Traffic Density Estimation Using Radar Sensors ...

5 downloads 460 Views 1MB Size Report
Traffic Density Estimation Using Radar Sensor Data from Probe Vehicles. 1 ... Seo and Kusakabe (2015) propose a method to estimate the traffic state (i.e., flow, ...
ITS World Congress 2017 Montreal, October 29 – November 2

Traffic Density Estimation Using Radar Sensor Data from Probe Vehicles Paper ID # [This number is only required when submitting your final paper)

Traffic Density Estimation Using Radar Sensor Data from Probe Vehicles Daisik Nam1, Riju Lavanya1, Inchul Yang2, 3*, Woo Hoon Jeon2, R Jayakrishnan1 1. Department of Civil and Environmental Engineering, University of California, Irvine, USA 2. Highway & Transp. Research Institute, Korea Institute of Civil Eng. and Building Tech, Korea 3. Dept. of ITS at Korea University of Science and Technology, Goyang-Si, Korea Corresponding author: Inchul Yang, E-mail: [email protected]; Phone: +82-31-910-0489; ORCID: 0000-0002-3882-0157;

Abstract High performance traffic management systems require accurate traffic density information since traffic density can be efficiently detected at the onset of congestion. The aim of this research is to develop a new algorithm for traffic density estimation using probe vehicles equipped with various sensors. A radarequipped probe vehicle has multiple sensors at various points. With the combination of high resolution 77 GHz radars, Cameras, Global Positioning System (GPS), digital map aided systems, and telecommunication technologies, probe vehicles can sense current traffic conditions in their vicinity. We develop a methodology to estimate traffic density of a road section by capturing the local densities detected by radar-equipped probe vehicles. Vehicle trajectory datasets provided by NGSIM were used to evaluate our methodology for different market penetration levels of the probe vehicles. The results confirm that the proposed method can estimate traffic density reasonably well even under low market penetration rates.

KEYWORDS: Radar Sensor Data from Probe Vehicles, NGSIM datasets, and Traffic Density

1

ITS World Congress 2017 Montreal, October 29 – November 2

Traffic Density Estimation Using Radar Sensor Data from Probe Vehicles Introduction Traffic congestion imposes tremendous economic and environmental costs on society. In 2013, the direct and indirect costs of congestion in the U.S were estimated to be approximately $124 billion. This number is expected to rise steadily due to factors such as population and GDP growth, decline in fuel prices, and a corresponding increase in car ownership [6]. Relieving traffic congestion has been the primary focus of most traffic management applications over the past few decades. Traffic density is a key control variable in high performance traffic monitoring and management systems since it is a reliable indicator of system performance. Therefore, accurate real-time estimation of traffic density is a prerequisite to the successful deployment of any ITS application aimed at mitigating traffic congestion. The measurement of traffic density has relied on traditional data collection methods such as loopdetectors or photographic techniques, which have spatial or temporal limitations. Pavement-invasive detectors such as traditional loop detectors suffer from high installation and maintenance costs. Furthermore, data obtained from them are not always reliable. New technologies such as vehicle sensing, vehicle tracking, and wireless communication avoid these drawbacks. The concept of ‘Internet of Things’ (IoT) and advances in sensor technologies offer efficient ways to monitor and estimate the state of traffic. Modern cars already come equipped with sensors to assist drivers with functions such as adaptive cruise control, parking, emergency stops, and lane-changing. Such technology is expected to be even more ubiquitous with the advent of autonomous cars. Therefore, it is evident that vehicle sensing technologies and data derived from them are poised to play a key role in the future of ITS applications. The focus of our research is on the real-time estimation of traffic density from data obtained solely from radar-equipped probe vehicles in a traffic network. A radar-equipped probe can capture “local density” by detecting cars in their sensing area. The link traffic density can then be inferred from this local density. In other words, link traffic conditions, which cannot be gauged by point detectors such as loop and video detectors, can now be captured on a continuous road section from moving sensor vehicles. Herring et al., (2010) estimate and predict traffic conditions in arterial networks using probe data. The proposed model for traffic condition estimation was evaluated using data from a fleet of 500 taxis in San Francisco, CA. Seo and Kusakabe (2015) propose a method to estimate the traffic state (i.e., flow, density and speed) based on probe vehicle data. In their study, two probes drive on the same road stretch and the following probe observes the number of vehicles between the leading probe and the following probe. From this data, they calculate the cumulative count at the probe vehicle’s current position and then estimate traffic density. Al-Sobky and Mousa (2016) introduce a similar approach, except that they use smart phones to capture the number of vehicles. To our best knowledge, this paper is one of the first attempts to estimate traffic density by using only radar sensors on vehicles. We rely on the potential for having high penetration rates of radar-equipped vehicles in the near future. Government investment in infrastructure is not always a given, and recent trends indicate that it is likely to only decrease in the future in most countries. Therefore, alternate sources of traffic data such as obtained from probe vehicles become more valuable for transportation researchers. Current research efforts for monitoring traffic have not kept pace with the expected growth in vehiclesensing market. In this study, we develop an algorithm that harnesses this new category of data to estimate traffic densities. The next section provides a general overview of the practical aspects of traffic density estimation.

