Unsupervised Driving Performance Assessment Using ... - IEEE Xplore

6 downloads 16949 Views 608KB Size Report
Abstract— In-vehicle smartphones contain rich information of driving and therefore can be utilized to establish driver behavior profile. The GPS data indicates the ...
2016 IEEE 19th International Conference on Intelligent Transportation Systems (ITSC) Windsor Oceanico Hotel, Rio de Janeiro, Brazil, November 1-4, 2016

Unsupervised Driving Performance Assessment using FreePositioned Smartphones in Vehicles Yang Zheng, Student Member, IEEE, and John H.L. Hansen, Fellow, IEEE 

Abstract— In-vehicle smartphones contain rich information of driving and therefore can be utilized to establish driver behavior profile. The GPS data indicates the occurrence of driving events like gas-hit, brake-hit, turn-right, turn-left, and forward-driving. The inertial measurement unit (IMU) data, provides good estimations of vehicle lateral, longitudinal, and vertical accelerations and rotations. One of the challenge using the smartphone as the estimator is the variety of device poses. This study proposes a coordinate transformation approach to address this task, which allows smartphones to be freepositioned in vehicles. With the converted vehicular 3-axis accelerations, this study also employs unsupervised clustering techniques to grade driving events by iteratively detect outliers. Discussion on risky events is also conducted.

I. INTRODUCTION Towards achieving the “intelligent vehicle” which allows the vehicle to understand its driver and therefore providing human-centered assistance, establishing “driver behavior profile” has been becoming a hot topic in naturalistic driving studies. One popular application comes from the auto insurance market (e.g., Drive Safe & SaveTM [1] and Ingenie [2]). Typically, a data-logging unit should be plugged into the vehicle’s OBD-II port, the CAN-Bus data will then be retrieved and wirelessly transmitted to the remote server via Telematics. The CAN-Bus data contains rich information for tracking driver behavior, and later the driver will see a score of his driving performance or an assessment report for specific driving events (e.g., acceleration, deceleration, left-turn, and right-turn). This idea makes it possible for the driver as well as his vehicle to understand their performance, but the main drawback is low customer acceptance and hence limit the platform deployment. On the other hand, the proliferation of smartphone application has made a great impact in the automotive industry. Besides infotainment and telecommunication functionalities, smartphones contain a variety of useful sensors such as cameras, microphones, as well as their Inertial Measurement Units (IMU) such as accelerometer, gyroscope, and GPS. These multi-channel signals would also be synchronized together, providing a comprehensive description of driving scenarios. Therefore, the smartphone could potentially be leveraged as a cost effective approach for invehicle data collection, monitoring, and added safety options/feedback strategies. It lowers the entry barrier and allows for a wider range of naturalistic data collection opportunities for vehicles and driver’s operating their own vehicles. But compared to the CAN-Bus signal, smartphone

Yang Zheng is with the Center for Robust Speech System (CRSS) – UTDrive Lab in the Electrical Engineering Department, University of Texas at Dallas, Richardson, TX 75080 USA (e-mail: yxz131331@ utdallas.edu). 978-1-5090-1889-5/16/$31.00 ©2016 IEEE

IMU data is highly incorporated with vibration noise, which requires to be filtered beforehand. The problem being focused in this study is to assess driving behavior using the data retrieved from in-vehicle smartphone IMU and GPS sensors. Specifically, to grade the performance on the five actions: left-turn, right-turn, gas-hit, brake-hit, and normal forward-driving. In classical naturalistic driving studies, driving scenarios are manually annotated and subjectively evaluated by reviewing videos, which requires huge amount of human effort. However, an appropriate processing of smartphone GPS and IMU data can effectively reduce the human labor by automatically recognizing the five events above. Furthermore, this study proposed a completely unsupervised clustering method to grade the performance into A, B, C, and D four levels, as well as indicate risky driving moments by detecting cluster outliers. In the processing of in-vehicle smartphone data, the most challenging part comes from the smartphone pose. Since the goal is to expand the platform deployment, it is important to consider any position and any orientation of the smartphone placed in vehicle. To overcome this difficulty, this study suggests a coordination transformation technique to convert smartphone IMU accelerations to the corresponding vehicle accelerations, and then align with gravity and GPS data to specify the vertical, lateral, and longitudinal movement of the vehicle. The remaining part of this paper is organized as follows. Section II will briefly introduce the related work from literature and our previous studies. Section III will highlight a brief overview of the system. Section IV and Section V will describe in details on the proposed methods for data processing and driving performance assessment. Section VI will present the experiment result and discussion. Section VII will conclude the paper with a summary of contributions and future work. II. RELATED WORK In the literature, smartphone-based sensing in vehicles have been widely discussed, and can be categorized as the following four types of applications [3]:   

