Granular Quantifying Traffic States Using Mobile Probes Quang Tran Minh
Eiji Kamioka
Graduate School of Engineering Shibaura Institute of Technology Tokyo, Japan
[email protected]
Graduate School of Engineering Shibaura Institute of Technology Tokyo, Japan
[email protected]
Abstract— This paper proposes a novel method for detecting traffic congestions, qualifying and quantifying congestion levels using mobile phones as traffic probes. The system provides a robust mechanism for granularly comparing the seriousness of different congested areas. Congested areas are detected in a detailed manner by which exact congested positions are reported. Moreover, congestions can be detected even though no complete traffic trace due to the traffic jam is collected. This feature is quite different from, and makes the system more robust compared to the previous ones. This project also consists of a reasonable vehicle classification method based on only GPS data. This mechanism improves not only the effectiveness and the accuracy but also the scalability, thus the system is flexibly applicable for any traffic system structure, especially in developing countries where a lot of motorbikes are travelling on the roads. The evaluation reveals that the proposed ideas are novel which are not discussed in the existing work. Keywords-traffic estimation; mobile phone probes; vehicle classification; quantifying traffic state, real-time traffic data; GPS.
I.
INTRODUCTION
Traffic jam is a serious social issue in almost every country. It causes economic loss, air pollution as well as other socialrelated issues. Obviously, this issue becomes more serious in developing countries where the traffic infrastructure has not caught up the transportation demands. Recent years, several researches on Vehicular Technology Systems (VTS) and Intelligent Transportation Systems (ITS) are dedicated for finding suitable solutions aiming at reducing the traffic congestion to be occurred. These studies majorly focus on estimating traffic states, disseminating discovered information to drivers thus help them to avoid from entering congested areas. Traditional systems relied on road-side fixed sensors to record the times when transponders cross sensors’ locations. The sensors might be loop detectors [1], [2], RFID readers [3], [4], etc. The essential drawbacks of these systems are their coverage limitation and their sensitiveness to errors and malfunctioning. To solve these drawbacks, on-board devices should be utilized. One may employ ad-hoc networks which includes wireless sensors equipped on vehicles [5], or utilize GPS receivers equipped on navigation systems [6], [7], [8], [9], [10]. However, these systems required costly devices equipped
on each vehicle and ad-hoc networks may not work properly when the density of vehicles is inadequate. Nowadays, with the advance of the mobile technology, mobile phones are investigated to be utilized as traffic probes recording real-time traffic data for the traffic estimation [11], [12], [13], [14]. This approach might help to solve the issues such as the coverage limitation, the real-time effect, the installation and maintenance cost, tec., since mobile phones are available everywhere. Nevertheless, several issues emerging from this technology are not thoroughly solved in the existing work. This paper aims to propose solutions for following essential issues: 1) the issues on vehicle classification (e.g. motorbike, car, bus, truck,…) to improve the effectiveness and the accuracy of the traffic estimation; 2) the issues on granularly detecting congestion areas without waiting for the traffic trace data; and 3) the issues on quantifying traffic states for evaluating congestion levels. As our best knowledge, no reliable method utilizing GPS data to quantify the traffic states in a detailed level was proposed. This work proposes a robust mechanism for granularly quantifying traffic states into continuous values, hence even the slightly different traffic states can also be recognized and granularly comparable. The remaining of the paper is organized as follows: Section 2 overviews the related work. An overviewed architecture of the proposed system is described in section 3. Section 4 summaries the challenges in traffic estimation using mobile probes and briefly discusses the proposed solutions. Section 5 describes the novel model for qualifying and quantifying traffic states. The evaluation is presented in section 6, and section 7 concludes this work. II.
RELATED WORK
VICS (Vehicle Information and Communication System) [15] is one of the well-known systems in Japan that provides real-time road traffic information. This system collects traffic data based on a huge number of fixed sensors of infrared ray, quasi-micro and FM wave techniques. A server at the VICS center processes data to estimate traffic states and distribute the information to the car navigation devices or to the Internet. This system works well but it requires a huge cost for the installation and maintenance. Moreover, it also suffers from the coverage limitation. This work is different, it focuses on utilizing on-board, mobile sensors to improve the coverage of the system and reduce the cost.