Traffic Density Estimation Method from Radar Sensors data Specification of a Radar Sensor-equipped Probe The density estimation schemes in this paper were developed for use in data collection for traffic management in South Korea, Figure 1 depicts an actual probe car used by KICT (Korea Institute of Civil Engineering and Building Technology) for observing traffic. This probe vehicle has multiple sensing equipment that detects nearby vehicles, a Global Positioning System, high resolution radars working with

2

ITS World Congress 2017 Montreal, October 29 – November 2

Traffic Density Estimation Using Radar Sensor Data from Probe Vehicles 77-gigahertz microchips, cameras, and computing processors to calculate traffic density. The application in the main processor calculates traffic density by using sensing technologies, image processing, and GIS.

Fig. 1. Probe vehicle with multiple sensors The above probe vehicle is capable of tracing adjacent vehicles’ trajectory via data fusion from multiple sources such as camera and radars. Outputs from the cameras and radars are collected and displayed in the onboard real-time traffic monitoring tool called TRADOS (TRAffic Density Observation System), shown in Figure 2. The various components of the monitoring tool are labelled as follows: 1. Realtime/Simulation Toggle Button, 2. Data storage location selection, 3. Radar data logging frequency, 4. Driver profile, 5. Data logging start/stop button, 6. Status bar, 7. Detected objects display, 8. Front camera video display, 9. Rear camera video display, 10. Probe vehicle location display, 11. GPS, vehicle dynamics, temperature and humidity data display, 12. Simulation play intervals, 13. Radar data logging interval, and 14. (Real-time mode) system start/stop; (Simulation mode) simulation play/pause

Fig. 2. Traffic monitoring tool onboard the probe vehicle

3

ITS World Congress 2017 Montreal, October 29 – November 2

Traffic Density Estimation Using Radar Sensor Data from Probe Vehicles Figure 3 is a conceptual diagram of how the sensor-equipped vehicle detects adjacent vehicles. Radars are located at various points on the probe vehicle. In the front, there are two radars: a forward long distance radar, and a forward wide range radar. The forward long distance radar generally has been used for “Adaptive Cruise Control” applications. In our research, the forward long radar has a role in detecting vehicles in the same lane. Forward wide range radar is used for object detection and accurate distance estimation for “Collision Avoidance” in modern cars. This sensor has a short sensing distance but has wide sensing range which allows us to detect vehicles in the front side sensing range. The probes have three radars in the rear. Two side backward sensors -- designed for “Blind Spot Detection” -- can detect vehicles driving the rear side lanes. Finally, the backward long distance sensor is used to detect vehicles following the probe. With these specifications and image processing capabilities, the application can now trace trajectories of detected vehicles in real-time. By differentiating the detected vehicles’ movement with respect to time, the probe can capture the relative travel characteristics of detected vehicles, which enables us to measure the detected vehicle’s travel speed. Similarly, the probe can also capture acceleration/deceleration of the detected vehicles, if required.

Space

1

Forward Long

Sensing Distance(m,

)

1 2

Forward Wide

Undetected vehicle Sensing Range(°,

)

Sensing Vehicle Detected Vehicle

Target Angle(°)

Backward Left

4

3

Backward Right

Backward Long

Space

Space

(a) Middle lane

Time

2

4 3

Time

Time

(b) Left lane

(c) Right lane

Fig. 3. Sensing vehicle and time-space diagram example

4

ITS World Congress 2017 Montreal, October 29 – November 2

Traffic Density Estimation Using Radar Sensor Data from Probe Vehicles Methodology to Estimate Traffic Density from Radar Sensor Data At a fundamental level, traffic density (𝑘𝑡 ) at time t is calculated by using some variation of “counting” the number of vehicles (𝑛𝑡 ) on a road unit(𝑋). In other words, traffic density is calculated by counting the nu mber of vehicles over a unit length of the road at a specific time t that we can infer from a snapshot aerial photo. Without loss of generality, traffic density ( k) can also be derived from cumulative number of vehicle s (n(x,t)).

