersect. Some found stopping locations closely match the expectation reconfirming a priori knowledge. Other locations, however, were surprising. In particular, the identified area just outside the shop as a place where people frequently stopped came as a surprising result. It can be explained by the fact that shoppers buying provisions for the journey are often accompanied by fellow travellers waiting in front of the shop to look after the luggage. The results for the second camera (covering a part of the upper floor) show stopping areas at the information booth on the left part of the plots, underneath both sides of the information display and around the wagon information stand. Again these findings are in agreement with expectations. More results of stopping detection, also in terms of velocity calculations from the track data can be found in [9]. The findings of video-based analysis have been used for supporting the design of a guiding system.

4. Conclusion: Applicable Methods for Specific Research Foci The presented projects and case studies exemplify that pedestrian spatio-temporal behaviour is interesting for a variety of different research fields. The complexity of spatial

Table 1. Comparison of data collection methods for pedestrian monitoring. Covered region

Indoor / outdoor

Accuracy 1

Counts / tracks / determinants

Main cost factor

Comments

Accuracy depending on environment (performance decreases in urban areas), world wide availability, georeferenced data, GPS modules integrated in mobile devices, cheap high sensitive receivers available.

medium, large

outdoor

approx. 3 – 50m

tracking

technical equipment

Cell-based positioning passive

large

indoor & outdoor

approx. 100m – 3km

tracking

provider licences

Cell-based positioning active

large

indoor & outdoor

approx. 100m – 3km

tracking

technical equipment

Accuracy depending on provider network density, software on mobile client is required.

Bluetooth passive

small, medium

indoor & outdoor

approx. 5 – 10m

tracking

technical equipment

Infrastructure equipment is required, observable mode has to be activated on clients, central data collection.

Bluetooth active

small, medium

indoor & outdoor

approx. 5 – 10m

tracking

technical equipment

Infrastructure equipment and software on mobile client is required.

Video analysis

small, medium

indoor & outdoor

varying

tracking

technical equipment

Accuracy depending on frame size, camera perspective; calibration. outdoor performance strongly depending on environmental conditions.

Observations unobtrusive

medium, large

indoor & outdoor

varying

tracking

manpower

Information about “natural” behaviour, accuracy depending on conditions, technique and observer skills, potential ethical issues, laborious analysis.

Observations non-disguised

medium, large

indoor & outdoor

varying

tracking

manpower

Strong observer effects, accuracy depending on conditions and technique, laborious analysis.

Questionnaires

medium, large

indoor & outdoor

low

determinants

manpower

Large sample possible, strong observer effects, information limited due to predefined categories.

Interviews

medium, large

indoor & outdoor

low

determinants

manpower

Detailed information, strong observer effects, accuracy depending on participants’ memory and awareness of their own behaviour, laborious analysis.

Trip diaries

medium, large

indoor & outdoor

medium, low

determinants

manpower

Detailed information, real-time diaries labour-intensive for subjects, recall diaries dependant on subjects’ memory (less accuracy).

Laser-scans

small

indoor & outdoor

few centimetres

tracking

technical equipment

High accuracy, crowded scenes require many scanners due to occlusion problems (costly).

Sensor mats

small

indoor & outdoor

?

counting

technical equipment

Counting accuracy depending on conditions and technology, mostly limited to single lanes. Tracking applications are being investigated.

RFID passive

large

indoor & outdoor

few centimetres

tracking

RFID readers

Length of tracking gaps depending on costly RFID-reader distribution. Infrastructure equipment is required. Cheap RFID – tags.

RFID active

large

indoor & outdoor

approx. 10 – 100m

tracking

technical equipment

Accuracy depending on environment condition and positioning technology. Infrastructure equipment is required.

WLAN

large

indoor & outdoor

approx. 30 – 300m

tracking

technical equipment

Accuracy depending on environment condition and positioning technology. Infrastructure equipment is required.

GPS

1 Experienced

values.

Accuracy depending on provider network density, central data collection using network provider interfaces.