978-1-4244-3574-6/10/$25.00 ©2010 IEEE
Ad-hoc networks [5] might help to overcome the coverage issues but such systems require navigation devices equipped on vehicles. Even though these devices’ price is getting cheaper, the accumulated money required for every vehicle makes the system costly. Besides that, the system may not work properly in case of sparse travelling vehicles since the data cannot be transmitted. Mobile Millennium Project (MMP) [14] is closely related to our work in which GPS-enable mobile phones are utilized as traffic probes for real-time data collection. The differences, however, between this project and our work are as follows: 1) MMP did not classify vehicles (e.g. bus, truck, car, motorbike) travelling on the roads thus it might not yield a detailed traffic state estimation; 2) MMP estimated the congestion level by employing the dynamical theory which analyzes vehicle flows on roads. The density inference based on the current traffic state and the vehicle flows on a long road may cause the accuracy decline. J.Yoon, et al. [11], proposed to divide the road system into shorter segments and quantify the traffic state based on segment basis. The traffic state, however, is analyzed based on the complete traces of vehicles on the considered segment. Obviously, this mechanism could not estimate the traffic state of serious congested segments since no traffic trace due to the traffic jam can be reported to the server. In additions, this work did not discuss the issues of vehicle classification either. III.
SYSTEM ARCHITECTURE M o b ile p h o n e s , P D A o r N a v ig a tio n d e v ic e s
T ie r 1 E s tim a tin g S erver
T ie r 2 D a ta b a s e S erver
T ie r 3
Figure 1. The architecture of the proposed system
The architecture of our system consists of three tiers as described in Fig.1. Mobile devices such as mobile phones or PDAs (1st tier) collect and report the real-time traffic data to the server. The server at the 2nd tier does the data preprocessing, estimates and disseminates the traffic information to the mobile devices or to the Internet. The database server (3rd tier) stores the real-time traffic data and a digital map which is needed for later traffic estimation and information dissemination processes. The distributed information consists of not only the congestion areas but also the granularly quantitative information of such congestions. The major purpose of this work is to introduce a sound method for quantifying traffic states. If the system provides a scaling metric for congestion levels, drivers can figure out how
serious the congestion is, and the congestion levels can be granularly comparable. IV.
CHALLENGES IN TRAFFIC ESTIMATION
Intuitively, the seriousness of a traffic state may be represented by two major factors: 1) the mean speed of vehicles, and 2) the number of participating vehicles (i.e. the density of the traffic flow). However, in this work, only GPS data is handled since GPS is the only common sensor equipped on commercial mobile phones. In additions, several difficulties are waiting as follows: 1) since the mean speed of vehicles may vary from road segment to segment, it cannot be interpreted directly to a detailed traffic state; 2) how to calculate the mean speed of various types of vehicles when the limitation speeds of different vehicle modes (i.e. bus, car, motorbike,…) in the same segment are different?; 3) how to classify vehicle modes and take them into account the calculation for the traffic density?; and 4) how to combine the affect of both the factors “mean speed” and “density” for quantifying congestion levels? How to compare traffic states of two segments in a granular manner? The solutions for solving aforementioned difficulties will be introduced and analyzed thoroughly in following sections. A. Estimating Traffic States Based on the Segment Basis As mentioned, the traffic characteristics are commonly varied from segment to segment thus it should be estimated based on the segment basis. Here, the segment is defined based on those features that contribute to the improvement of the traffic estimation’s accuracy. Obviously, road-land marks such as intersections, crosswalks, curved places, places where the width of the segment or the number of lanes change, etc., should be the start points of segments. In the case of the long roads whose characteristics do not change frequently such as the highways, the segments are divided in every 1km. When data is collected from a segment, a corresponding data structure called the data-segment is created and stored in the database server for later traffic estimation as depicted in Fig.2. Figure 2 shows that each travelling vehicle is represented by one and only one data point which consists of the position, heading, velocity, timestamp, the virtual ID of the phone sending this data and the vehicle mode.