𝑘𝑡 =

∂nt ∂x

𝑛(𝑥,𝑡)−𝑛(𝑥+∆𝑥,𝑡)

= lim

∆𝑥

∆𝑥→0

1 𝑇 𝑘̅ = 𝑇 ∫0 𝑘𝑡 𝑑𝑡 =

=

𝑁([𝑥,𝑥+∆𝑥];𝑡)

1 𝑇 𝑁([𝑥,𝑥+∆𝑥];𝑡) 𝑑𝑡 ∫ 𝑇 0 𝑋

(1)

𝑋

=

𝑇 1 ∫ 𝑁([𝑥, 𝑥 𝑋∙𝑇 0

+ ∆𝑥]; 𝑡)𝑑𝑡

(2)

Traffic density (𝑘(𝑑×𝑡)) of a section d is closely correlated with the average number of detected vehicles on a road stretch (𝑑). A probe vehicle can identify the number of detected vehicles (𝑛 𝑆 ) in the sensing ar ea (𝐴𝑆 , vehicles per square meter), which we call “Area based density”. However, the unit (vehicles/m2) in which this area-based density is represented is not a commonly used unit for traffic density (vehicles/m) i n ITS applications. Thus, this study proposes a method to convert the area-based density to the traffic loc al density (𝑘(𝑑𝑆 ×𝑡)) by using an Ordinary Least Squares conversion method. 𝑛𝑆

𝑘(𝑑𝑆 ×𝑡) ≈ 𝛼 ∙ 𝐴

𝑠

where 𝛼 is a conversion parameter

(3)

When we consider multiple sensors (𝑠 ∈S) in a car, the local density of a probe (p) can be estimated by Eq (4), which is an average of traffic density of all equipped sensors. 𝑆

∑ 𝑘 (𝑑 ×𝑡) 𝑘̅𝑝 (𝑑𝑆 ×𝑡) = 𝑠 𝑠 S 𝑠

(4)

After a probe car passes the road section (𝑥𝑎 , 𝑥𝑏 ) during a time period ( [𝑡𝑎 , 𝑡𝑏 ]), the section density can be estimated by averaging local densities at every time step j. In our study, the density is calculated every 10 seconds by using data points obtained every one tenth of a second. Thus, in this case, J becomes 100 frames, meaning that traffic density of a road section for one time period (10 seconds) is calculated by averaging 100 local density data points from probe vehicles. ̅ 𝑝 (dS ×t) ∑𝐽𝑗=1 𝑘

𝑘̂([𝑥𝑎 , 𝑥𝑏 ]×[𝑡𝑎 , 𝑡𝑏 ]) ≒ where 𝐽 =

(5)

𝐽

𝑡𝑏 −𝑡𝑎 𝑝𝑢𝑙𝑠𝑒 𝑟𝑎𝑡𝑒

, 𝑗 = 1,2,3,4, … , 𝐽

When we consider multiple probes (𝑝 ∈ 𝑃) in a study area, Eq (5) can be extended by averaging the traffic densities captured by multiple probes, as in Eq (6)

k̂([xa , xb ]×[t a , t b ]) ≒

𝐽 ̅ 𝑝 (dS ×t) ∑𝑃 𝑝 ∑𝑗=1 𝑘

(6)

𝑃𝐽

Figure 3 also shows an example of vehicle detection in such a case. There are five vehicles near a probe car, which is driving in the middle lane. Out of the five vehicles, four are detected and the trajectories of detected vehicles are also depicted in the figure. The local density is calculated by counting the number of detected vehicles over the sensing area. The shaded area in each graph represents the time-space domain of the sensing area of each lane. Some portion of traffic conditions in the road section are sampled as a probe car drives on the road stretch. From the captured number of vehicles in the shaded area, we can calculate traffic density by summing up all detected vehicles’ travel times over the shaded area as shown by Eq (6).