activities and the interdependencies between those activities and the surrounding physical settings raise numerous questions which can only be tackled by accurately investigating pedestrian motion behaviour and related influence factors. A great number of empirical methods and technologies have been developed for the purpose of collecting and analysing data in order to understand, describe, model and predict pedestrian spatiotemporal behaviour. In the field of ambient assisted living the sub-domains of people tracking and event detection are particularly important, because they provide motion trajectory data of people and semantic interpretation, respectively. The applicability of particular data collection methods strongly depends on the specific focus of the research questions. Therefore, suitable methods and techniques need to be carefully selected. Table 1 juxtaposes the empirical data collection methods described in this chapter in a qualitative manner according to the following criteria: • Covered region: refers to the dimension of the investigation area: “small” (e.g. a room), “medium” (e.g. a building, a street), “large” (e.g. city, county). • Indoor / outdoor: indicates whether the method can be used in indoor or outdoor areas (or both). • Accuracy: refers to the positioning accuracy level a method can achieve under normal conditions. • Counts / tracks / determinants: refers to the kind of collected data. Methods which can be used for tracking are usually applicable for counting as well. “Determinants” is related to factors influencing spatial behaviour (e.g. individual preferences). • Main cost factor: most significant expense factor with growing sample size. In several cases the investigation of a topic by applying one particular empirical method may not be sufficient, as each one implies specific limitations. Hence, there have been several approaches combining two or more methods in order to overcome their drawbacks and maximise their benefits. Early examples can be found in [19] using video and behavioural mapping techniques for analysing city tourists’ behaviour, and in [25] combining unobtrusive tracking methods and inquiries to analyse urban tourism. More recently, localisation technologies have been applied in projects aiming at investigating mobility patterns, e.g. for the development of activity-based transportation models by collecting data with the help of GPS enhanced self-administered diaries recorded on PDAs [22]. With the rapidly growing technical advances in the field of localisation technologies, further developments of more accurate, more reliable and cheaper technologies can be expected. Still, especially regarding the analysis and interpretation of particular datasets, many problems remain unsolved and achieving profound and comprehensive knowledge about human spatio-temporal behaviour is still challenging.

References [1] [2] [3]

E.R. Babbie. The Practice of Social Research. Thomson/Wadsworth, 2006. D. Bauer, N. Br¨andle, and S. Seer. Finding Highly Frequented Paths in Video Sequences. In Proc. 18th International Conference on Pattern Recognition (ICPR2006), Hongkong, 2006. D. Bauer and K. Kitazawa. Using laser scanner data to calibrate certain aspects of microscopic pedestrian motion models. In Proceedings 4th Intl. Conf. on Pedestrian and Evacuation Dynamics (PED2008), Wuppertal, Germany., 2008.

[4]

[5] [6] [7] [8] [9]

[10] [11] [12] [13] [14] [15] [16] [17] [18] [19] [20] [21]

[22] [23] [24] [25] [26]

[27] [28] [29]

[30]