1km
1km
Figure 2. Road segmentation and a corresponding data-segment which consists of data points of 3 vehicle modes
B. Reporting the Right Data at the Right Time Different from the existing systems, a vehicle neither reports its information in a time-interval setting nor whenever it enters a segment [13]. Instead, the mobile phone recognizes and reports only the utilized data which is the one collected
when the speed of a vehicle significantly change. It is clear that, a vehicle reduces velocity when it enters a bad traffic state segment and it accelerates when the traffic condition is getting better. The velocity change rate is calculated as in (1) where v and v0 are the current and previous speeds of a vehicle. As shown in Fig. 3, data is reported only when the velocity change (increase or decrease) significantly at time t1, t3 and t4, and the new data point replaces the previous one at the server’s storage.
r =|
v − v0 | v0
(1)
Speed P1
t1
Input: GPS data includes the default unknown mode: “U” Output: A special mode is assigned to the vehicle Method: 1. GPS data is sent to the server 2. Sever recognizes and assigns a specific mode to mobile phone 3. If a specific mode is assigned to client then STOP Else the mobile phone will try to indentify its mode using “acceleration pattern” analysis. 4. The server classifies the remaining unidentified vehicles based on their “passing time patterns” at the congestions Figure 4. Pseudo code for vehicle classification based on only GPS data
P2
t0
decides whether to report its information when reaching a new position. This approach not only significantly reduces the data transmission load but also helps to avoid the estimation’s bias.
Default mode: U
P3 t2
Data segment 1
t3
t4
Mobile phone sends data to server
t5 Time/Road
Data segment 2
On the road
Y
N
Mode: W Sever sends boundary to MB
Y
Figure 3. A vehicle reports its information based on its velocity change rate
MB out of the boundary
Count the number of MB on the same vehicle
C. Vehicle Classification Using Only GPS Data Different vehicle modes, such as car, bus, truck, motorcycle, etc., differently influence the traffic state. Existing systems did not mention this factor, thus their accuracy is not clarified, especially when they are applied in the places where several vehicle modes present. Depends on the purpose of the system, there are several methods for vehicle classification. In the context of traffic estimation, obviously, vehicles should be differentiated into groups based on their influence the traffic states. According to our field studies, there are three groups of vehicles which differently influence the traffic state: 1) light vehicles include {Sedan, Pickup, SUV, Merged, Van}, denoted as “C”; 2) heavy vehicles include {Semi, Truck, Bus}, denoted as “B”; 3) bikes include {Motorbike, Bicycle}, denoted as “M”. In addition, the walker mobile phones, denoted as “W”, must be recognized. Obviously, the shapes of vehicles give a straight direction for classification. Several works utilized this trait and developed specific sensors such as video cameras, micro wave sensors, etc., [16], [17] for recognizing vehicles’ shapes. These approaches, however, lead to the usage of road-side fixed sensors which does not satisfy the economical feasibility and the coverage requirement. Our proposed framework aims at working with only GPS data collected by mobile phones to keep a low installation and maintenance cost. This setting reveals nontrivial challenges for investigation. Our classification model is described in Fig.4. The classification process at the server side depicted in step 2 (Fig.4) is described as a flowchart in Fig.5. Based on the digital map and the GPS data reported from the mobile phone, the server can detect if this client is on the road or not. If it is not on the road, its mode is set to “W” (walker) and its boundary is provided. According to this boundary, the client
N>7
N: 3-7 N
Mode: B
Mode: BC Else
Figure 5. Vehicle classification at the server side based on GPS data
On the other hand, if the client is on-road, the number of passengers in the same vehicle is counted. It is clear that the data points sent from these passengers consist of similar features such as timestamp, velocity, direction, etc., which are adequate for inferring different mobile phones in the same vehicle. If this number is greater than 7 then that vehicle should be a bus (“B”). Otherwise, if the number is from 3 to 7 then the vehicle is set to “BC” (i.e. it should be in “B” or “C” group but no specific one is identified yet). A vehicle consists of less than 3 passengers cannot be classified at this time. After this process, some buses (i.e. "B") or a more specified mode of "BC" are identified. In addition, the system can recognize the situation of several mobile phones on the same vehicle. The remaining vehicles in “U” and “BC” groups can be classified in the mobile phone itself based on the “acceleration pattern” analysis (step 3 in Fig.4) as depicted in Fig.6. Figure 6a shows that a vehicle whose acceleration pattern is large (Fig.6b) should be classified into “B”. The narrow acceleration pattern vehicle (Fig.6c) is classified into “C” or “MC” (i.e. motorbike or car) depending on its initial value as “BC” or “U”, respectively. At this stage, almost every vehicle is classified into a specific group of “B”, “C” or “MC”. Until this stage, almost every vehicle is classified into a specific group of “B”, “M” or “C”. A small fraction of remaining unclassified vehicles are labeled as “MC” (in Fig.6a) which is more specific than the unknown class, “U”.