5

ITS World Congress 2017 Montreal, October 29 – November 2

Traffic Density Estimation Using Radar Sensor Data from Probe Vehicles Evaluation of Proposed Algorithm Data Sets and Simulation Though the technical viability of our methods could be established with actual sample data collection using the KICT probe vehicle shown above, a sufficient number of them are required to derive the true benefits of the scheme. Thus simulation using observed vehicle trajectories, but assuming some of the overserved vehicles to be radar-equipped probe vehicles is one option to evaluate the efficacy of the scheme at various market penetration rates of equipped vehicles. For this study , we consider a six-lane freeway stretch on I-80 in Emeryville, California depicted in Figure 4. This stretch is 0.9 km in length, and contains one on-ramp (at Powell St.) and one off-ramp (at Ashby Avene). The median or innermost lane on this freeway stretch is an HOV lane. As part of the NGSIM (Next Generation Simulation) project undertaken by the Federal Highway Administration (FHWA) of USA (FHWA, 2006), video data on vehicle movements on this stretch was recorded by multiple cameras. This video data was then converted to vehicle trajectories that show positions every one-tenth of a second. The NGSIM datasets contain the following important information relevant to our study: 1. VehicleID, 2. Vehicle Type (motorcycle is removed), 3. Vehicle length and width, 4. Frame ID, 5. Coordination. Due to the high granularity of vehicle trajectories present in this data set, it was chosen as a suitable candidate for the implementation of our proposed study. Four datasets are now publicly available (Cambridge Systematics, 2005). Among them, this study utilizes the trajectories on April 13, 2005 between 5:15pm to 5:30pm. The number of vehicles in this dataset is 1,790, and a total of 17,533,791 vehicle positions and vehicle attributes are recorded for a time interval of 1,168 seconds (11,684 frames). Vehicle trajectories in the NGSIM data are assumed to start at the beginning of the road stretch. However, there are vehicles that are already present at various locations on the road when the NGSIM system starts recording. Therefore, we removed the first 150 seconds from our analysis, since the data from this time period does not accurately represent the state of the traffic. Similarly, traffic conditions at the end of the time period are not representative of the real traffic conditions, and thus the last 118 seconds data was also discarded.

Fig 4. I-80 Study area and simulation example NGSIM datasets give us a valuable test-bed to evaluate our algorithm. Using these datasets, we created a Python application to simulate and construct vehicle trajectories and overlaid them over the road stretch under consideration. This application allowed us to visualize the performance of sensor-equipped vehicles. For instance, for better visualization, we coded probe vehicles in red, and vehicles sensed by the probes in real time as yellow. The rest of the vehicles stay blue and only change their color to yellow when they are in the sensing area of radar-equipped probe vehicles. A zoomed-in screenshot of the road segment near the Powell street on-ramp is shown in figure 5.

6

ITS World Congress 2017 Montreal, October 29 – November 2

Traffic Density Estimation Using Radar Sensor Data from Probe Vehicles

Fig 5. Depiction of probe vehicles (red) and sensed vehicles (yellow)

Simulation Results In our simulation application, we can set the penetration ratio of probe vehicles, and the performance characteristics of radars. With the sensor data, the local traffic density is calculated with the number of observed vehicles and the area of observation. Each probe vehicle also sends collected data to a central computing server, which then estimates the road section’s traffic density using our developed algorithm. For this study, the radar configuration on the probe vehicles is assumed as shown in Table 1.

Table 1. Radar configuration Sensor Code

Name

Target Angle (°)

Sensing Range(°)

Sensing Distance(m)

1

Forward Long

0

10

30

2

Forward Wide

0

60

10

3

Backward Left

160

40

10

4

Backward Right

200

40

10

5

Backward Long

180

20

20

The plots in Figure 6 show the estimated densities on the study area in comparison to the real density over the entire time period, for different probe vehicle penetration rates. The solid line indicates the real (i.e. ground truth, Eq 2) density that is calculated by averaging total number of vehicles over the road stretch during time periods of 10 seconds. The dotted teal, green, blue, and red lines represent densities estimated using our proposed method (Eq 6) for penetration rates of 1%, 5%, 10%, and 25% respectively. The overall pattern that we see is that the estimated densities explain the ground truth density and that the fluctuation of estimated density is dampened at higher penetration rates. It is interesting to note that the patterns of under-estimation and over-estimation remain the same for most penetration scenarios.

7

ITS World Congress 2017 Montreal, October 29 – November 2

Traffic Density Estimation Using Radar Sensor Data from Probe Vehicles

Fig 6 Estimated densities in comparison with real density The estimation improves drastically when the penetration rate changes from 1% to 5%. At extremely low values of penetration (1%), the density estimates show larger error, which is to be expected since there simply aren’t enough probe vehicles in the network. Since our method does not rely upon any other source of data other than probe vehicles, it is evident that there needs to be a minimum threshold of probe vehicle penetration for accurate density estimation. Table 2 depicts the performance of our proposed algorithm with respect to the Percent Root Mean Square Error (RMSE, Eq 7), and Percent Relative Error (Eq 8). 2