Traffic information, such as the location and movements of other vehicles or pedestrians. Vehicle information, such as vehicle diagnostics. Environmental information, such as road conditions and weather conditions.

John H.L. Hansen is the director of CRSS-UTDrive Lab in the Electrical Engineering Department, University of Texas at Dallas, Richardson, TX 75080 USA (e-mail: john.hansen@ utdallas.edu).

1598

IMU: 3-axis Acceleration 3-axis Gyroscope GPS: Latitude, Longitude Speed, Bearing

Data Processing

Driving Assessment

Filtering

Feature Extraction

Coordinate Transformation

Event Identification

Report: Grade, Score

Clustering – Outlier Detection

Axis Alignment

Figure 1. Overall procedure flow diagram



Driver behavior information, such as insurance telematics.

For the studies on driver behavior aspect, You et al. [4] proposed a smartphone App – CarSafe to alert drowsy and distraction with the utilization of dual cameras. Similarly, Bergasa [5] published another App – DriveSafe for alerting inattention and scoring driving behavior, based on computer version as well. Instead of video stream processing, Johnson and Trivedi [6] developed a system which can detect and classify aggressive versus non-aggressive driving maneuvers, by only using the smartphone IMU and GPS data. Eren et al. [7] used a similar approach based on the same algorithms, but expanded their system by adding a driving style classification feature. Fazeen et al. [8] contributed a driver behavior monitoring system that advises a driver on dangerous vehicle maneuvers. However, all of these studies assume the device should be placed at a fixed position, which becomes a common limitation of their systems. Taking the consideration of in-vehicle smartphone pose variety, Dai et al. [9] suggested an IMU-based calibration procedure to obtain pitch angle and roll angle of the smartphone relative to the vehicle in order to detect drunk driving. However, they assume that the smartphone is aligned with the longitudinal axis of the vehicle, so the main restriction of this system is that the yaw angle must be null. To compensate, Almazan et al. [10] attempted a full autocalibration method to estimate the yaw angle of a smartphone relative to a vehicle in every case. Castignani et al. [11] proposed a fuzzy system for driver behavior monitoring, which enables the distinction between calm and aggressive driving to be made. In his study, the magnitude of acceleration vector is computed to mitigate the problem of decomposing vehicle’s longitudinal and lateral movements. In our previous studies since 2012, an Android App – MobileUTDrive [12] has been designed and continuously updated for the in-vehicle data collection and driving monitoring. It has been shown that the yaw of a portable device gyroscope is closely related with the steering wheel angle extracted from the CAN-Bus [13], therefore providing a cost effective and widely applicable approach for driving behavior analysis [14]. Using this approach has resulted in several studies to detect maneuvers, and design driving safety systems that combine in-vehicle speech and video analysis and driving performance evaluation [15, 16]. However, most of the previous work employed supervised machine-learning algorithms, which require large amount of time for human annotation beforehand. This study will move forward with unsupervised clustering techniques for the driving behavior analysis.