These vehicles can be classified at the server side based on their “passing time patterns” at the congestion areas.
Vavgi =
¦ Speedin 5 Percentitl eto 95 Pecentileo fMode _ i ni ( from 5 Percentile to 95 Percentile )
Speed
M vi =
Initial mode: U/BC
Time (s) U -> B Analyzes the Large BC -> B b) Acceleration pattern of heavy vehicle acceleration pattern Speed a) Classification flow Narrow chart on the client side U -> MC BC -> C
Time (s) c) Acceleration pattern of light vehicle
Figure 6. Vehicle classification at clients based on acceleration patterns
The field studies revealed that a motorbike travels faster than a car at congested areas. In our model, the server handles the passing information of the “MC” vehicles. According to this data, the server can compare the time a vehicle needs to pass in every equivalent interval, says 15m for instance. An “MC” vehicle whose passing time matches the motorbike passing time pattern (i.e. the fast one) is classified into the “M” class; otherwise it is classified into the “C” class. At this stage, every vehicle is classified into one of the three desired groups “B”, “M” or “C” and the walkers are labeled as “W”. V.
QUALIFYING AND QUANTIFYING TRAFFIC STATES
After collecting the real-time traffic data, the system estimates
traffic states and disseminates this information to drivers. The requirement here is that the system must not only distribute the qualitative information but also the quantitative information about the traffic states. Obviously, a congestion occurs when a cluster of vehicles travel slowly and occupies the “room” of a road segment. This section discusses the relationship between the cluster’s characteristic such as the mean speed, the density, etc., to the traffic state. A. Qualifying Traffic States For quantifying traffic states, the mean speed and the density of vehicles are two major factors should be investigated. For calculating the mean speed of influence vehicles, only the 5th to 95th percentiles of speed are considered for avoiding any bias. Moreover, the average speeds of individual vehicle modes should be calculated separately before synthesizing them in a suitable way to represent the mean speed capacity of a segment. Let ni the number of vehicles in mode i, Vmaxi the limitation speed of this mode at the considered segment (e.g. 80km/h for cars, 60km/h for buses, etc.). We define the “average speed”, Vavgi; the “mean speed capacity”, Mvi of mode i; and the relative “mean speed capacity” of all vehicles travel on the current segment, Mv as in equations (2), (3) and (4). Qualitatively, Mv represents the traffic state in the manner of velocity. The high value of Mv represents a good traffic condition and vice versa. Here, the threshold of Mv is set to 60% for a good traffic condition.
Mv =
Vavgi Vmax i
* 100%
¦M *n ¦n vi
i
*100%
i
(2) (3) (4)
In addition to the mean speed capacity, the density should be analyzed. It can be defined as the ratio of the number of participating vehicles to the occupation capacity of that area (i.e. the number of vehicles which can be absorbed by that area). Let C the capacity of the covered cluster in the considered segment. This value is calculated as in (5), where l is the length of the cluster, lcar is the average length of a car which is set to 5m [18], and 1.5 is the coefficient describing the space which must be yielded between two cars at the congested area. Let S the approximated occupation space in the covered cluster which is calculated as in (6), where ncar, nbus, nmotorbike are the number of cars, buses, motorbikes, respectively. Here, it is approximated that a bus occupies a space relatively equals to that of 4 cars, and in turn a car occupies a space equals to that of 4 motorbikes. Let įs the FREE space ratio, which is calculated in (7), representing the ratio of the free space in the considering cluster. Intuitively, the high value of įs represents a good traffic condition and vice versa. In this work, the threshold of įs is set to 40% for a good traffic condition.