𝑛

∑ (𝑘 −𝑘𝑒𝑠𝑡,𝑖 ) % 𝑅𝑀𝑆𝐸 = √ 𝑖=1 𝑜𝑏𝑠,𝑖 ×100 𝑛 1

𝑘𝑜𝑏𝑠,𝑖 −𝑘𝑒𝑠𝑡,𝑖

% 𝑅𝐸 = ∑𝑛𝑖=1 𝑛 |

𝑘𝑒𝑠𝑡,𝑖

(7)

| ×100

(8)

Both these MOEs show consistent patterns, with the errors dropping by almost 60% between penetration values of 0.01 and 0.05, and then remaining relatively constant for higher penetration rates. In the particular traffic conditions on this road stretch, the probe vehicle penetration rate of 0.1 is sufficient to estimate traffic density. It is perhaps notable that radar data from such a relatively small fraction of vehicles is enough to obtain very accurate density estimates, as it points to potential practical success of the method in future.

Table. 2 MOEs by penetration rates

MOE % RMSE % RE

1% 28.37 21.96

Penetration Rate 5% 10% 12.51 11.20 8.55 7.72

8

25% 11.50 7.45

ITS World Congress 2017 Montreal, October 29 – November 2

Traffic Density Estimation Using Radar Sensor Data from Probe Vehicles Conclusion Vehicle sensing technology is expected to play a pivotal and active role in the near future of traffic estimation, management, and control. This technology gives rise to alternate sources of traffic data such as probe vehicles, which are independent from expensive and invasive infrastructure such as loop detectors. In this study, we develop an algorithm using data obtained only from probe vehicles. Our method takes into account various factors such as geometries of sensors, sensing performance, and dynamics of traffic flow. Our algorithm was motivated by state-of-the-art specifications of sensor vehicles currently deployed as part of a research project conducted by the Korea Institute of Civil Engineering and Building Technology. The proposed method is evaluated by creating an application that visualizes NGSIM data and depicts sensor functionality. Measures of effectiveness such as RMSE and Relative Errors, were computed for the study area with radar-vehicle penetration ratios of 1%, 5%, 10%, and 25%. Results show that our estimation improves significantly when the penetration rate changes from 1% to 5%. Furthermore, estimation errors from 10% to 25% penetration scenarios are not significantly different in the studied network, but further studies are required before we can make generalizations. Future work would involve the simulation of more realistic traffic conditions. The NGSIM dataset, while useful, is constrained in terms of time and location. It contains vehicle trajectory information only for a 15 minute period on a linear road stretch. A detailed simulation of a real life network for a longer time period (i.e. at least an hour), under varying traffic flow regimes is needed, and is the subject of ongoing research by the authors.

Acknowledgements This research was supported by a grant from the Strategic Research Project (Development of Driving Environment Observation, Prediction and Safety Technology Based on Automotive Sensors) funded by the Korea Institute of Civil Engineering and Building Technology. References 1. Al-Sayed Ahmed AL-Sobky and Ragab M. Mousa. (2016) Traffic Density Determination and Its Applications using Smartphone, Alexandria Engineering Journal, Vol: 55, p.p. 531-523 2. Cambridge Systematics, Inc. NGSIM I-80 Data Analysis (4:00 p.m. to 4:15 p.m.). Technical report, September 2005. Summary Report, Prepared for Federal Highway Administration. 3. Cambridge Systematics, Inc. NGSIM I-80 Data Analysis (5:00 p.m. to 5:15 p.m.). Technical report, September 2005. Summary Report, Prepared for Federal Highway Administration. 4. Cambridge Systematics, Inc. NGSIM I-80 Data Analysis (5:15 p.m. to 5:30 p.m.). Technical report, September 2005. Summary Report, Prepared for Federal Highway Administration. 5. Federal Highway Administration. Next Generation SIMulation Fact Sheet. Technical report, December 2006. FHWA-HRT-06-135 6. INRIX and Centre for Economics and Business Research. (2014). 50% Rise In Gridlock Costs by 2030. 7. Ryan Herring, Aude Hofelitner, Pieter Abbeel, and Alexandre Bayen, (2010), Estimating Arterial Traffic Conditions using Sparse Probe Data, 13th International IEEE Conference on Intelligent Transportation Systems, Madeira Island, Portugal, 2010 8. Toru Seo and Takahiko Kusakabe. (2015), Probe Vehicle-based Traffic Flow Estimation Method without Fundamental Diagram, Transportation Research Procedia 9, 21st International Symposium on Transportation and Traffic Theory, Kobe, Japan, 2015, p.p. 149-163

9