III. SYSTEM OVERVIEW Fig. 1 illustrates the overall working procedure presented in this study. The input signals are given as the IMU 3-axis acceleration and 3-axis gyroscope value relative to the smartphone, as well as GPS information including latitude and longitude coordinates, and current moving speed and bearing angle (North based). There are two main modules developed within system – data processing module, and driving assessment module. In the “Data Processing” module, the first step is to filter out the noise of raw input signals. It then performs coordinate transformation to convert the smartphone centered X, Y, Z – accelerations to vehicle centered accelerations, as well as convert the global-referenced GPS bearing directions to vehicle-referenced moving angles. So far, it has not yet been decided which axis is corresponded with the vehicle lateral, longitudinal, or vertical movement. An additional alignment step is required to assign the axis, by referring to gravity and GPS information. In the “Driving Assessment” module, a long sequence driving data is first framed by a fixed time window, and a set of features are extracted from each frame. These features are sufficient to recognize the four typical driving events – leftturn, right-turn, gas-hit, brake-hit, and the remaining sequences are indicated as forward-moving. An unsupervised clustering method is applied on each event, and based on the result of outlier detection, each event is graded with A, B, C, and D four levels. The final output will be an overall driving performance report for this driver. IV. DATA PROCESSING A. Pre-processing Due to the road surface vibration, the raw acceleration and gyroscope signals are badly infected with random noise. Therefore, the first step of data pre-processing should be noise reduction. A median filter is employed. The original GPS data is typically sampled at 1Hz, but the frequency of smartphone IMU data may vary from 5-50 Hz, depending on the hardware and developers setting. To be consistent, both the IMU and GPS data is re-sampled at a uniform rate – 10 Hz, utilizing linear interpolation. The GPS data contains valuable information like current moving speed and bearing angle, therefore, it is straightforward to take their derivative values as indicates of vehicle longitudinal and rotational accelerations. The GPS bearing angle is North-based ranging from 0 to 360 degree, if a vehicle

1599

Figure 2. Typical situations when a smartphone is placed in vehicle

is heading to the North, a small change in direction will cause a “big jump” between 0 and 359 of the bearing output, which will result with an undesired derivative value. Therefore, it is important to unwrap the radian phase to compensate the discontinuity. B. Coordinate Transformation The smartphone IMU data can be an effective estimate of vehicle dynamics. However, the orientation and relative movement of the smartphone inside the vehicle yields the main challenge for platform deployment. Fig. 2 depicts three typical positioning scenarios. First, the smartphone device could be mounted at a fixed position; a low-level calibration is needed to compensate for the mounting variance. In the second case, the device is stationary and sitting in the vehicle, but its orientation is unknown. The phone-referenced sensors reading should therefore be converted to the car-referenced dynamic outputs. In case 3, hand-held or any freestyle situation, the relative movement of the device inside the vehicle should be decoupled if it exists. This study proposes a coordinate transformation technique that can be applied for the first two scenarios where no relative movement exists.

ACCP  X P YP

ZP 

GYROP  

 

YC

(1)

T

(3)

Coordinate transformation is considered as a combination of 3-step rotations [18], resulting with three transformation matrices.

0  X C  1  Y   0 cos   C   Z C  0  sin 

sin  cos  0

0  X P  XP     0  YP   [T1 ] YP   Z P  1  Z P 

(5)

(4)

0   x' '  x' '    sin    y ' '  [T3 ] y ' '  z ' '  cos    z ' ' 

(6)

Finally, the complete coordinate transformation is computed by XC  XP   Y   [T ][T ][T ] Y  3 2 1  P   C  Z C   Z P 

 Step 1: rotate about Z-axis.  x'  cos   y '   sin      z '   0

0 sin    x'  x'    1 0   y '  [T2 ] y '  z '  0 cos    z ' 

 Step 3: rotate about X-axis. (2)

ZC 

(d)

 x' '  cos   y ' '   0     z ' '   sin 

The goal is to compute the car-referenced accelerations, denoted as ACCC  X C

(c)

 Step 2: rotate about Y-axis.

as well as the gyroscope ration angles along the 3-axis T

(b)

Figure 3. Car-referenced coordinate system (a,c) versus Phonereferenced coordinate system (b,d), where Xc, Yc, Zc denote the vehicle lateral, longitudinal, and vertical movement, and Xp, Yp, Zp denote the phone movements.