D. Bauer, M. Ray, N. Br¨andle, and H. Schrom-Feiertag. On Extracting Commuter Information from GPS Motion Data. In Proceedings International Workshop on Computational Transportation Science (IWCTS08), 2008. R. Bechtel. Environmental Psychology, chapter Human movement in architecture, pages 642–645. New York: Holt, Rinehart & Winston, 1967. C. Beleznai, B. Fr¨uhst¨uck, and H. Bischof. Human Tracking by Mode Seeking. In Proc. 4th International Symposium on Image and Signal Processing an Analysis, 2005. S. Blivice. Pedestrian Route Choice: A Study of Walking to Work in Munich. PhD thesis, University of Michigan, 1974. W. Bohte, K. Maat, and W. Quak. Urbanism on Track, chapter A Method for Deriving Trip Destinations and Modes for GPS-based Travel Surveys, pages 129–145. IOS Press Amsterdam, 2008. N. Br¨andle, D. Bauer, and S. Seer. Track-based Finding of Stopping Pedestrians – A Practical Approach for Analyzing a Public Infrastructure. In Proceedings 9th IEEE International Conference on Intelligent Transportation Systems (ITSC2006), Toronto, Canada, 2006. N. Caceres, J.P. Wideberg, and F.G. Benitez. Review of traffic data estimations extracted from cellular networks. IET Intelligent Transport Systems, 2:179–192, 2008. P. Corbetta and B. Patrick. Social Research: Theory, Methods and Techniques. SAGE, 2003. P. de Bois and R. de Haan. GPS experiment in Almere. to be published 2009. N.K. Denzin and Y.S. Lincoln. The SAGE Handbook of Qualitative Research. SAGE, 2005. H. Esser. Befragtenverhalten als ”rationales Handeln” – Zur Erkl¨arung von Antwortverzerrungen in Interviews. Arbeitsbericht 85/01, ZUMA, 1985. U. Furbach, M. Marun, and K. Read. Street Level Desires, chapter Information systems for Spatial Metro, pages 74–79. Urbanism: Delft, 2008. B. Glover and H. Bhatt. RFID Essentials. O’Reilly, 2006. N. Haering, P.L. Venetianer, and A. Lipton. The evolution of video surveillance: an overview. Machine Vision and Applications, 19:279–290, 2008. R. Hariharan and K. Toyama. Project Lachesis: Parsing and Modeling Location Histories. In Proceedings of the Third International Conference on GIScience, Adelphi, MD, USA, 2004. R. Hartmann. Combining Field Methods in Tourism Research. Annals of Tourism Research, 15:88–105, 1988. M. Hill. Stalking the Urban Pedestrian: A Comparison of Questionnaire and Tracking Methodologies for Behavioral Mapping in Large-Scale Environments. Environment and Behavior, 16:539–550, 1984. H.H. Hovgesen, T.A.S. Nielsen, P. Bro, and N. Tradisauskas. Urbanism on Track, chapter Experiences from GPS tracking of visitors in Public Parks in Denmark based on GPS technologies, pages 65–77. IOS Press Amsterdam, 2008. D. Janssens, E. Hannes, and G. Wets. Urbanism on Track, chapter Tracking Down the Effects of Travel Demand Policies, pages 147–159. IOS Press Amsterdam, 2008. T. Kanda, M. Shiomi, L. Perrin, and H. Hagita H. Ishiguro. Analysis of People Trajectories with Ubiquitous Sensors in a Science Museum. In Proc. IEEE Int. Conf. on Robotics and Automation, 2007. M.D. Karunanayake, M.E. Cannon, and G. Lachapelle. Evaluation of Assisted GPS (AGPS) in Weak Signal Environments Using a Hardware Simulator. In ION GNSS, 2004. A. Keul and A. K¨uhberger. Tracking the Salzburg Tourist. Annals of Tourism Research, 24:1008–1024, 1997. F. Kleijer, D. Odijk, and E. Verbree. Location Based Services and TeleCartography II, chapter Prediction of GNSS Availability and Accuracy in Urban Environments Case Study Schiphol Airpor, pages 387– 406. Springer: Berlin, Heidelberg, 2009. H. Kuusniemi and G. Lachapelle. GNSS Signal Reliability Testing in Urban and Indoor Environments. In ION NTM, 2004. A. LaMarca and E. DeLara. Location systems: An introduction to the technology behind location awareness. Synthesis Lectures on Mobile and Pervasive Computing, 4:1–122, 2008. A.J. Lipton, J.I.W. Clark, B. Thompson, G. Myers, S.R. Titur, Z. Zhang, and P.L. Venetianer. The Intelligent Vision Sensor: Turning Video into Information. In Proceedings IEEE International Conference on Advanced Video and Signal based Surveillance (AVSS2007), 2007. A. Millonig and G. Gartner. Shadowing - Tracking - Interviewing: How to Explore Human SpatioTemporal Behaviour Patterns. In Technical Report 48: Workshop on Behaviour Monitoring and Interpretation, pages 1–15, 2008.

[31] [32] [33] [34] [35] [36] [37] [38]

[39] [40] [41] [42] [43] [44]

[45] [46] [47] [48] [49] [50]

[51]

[52] [53] [54] [55] [56]

[57]