C = numberoflanes *
l 1.5 * l car
(5)
S = ncar + nbus * 4 +
nmotorbike 4
(6)
σS =
C−S C
(7)
The combination of the mean speed capacity and the density analysis creates a so-called “traffic-state" quadrant, which can be used to qualify traffic states in a more detailed manner as shown in Fig.7. The 1st quadrant represents good traffic states, the 3rd one for the bad states while the 2nd and 4th quadrants represent the threat conditions. For more detailed the 2nd quadrant expresses situations where few vehicles are travelling in a slow speed while the 4th one represents for the opposite situations (i.e. good speed but in high density).
2
1
3
4
Figure 7. Quadrant space qualifying traffic states
B. Quantifying Traffic States Obviously, a point that is far apart from the new origin (the origin forms the quadrant space – Fig.7) upward and rightsided ward represents a better state and vice versa. The ideal traffic state is represented by the point at the coordination (1,1) while the origin (0,0) represents the worst traffic condition. For convenience, let x and y the axes represent for the mean speed capacity (Mv) and the free space ratio (įs), respectively. The new origin is (x0, y0) where x0, y0 are the aforementioned thresholds of Mv and įs, respectively. We propose a coefficient called the “Goodness metric – G(x,y)” to quantify the traffic state of a segment which is mapped to the quadrant space at the point (x,y). This coefficient is defined as in (8) which is a continuous value, ranging from -1 to 1, representing traffic states from the worst to the best ones. For instance, the Goodness values of the worst and the best traffic states are calculated as: Gworst = G(0,0) = (0-0.6) + (0-0.4) = -1 and Gbest = G(1,1) = (1-0.6) + (1-0.4) = 1. G(x,y) = (x-x0) + (y-y0)
by the mobile phone and the velocity change rate of every two consecutive values is extracted using expression (1). After that, four statistical values, namely the maximum, minimum, average and the standard deviation of the 24 velocity change rates are utilized as the input of a 4 input nodes and 6 hidden nodes artificial neural network (ANN) [19]. The data of 23 bus and 7 car driving on the highway in the middle of the way from Omiya Campus to Toyosu Campus (about 70km), Shibaura Institute of Technology, was collected. This experiment classifies only two kinds of vehicle (i.e. car and bus) thus only one output node is set for the aforementioned ANN and the result is depicted in Fig.9. The figure shows that the result is really good if the threshold is set to 0.2 (i.e. the vehicle is a car if the output from 0 to 0.2) and 0.8 (i.e. output of 0.8 to 1 represents for a bus). In this case about 80% of vehicle can be differentiated based on the velocity change rates.
(8)
According to the Goodness metric, traffic states can be qualified in a more detailed manner. Figure 8a shows that the Goodness values of all the points on the cross line are identical and equal to that of the new origin (i.e. G(x0,y0) and is 0). The points upper the cross line which are not in the 1st quadrant represent the fairly good traffic state segments. Otherwise, the points under this line represent the bad or the threat ones. Figure 8b shows the Goodness values of some selected points in the quadrant space. The vertical axis represents the Goodness values of road segments which are mapped onto the quadrant space on the bottom of the chart. It should be recalled that the Goodness value is continuous so that it is quite relevant for granularly comparing traffic state levels.
Figure 9. The vehicle classification result using an ANN
The special feature of this work, compared to the existing ones, is that different vehicles modes are analyzed thoroughly in the manner of their influence the traffic states. Obviously, this feature improves the effectiveness as well as the accuracy of the proposed system significantly. In additions, this vehicle classification mechanism is actually robust even though only the GPS data, which is less informative for the classification purpose, is handled. Using the Goodness metric is a practical approach to quantifying the traffic states. Traffic states are quantified into continuous values ranging from -1 to 1 so that drivers can realize how serious the traffic state of a segment (or of a cluster for more detailed) is and the traffic state levels of different segments can be compared in a granular manner. This approach improves the effectiveness of the quantification results. For our best knowledge no system like this has been reported before. Clip 1 (v1)
(a)
(b)
Figure 8. Qualifying the traffic states based on the Goodness metric
VI.