As depicted in Fig. 3 [17], the input signal retrieved from smartphone IMU contains the phone-referenced 3-axis accelerations, given as T

(a)

(7)

It is worthwhile to note that the final transformation result depends on the order of rotations. In practice, however, the variation is small. To be consistent, Eq. (7) is taken as the unique form, and then a median filter is applied for data smoothing.

1600

2nd dimension of Principle Component Analysis

C. Axis Alignment If the smartphone pose in vehicle is known as a prior knowledge, the step of coordinate transformation can be regarded as a calibration procedure to compensate the small variations of device misplacement. However, if the pose is unknown, it is still not able to determine which axis is corresponded with vehicle’s longitudinal, lateral, or vertical movement. Therefore, an additional axis alignment step is needed. In general, the vehicle vertical acceleration should be close to the gravity (i.e., g  9.8m/s). Comparing the normalized values of ACCC and GPS speed and bearing derivatives, to find the axis indices becomes a problem of minimizing the cost function of mean square errors (MSE), given as longi

k

axis  arg min lati

k

1 N 1  [( ACCC {k}(i)  t GPSspeed (i))2 ] N i

(8)

1 N 1  [( ACCC {k}(i)  t GPSbearing(i))2 ] N i

(9)

axis  arg min verti

k

1 N 1 [( ACCC {k}(i)  g )2 ] N i

(10)

where N, i=0,1, …, N-1 denotes the data length of a given sequence, and ACCC {k}, k  0,1,2 denote the accelerations

X C

YC

T ZC  resulted from coordinate transformation.

1st dimension of Principle Component Analysis

Figure 4. Scatter plot of (a) OneClassSVM and (b) TAD cluster output, visualization on the first two PCA dimensions. Grade A samples are closest to the cluster centroid, and represent the best driving events. Grade D samples are the outermost anomalies representing the least rated events.

V. DRIVING ASSESSMENT

A. Feature Extraction A long sequence of driving data is first segmented into frames, with the fixed window length of 1 second. Within each frame of transformed IMU accelerations (i.e., ACCC), a group of feature set should be extracted to represent the vehicle dynamic movements, which are sufficient to reflect the driver behavior performance at this moment. Referring to the suggestions of [19], we select three time-domain features (mean, variance, and mean-crossing radio) and four frequency-domain features (peak frequency, spectral energy, entropy, and correlation) for each axis, resulting with a 21dimensional feature space. Unsupervised feature selection

Grade A Grade B Grade C Grade D

(b) Topological Anomaly Detection (TAD) Cluster

So far, the in-vehicle smartphone signals after the “Data Processing” module should become good estimates of vehicle kinematics, regardless of the variety of device poses.

One convenient approach for driving assessment is to identify the “context”, and then quantify if there is any deviation from the expected or “neutral” behavior. Theoretically, the good, safe, or convenient driving behavior should be reflected with the stable, steady, or low-variance of vehicular dynamical performance. This underlies the first assumption of driving behavior assessment. In addition, an experienced driver should act as “good driving” for most of the time, whereas “bad driving” may occur at a limited number of moments, which may be caused by visually or cognitively inattention, unfamiliar with the route, distracted by secondary tasks, and so on. This underlies the second assumption in this section.

1st dimension of Principle Component Analysis

(a) One-Class SVM Cluster 2nd dimension of Principle Component Analysis

axis  arg min

Grade A Grade B Grade C Grade D

techniques are also considered for dimension reduction [20, 21, 22]. B. Event Identification In contrast with the IMU accelerations, the GPS data is only adopted for event identification. The peaks/valleys of GPS speed derivative indicate the occurrences of gashit/brake-hit events, whereas the unwrapped GPS bearing derivative indicate the right-turn/left-turn events. The remaining part of the frames are considered as normal forward-driving events. Therefore, a long duration driving data is constructed with a sequence of small blocks, and each block is categorized within the five classes – gas-hit, brake-hit, right-turn, left-turn, and forward-driving. The next step will establish five clusters on these classes, and the distance from each individual event to the cluster centroid will be measured for driving assessment. C. Clustering – Outlier Detection Based on the assumption of experienced drivers, most of their driving time should be labeled as “good” behavior and “bad” behavior only occur at a small number of moments. If the extracted feature set is sufficient to represent vehicle dynamic movements, which are corresponding with driving performance, there should be common characteristics for the “good driving” events, whereas “bad driving” events will turn