B.T. Morris and M.M. Trivedi. A Survey of Vision-Based Trajectory Learning and Analysis for Surveillance. IEEE Transactions on Circuits and Systems for Video Technology, 18(8):1114–1127, 2008. S. Munder and D.M. Gavrila. An Experimental Study on Pedestrian Classification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 28(3):1–6, 2006. T.S. Nielsen and H.H. Hovgesen. GPS in pedestrian and spatial behaviour surveys. In Conference Proceedings 5th International Conference on Walking in the 21st Century: Cities for People, 2004. R.E. Nisbett and T.D. Wilson. Telling more than We can Know: Verbal Reports on Mental Processes. Psychological Review, 84:231–259, 1977. I. Pavlidis, V. Morellas, P. Tsiamyrtzis, and S. Harp. Urban Surveillance Systems: From the Laboratory to the Commercial World. Proceedings of the IEEE, 89(10):1478–1497, 2001. R. Poppe. Vision-based human motion analysis: an overview. Computer Vision and Image Understanding, 108:4–18, 2007. J. Raper, G. Gartner, H. Karimi, and C. Rizos. A critical evaluation of location based services and their potential. Journal of Location Based Services, 1:5–45, 2007. M. Ray and H. Schrom-Feiertag. Cell-based Finding and Classification of Prominent Places of Mobile Phone Users. In The 4th International Symposium on Location Based Services & TeleCartography, 2007. P. Remagnino, S.A. Velastin, G.L. Foresti, and M. Trivedi. Novel concepts and challenges for the next generation of video surveillance systems. Machine Vision and Applications, 18:135–137, 2007. B. Rinner and W. Wolf. An Introduction to Distributed Smart Cameras. Proceedings of the IEEE, 96(10):1565–1575, 2008. S. Schneegans, P. Vorst, and A. Zell. Using RFID Snapshots for Mobile Robot Self-Localization. In 3rd European Conference on Mobile Robots (ECMR 2007), 2007. H. Schrom-Feiertag and M. Ray. Learning Interests for a Personalised City Tour based on Motion Behaviour. In The 4th International Symposium on Location Based Services & TeleCartography, 2007. M. Shah, O. Javed, and K. Shafique. Automated Visual Surveillance in Realistic Scenarios. IEEE Multimedia, 14(1):30–39, 2007. X. Shao, H. Zhao, K. Nakamura, K. Katabira, R. Shibasaki, and Y. Nakagawa. Detection and tracking of multiple pedestrians by using laser range scanners. In IEEE International Conference on Intelligent Robots and Systems, 2007. N. Shoval. Tracking technologies and urban analysis. J.Cities, 25(1):21–28, 2008. N. Shoval and M. Isaacson. Tracking tourists in the digital age. Tracking tourists in the digital age, 34(1):141–159, 2007. B. W. Silverman. Density Estimation for Statistics and Data Analysis. Chapman and Hall, London, 1986. D. Singel´ee and B. Preneel. Enabling Location Privacy in Wireless Personal Area Networks. Cosic internal report, Katholieke Universiteit Leuven, 2007. skiline marketing gmbh. Skiline. http://skiline.cc/. last accessed 2009/04/02. A. Steinhage and C. Lauterbach. Monitoring Movement Behaviour by means of a Large Area Proximity Sensor Array in the Floor. In B. Gottfried and H. Aghajan, editors, 2nd Workshop on Behaviour Monitoring and Interpretation (BMI08), pages 15–27. CEURS Proceedings, 2008. J. Suutala and J. Rning. Towards the Adaptive Identification of Walkers: Automated Feature Selection of Footsteps Using Distinction-Sensitive LVQ. In Int. Workshop on Processing Sensory Information for Proactive Systems (PSIPS 2004), 2004. P. Thornton, A. Williams, and W.G. Shaw. Revisiting Time-Space Diaries: An Exploratory Case Study of Tourist Behavior in Cornwall, England. Environment and Planning A, 29:18471867, 1997. E. Trevisani and A. Vitaletti. Cell-ID technique, limits and benefits: an experimental study. In Sixth IEEE Workshop on Mobile Computing Systems and Applicaitons, 2004. S.C. van der Spek. Urbanism on Track, chapter Tracking Technologies, pages 25–33. IOS Press Amsterdam, 2008. S.C. van der Spek. Urbanism on Track, chapter Spatial Metro: Tracking pedestrians in historic city centres, pages 79–101. IOS Press Amsterdam, 2008. P. Vorst, J. Sommer, C. Hoene, P. Schneider, C. Weiss, T. Schairer, W. Rosenstiel, A. Zell, and G. Carle. Indoor Positioning via Three Different RF Technologies. In 4th European Workshop on RFID Systems and Technologies, 2008. R.S. Weiss and S. Boutourline Jr. A summary of fairs, pavilions, exhibits, and their audiences. Technical

[58]

[59]

[60] [61] [62] [63] [64]

report, Cambridge, Mass., 1962. S. Wiesenhofer, H. Feiertag, M. Ray, L. Paletta, P. Luley, A. Almer, M. Schardt, J. Ringert, and P. Beyer. Mobile City Explorer: An innovative GPS and Camera Phone Based Travel Assistant for City Tourists, pages 557–573. Lecture Notes in Geoinformation and Cartography. Springer Berlin Heidelberg, 2007. A. Willis, C. Gjersoe, N.and Havard, J. Kerridge, and R. Kukla. Human movement behaviour in urban spaces: implications for the design and modelling of effective pedestrian environments. Environment and Planning B: Planning and Design, 31:805–828, 2004. G. Winkel and R. Sasanoff. Environmental Psychology, chapter An approach to an objective analysis of behavior in architectural space, pages 619–631. New York: Holt, Rinehart & Winston, 1966. A. Yilmaz, O. Javed, and M. Shah. Object Tracking – A Survey. ACM Computing Surveys, 38(4):pp 1–45, 2006. V. Zeimpekis, G.M. Giaklis, and G. Lekakos. A Taxonomy of Indoor and Outdoor Positioning Techniques for Mobile Location Services. ACM, 2003. B. Zhan, D.N. Monekosso, P. Remagnino, S.A. Velastin, and L. Xu. Crowd analysis: a survey. Machine Vision and Applications, 19:345–357, 2008. H. Zhao and R. Shibasaki. A novel system for tracking pedestrians using multiple single-row laserrange scanners. IEEE Transactions on Systems, Man, and Cybernetics Part A:Systems and Humans, 35:283–291, 2005.