EVALUATIONS
This section analyzes the effectiveness and the scalability of the proposed traffic states quantifying method using mobile probes. A. Effectiveness Evaluation To evaluate the effectiveness of the vehicle classification model, the velocity change rate (a kind of relative acceleration) is collected. A soft ware is installed on an iPhone to receive the GPS signal and calculate the velocity of the carrying vehicle. Every 25 consecutive velocity values of a vehicle are recorded
Clip 2 (v2)
Clip 3 (v3)
- Segment: 2 lanes, 100m - Segment: 2 lanes, 100m - Segment: 2 lanes, - 4760 vehicles/h - 3360 vehicles/h 130m - Truck/car = 2/23 - Truck/car = 3/11 - 4800 vehicles/h - Truck/car = 4/16 - Mean Speed Capacity: - Mean Speed Capacity: - Mean Speed Capacity: 4.3% 10.26% 13.6% - Free Space Ratio: - Free Space Ratio: - Free Space Ratio: 0% 7.69% 11.54% Value: - Goodness Value:-0.82 - Goodness Value: - - Goodness -0.96 0.75 Evaluations from 12 people Giving a range of 1 to 10 represents from the best (free flow) to the worst (congested) traffic states, they are asked to quantify these traffic state levels. Most of them quantified v2-v1-v3 as 6-7-8, some ones quantified as 5-6-7. Figure 10. Effectiveness of the traffic state quantification model
To evaluate the effectiveness of the traffic state quantification model, its 1st prototype has been implemented. For testing, three video clips, say v1, v2, v3, of real road segments on the National Route No.16 (Higashi Omiya, Saitama, Japan) were taken. The table in Fig.10 shows the
consistency of the estimation using the proposed model compared to the human being evaluation. The model shows more detailed quantified Goodness values as -0.75, -0.8, -0.96 for v2, v1, v3, respectively. The human being evaluated these corresponding traffic states as the relatively ratio as 6 – 7 – 8 in a range of 1 to 10 (ranging from free flow to congested states). B. Scalability Evaluation In this work, the road system is divided into segments and the traffic state is estimated based on the segment basis. Here, the important characteristics of a segment such as the segment’s width, the number of lanes, the maximum allowed speed for each vehicle modes, etc., are taken into account. This mechanism allows the system to detect the congestion in a more detailed place (i.e. in the segment basis). In additions, the system can exactly identify at what subarea in the considering segment the congestion occurs. As analyzed, this method can identify the clusters of density vehicles in a segment thus more than one congested areas in a segment can be detected as it happens in the real world. This is one of the special characteristics of this work compared to the existing ones. Some existing systems work well for highway traffic estimation, others work well for the artery or the city traffic systems. This proposed system is a rare one that can work well regardless of traffic structures since it estimates traffic state based on the road segment basis. In additions, the system can cover a wide area such as the national wide and it can be applied in any country’s traffic structure, especially in developing countries where a lot of motorbikes participate in travelling on the roads.