1601

out to be the outliers. Therefore, manual or subjective annotation is not a required condition, and the unsupervised clustering method is proposed.

grade formula, so as to provide an overall score of the driving performance. Table I summarizes the six drivers’ final GPA score and frequency of risky events for the five typical driving

To generate a stratified output, a cluster scheme is applied iteratively for three times, resulting in three layers of outliers and one inner centroid layer. Each driving event sample is labeled with a grade from “A” to “D”, corresponding with the layers from innermost to outermost. The gradient of driving performance therefore forms a range from “good” to “bad”, and is therefore consistent with the grades from “A” to “D”. Illustrated in Fig. 4, this task employs two unsupervised clustering models – One-Class SVM [23] and Topological Anomaly Detection (TAD) [25]. One-Class SVM captures the shape of dataset, and separates non-linear scaled layers after iteration. TAD computes the distance between nearest neighbors, to generate the layers, the graph resolution is iteratively reduced in a linear scale.

(a) One-Class SVM Result

VI. EXPERIMENT RESULT For the naturalistic data collection, an experiment route is designed in the local district around the UT-Dallas campus. Each complete driving route requires roughly 15-18 minutes driving at the designed speed limit 40 mph, depending on traffic and stop light conditions. The test data for the current evaluation is collected with 6 adult drivers, each contributing 45-60 minutes of driving. All these adult drivers have +5 years of driving experience and have general familiarity of the roads and traffic conditions surrounding the 3.6 miles path outside of UTDallas campus.

(b) Topological Anomaly Detection (TAD) Result Figure 5. Percentage of driving event samples in each grade clustered by (a) One-Class SVM and (b) TAD. BRK=brake-hit, GAS=gas-hit, RTR=right-turn, LTR=left-turn, FWD=forward-driving. Risky Moments (a) Stop-signs and Pedestrians

During data collection, the smartphone is mounted against a vehicle front windshield (i.e., the device front camera faces the driver, while the back camera faces the road). This is intended to capture video streams and reserve for human annotation or subjective analysis. Regarding the verification of coordinate transformation for various device positions, it requires to be done in a separate study by comparing with other data source like CAN-Bus. It is therefore beyond the scope of this study, just a small extent of calibration is needed. Grading results for driving assessment depend on the cluster algorithm. As shown in Fig. 5, One-Class SVM maintains a large number of samples in grade “A” for all the five events, whereas TAD results a variant grade distribution – only a few “bad” in FWD but more “not good” or “bad” in other events. However, for the grade “D” groups resulted from the two models, the intersection part can be concluded as “well-accepted” anomalies. These samples are therefore considered as the riskiest events with more confidence. Fig. 6 highlights the risks occurrence on a test drive route. In area (a), several stop-signs exist and pedestrians pass by, so the driver needs to make frequent brake/gas hits. In area (b), the speed limit is low and traffic flow is reduced, so the vehicle can keep in a stable state and little or no risks. However, the speed limit and traffic flow increases a lot in area (c), which causes the occurrence of greater risks there. It can be inferred that risks have higher possibility to occur in the switching period between high-speed and low-speed. For each driver, all their driving events are assigned with the grades “A”, “B”, “C”, and “D”. It is interesting to calculate the “Grade Point Average (GPA)” using the standard college

(b) Low speed, Little traffic

(c) Speed limit increase, Traffic flow increase

Figure 6. Highlight of risky moments (red) on a testing round (green). Area (a) contains several risks because of two many stop signs and pedestrians; area (b) contains no risks because of low speed and little traffic; area (c) contains several risks because speed limit and traffic flow increase.

events. Specifically, for Driver #1, his auto insurance company – State Farm – reports a similar statement to be compared. The assessment grades in this study appear to be higher than the insurance report, it is probably because the insurance company keeps greater restrictions in their evaluation process. VII. CONCLUSION Smartphone IMU sensing provides a cost-effective data acquisition platform to estimate vehicle kinetics, which expand the deployment for naturalistic driving study. The existing main challenge is to overcome the variety of device poses. To address this problem, this study contributes a coordinate transformation approach to convert the devicereferenced 3-axis accelerations into the vehicle-referenced frame. This approach allows smartphones to be freepositioned in vehicle.

1602

TABLE I.

DRIVING ASSESSMENT GPA REPORT

Driver#1 Insurance Report

Driver#1

Driver#2

Driver#3

Driver#4

Driver#5

Driver#6

Forward-Driving Left-Turn Right-Turn Gas-Hit Brake-Hit

BBB+ B-

3.437, B+ 3.337, B+ 3.444, B+ 3.462, B+ 3.375, B+

3.44, B+ 2.667, C+ 2.625, C+ 3.4, B+ 3.565, B+

3.435, B+ 3.542, B+ 3.25, B 3.511, B+ 3.407, B+

3.436, B+ 3.423, B+ 3.25, B 3.469, B+ 3.42, B+

3.438, B+ 3.087, B 2.429, C+ 3.5, B+ 3.333, B+

3.439, B+ 3.56, B+ 2.333, C+ 3.571, B+ 3.464, B+

Risky Frequency

-

2.0758%

3.0025%

1.5200%

3.3814%

1.2652%

2.1220%

This paper assessed the driving performance for the five typical driving events– right-turn, left-turn, gas-hit, brake-hit, and forward-driving. Based on the assumption that an experienced should perform “good driving” for most of the time which should have characteristics in common, each event can be graded into four levels, by iteratively applying unsupervised clustering techniques to detect outliers. Moreover, the intersection of outermost anomalies obtained from different clusters can be confidently recognized as “wellaccepted” risky events. This unsupervised and objective driving assessment approach will effectively reduce the human works in processing huge amount of naturalistic driving data. In the future, it is worthwhile to continue validate the effectiveness of IMU coordinate transformation, with the examination on a protocol set of device positions, as well as the comparison of other data sources. In addition, with the accessible context information from the smartphone, it is possible to analysis the effect of weather and daylight to driving performance. Besides, it is also considered to build the clustering models within smartphone, and perform real-time in-device driving assessment. Or alternatively, establish “cloud computation” with the help of telematics.

[10]

[11]

[12]

[13]

[14]

[15]

[16]

REFERENCES [1] [2] [3]

[4]

[5]

[6]

[7] [8] [9]

State Farm Auto Insurance, Drive Safe & Save discount program, https://www.statefarm.com/insurance/auto/discounts/drive-safe-save/. Retrieved May 26, 2016. Ingenie Services Limited., Black box car insurance for young drivers, (2012) http://www.ingenie.com. Retrieved May 26, 2016. Engelbrecht, J., Booysen, M. J., van Rooyen, G. J., & Bruwer, F. J., "Survey of smartphone-based sensing in vehicles for intelligent transportation system applications." Intelligent Transport Systems, IET 9.10 (2015): 924-935. You, C. W., Lane, N. D., Chen, F., Wang, R., Chen, Z., Bao, T. J., ... & Campbell, A. T., "Carsafe app: Alerting drowsy and distracted drivers using dual cameras on smartphones." Proceeding of the 11th annual international conference on Mobile systems, applications, and services. ACM, 2013. Bergasa, L. M., Almería, D., Almazan, J., Yebes, J. J., & Arroyo, R., "Drivesafe: an app for alerting inattentive drivers and scoring driving behaviors." Intelligent Vehicles Symposium Proceedings, 2014 IEEE. IEEE, 2014. Johnson, Derick A., and Mohan M. Trivedi. "Driving style recognition using a smartphone as a sensor platform." Intelligent Transportation Systems (ITSC), 2011 14th International IEEE Conference on. IEEE, 2011. Eren, H., Makinist, S., Akin, E., & Yilmaz, A., "Estimating driving behavior by a smartphone." Intelligent Vehicles Symposium (IV), 2012 IEEE. IEEE, 2012. Fazeen, M., Gozick, B., Dantu, R., Bhukhiya, M., & González, M. C., "Safe driving using mobile phones." Intelligent Transportation Systems, IEEE Transactions on 13.3 (2012): 1462-1468. Dai, J., Teng, J., Bai, X., Shen, Z., & Xuan, D., "Mobile phone based drunk driving detection." Pervasive Computing Technologies for