essential barriers of the system to be realized, thus more efforts are needed in testing and finding suitable solutions for realworld applications. REFERENCES [1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
VII. CONCLUSIONS AND FUTURE WORK This research proposed a novel method for granularly qualifying and quantifying the traffic states. The model can detect, qualify and quantify congestions’ levels even though no complete passing trace is recorded at the congested areas. This approach is quite different and more robust compared to those systems which rely on the traffic traces. The quantification model can be set up as a tool for other work in this field in quantifying traffic states. The idea of vehicle classification based on only less informative GPS data (in the manner of vehicle classification) is not represented in previous work. This paper proposed a simple yet effective way for solving this issue. This technique improves the effectiveness, the computational performance as well as the scalability of the system. Nevertheless, the accuracy and the unavailability of GPS data at some restriction places such as in the tunnels, under big buildings, etc., are major remaining challenges. In the future, more detailed performance will be evaluated in order to confirm the robustness and the flexibility of the proposed approach. We are also planning to investigate other sensors such as vibrancy sensors, thermo-sensors, noise sensors, etc., which are possible to be quipped on mobile phones in the near future, in the combination with GPS receivers to improve the accuracy of the system. The complexity of the system and the requirement of the data sending from users’ mobile phones are
[10]
[11] [12]
[13]
[14] [15] [16]
[17]
[18] [19]
B. Coifman, “Identifying the onset of congestion rapidly with existing traffic detectors." In Transportation Research, volume 37 of Part A, pp. 277̄291. 2003. W. Lin and C. Daganzo, “A simple detection scheme for delay-inducing freeway incidents.” In Transportation Research, volume 31A of Part A, pp. 141̄155. 1997. J. Wright and J. Dahlgren, “Using vehicles equipped with toll tags as probes for providing travel times.” California PATH Working Paper UCB-ITSPWP-2001-13, Institute of Transportation Studies, University of California, Berkeley, CA, 2001. X. Ban, Y. Li, A. Skabardonis, and J.D. Margulici, “Performance evaluation of travel time methods for real time traffic applications.” In 11th World Conference on Transport Research, Berkeley, CA, June 2007. N. Shibata, T. Terauchi, T. Kitani, K.Yasumoto, M. Ito, T. Higashishino, “A Method for Sharing Traffic jam Inforamtion Using Inter-Vehicle Communication,” mobiquitous, 3rd Annual International Conference on Mobile and Ubiquitous Systems – Workshops, pp. 1-7, 2006. J. Santa, A. F. Gomez-Skarmeta, “Sharing Context-Aware Road and Safety Information,” IEE Pervasive Computing, vol. 8, no. 3, 8 -65, July – Sept. 2009, doi:10.1109/MPRV.2009.56. K.K. Sanwal and J.Walrand, “Vehicles as probes.” California PATHWorking Paper UCB-ITS-PWP-95-11, Institute of Transportation Studies, University of California, Berkeley, CA, 1995. R. Zito, G. D̉Este, and M. Taylor, “Global positioning systems in the time domain: how useful a tool for intelligent vehicle-highway systems?” Transportation Research C pp.193̄209, 1995. J. Ygnace, C. Drane, Y. Yim, and R. Lacvivier, “Travel time estimation on the san francisco bay area network using cellular phones as probes.” Technical Report UCB-ITS-PWP-2000-18, California Partners for Advanced Transit and Highways (PATH), Sept. 2000. M. McNally, J. Marca, C. Rindt, and A. Koos, “Tracer: In-vehicle, gps based, wireless technology for traffic surveillance and management..” Technical Report UCB-ITS-PRR-2003-23, California Partners for Advanced Transit and Highways (PATH), July 2003. J. Yoon, B. Noble, and M. Liu “Surface Street Traffic Estimation” MobiSys ’ 07, San Juan, Puerto Rico, USA, June 11-14, 2007. D. B. Work and A. M. Bayen, “Impacts of the Mobile Internet on Transportation Cyberphysical Systems: Traffic Monitoring using Smartphones.” In National Workshop for Research on High-Confidence Transportation Cyber-Physical Systems: Automotive, Aviation and Rail, Washington, DC, Nov.18-20, 2008. B. Hoh, M. Gruteser, R. Herring, J. Ban, D. Work, J.C. Herrera, A. Bayen, M. Annavaram, and Q. Jacobson, “Virtual Trip Lines for Distributed Privacy-Preserving Traffic Monitoring.” MobiSys’08, Breckenridge, Colorado, USA, June 17–20, 2008. http://traffic.berkeley.edu/theproject.html. Ministry of Land Infrastructure and Transport: “The System Outline of VICS,” http://www.vics.or.jp. C. Setchell, and E. L. Dagless, "Vision-based road traffic monitoring sensor." Vision, Image and Signal Processing, IEE Proceedings, pp. 7884, 2001. J. Gajda, R. Sroka, M. Stencel, A. Wajda, and T. Zeglen, “A Vehicle Classification Based on Loop Detectors.” Proceedings of the 18th IEEE Conference IMTC, pp. 460-464, 2001. http://en.wikipedia.org/wiki/Full-size_car. D. E. Rumelhart, B. Widrow, M. A. Lehr, “The basic ideas in neural networks,” Communications of the ACM, v.37 n.3, p.87-92, March 1994.