[17]

[18] [19]

[20] [21] [22]

[23]

[24]

[25]

1603

Healthcare (PervasiveHealth), 2010 4th International Conference onNO PERMISSIONS. IEEE, 2010. Almazan, J., Bergasa, L. M., Yebes, J. J., Barea, R., & Arroyo, R., "Full auto-calibration of a smartphone on board a vehicle using IMU and GPS embedded sensors." Intelligent Vehicles Symposium (IV), 2013 IEEE. IEEE, 2013. Castignani, G., Derrmann, T., Frank, R., & Engel, T., "Driver behavior profiling using smartphones: A low-cost platform for driver monitoring." Intelligent Transportation Systems Magazine, IEEE 7.1 (2015): 91-102. Zheng, Y., and Hansen, J.H.L., “MobileUTDrive: An Android Portable Device Platform for In-vehicle Driving Data Collection and Display”, in Proc. 3rd International Symposium on FAST-zero, Gothenburg, Sweden, September 2015. Sathyanarayana, A., Sadjadi, S. O., & Hansen, J. H.L., "Leveraging sensor information from portable devices towards automatic driving maneuver recognition." Intelligent Transportation Systems (ITSC), 2012 15th International IEEE Conference on. IEEE, 2012. Sathyanarayana, A., Sadjadi, S. O., & Hansen, J.H.L., "Automatic Driving Maneuver Recognition and Analysis using Cost Effective Portable Devices." SAE International Journal of Passenger CarsElectronic and Electrical Systems 6.2 (2013): 467-477. Zheng, Y., Shi, X., Sathyanarayana, A., Shokouhi, N., & Hansen, J. H.L., "In-vehicle speech recognition and tutorial keywords spotting for novice drivers' performance evaluation." Intelligent Vehicles Symposium (IV), 2015 IEEE. IEEE, 2015. Zheng, Y., Shokouhi, N., Thomsen, N., Sathyanarayana, A., & Hansen, J., Towards Developing a Distraction-Reduced Hands-Off Interactive Driving Experience using Portable Smart Devices. No. 2016-01-0140. SAE Technical Paper, 2016. Wang, Y., Yang, J., Liu, H., Chen, Y., Gruteser, M., & Martin, R. P., "Sensing vehicle dynamics for determining driver phone use." Proceeding of the 11th annual international conference on Mobile systems, applications, and services. ACM, 2013. Widnall, S. "Lecture L3-Vectors, Matrices and Coordinate Transformations." Dynamics (2009). Lu, H., Yang, J., Liu, Z., Lane, N. D., Choudhury, T., & Campbell, A. T., "The Jigsaw continuous sensing engine for mobile phone applications." Proceedings of the 8th ACM conference on embedded networked sensor systems. ACM, 2010. Jolliffe, Ian. Principal component analysis. John Wiley & Sons, Ltd, 2002. Frigui, Hichem, and Raghu Krishnapuram. "Clustering by competitive agglomeration." Pattern recognition 30.7 (1997): 1109-1119. Bingham, Ella, and Heikki Mannila. "Random projection in dimensionality reduction: applications to image and text data." Proceedings of the seventh ACM SIGKDD international conference on Knowledge discovery and data mining. ACM, 2001. Ma, Junshui, and Simon Perkins. "Time-series novelty detection using one-class support vector machines." Neural Networks, 2003. Proceedings of the International Joint Conference on. Vol. 3. IEEE, 2003. Riani, Marco, Anthony C. Atkinson, and Andrea Cerioli. "Finding an unknown number of multivariate outliers." Journal of the Royal Statistical Society: series B (statistical methodology) 71.2 (2009): 447466. Gartley, M. G. "Topological anomaly detection performance with multispectral polarimetric imagery." SPIE Defense, Security, and Sensing. International Society for Optics and Photonics, 2009.