The recent emergence of automatic vehicle identification (AVI) technology for use in electronic toll collection has provided an opportunity to develop automatic incident detection (AID) methods that rely on individual vehicle travel time data rather than loop detector data. This paper examines the performance of three AVI based AID algorithms. Travel time data for testing the algorithms was obtained by simulating a 12-km section of the collector facility of Highway 401 in Toronto, Canada. The results from the three AVI based AID algorithms are compared to the performance of a leading loop detector based algorithm, which was independently tested on similar simulated data. The AID performance results indicate that AVI based AID can provide similar incident detection performance as existing loop detector based AID methods. INTRODUCTION Most urban freeways throughout North America are heavily utilised and experience ever increasing congestion during the peak commuting periods. Recurrent congestion results from high traffic demands and limited roadway capacity. Non-recurring congestion results from the occurrence of unexpected events (incidents) such as collisions, stalled vehicles, or B. Hellinga and G. Knapp 2 material spills. The U.S. Federal Highway Administration estimates that approximately 60% of travel time lost to congestion is a result of incidents and the percentage is believed to be increasing (Lindley, 1987). The early detection of incident events minimises the delay experienced by drivers, wasted fuel, emissions, and lost productivity, and also reduces the likelihood of secondary collisions. The goal of automatic incident detection (AID) is to minimise the human requirements in the efficient and effective detection of incident events. The emergence of automatic vehicle identification (AVI) technology has provided a new and previously unavailable form of real-time traffic data, namely individual vehicle travel times. This paper examines three freeway AID algorithms that rely on vehicle travel time data obtained from AVI equipped vehicles. The performance of these algorithms is compared to a leading conventional AID algorithm that relies on data obtained from in-road inductive loop detectors. Nomenclature i 20-second interval j Road segment reference (section of roadway between 2 AVI antennae) t time of day τti segment travel time reported by an AVI equipped vehicle at time t during interval i ni number of AVI equipped vehicle reports received during interval i δ duration of the comparison window nδ number of intervals within comparison window of duration δ i τ mean interval travel time for all AVI equipped vehicles in interval i δ τ mean of all mean interval travel times i τ in comparison window δ var variance of all mean interval travel times i τ in comparison window δ σ log-normal variance of i τ in comparison window μδ log-normal mean of i τ in comparison window z z value associated with the level of confidence ULi upper confidence limit for the mean travel time for interval i B. Hellinga and G. Knapp 3 Structure of the Paper The following section describes the network and the data that were used for testing the proposed AVI algorithms. This is followed by a description of the three proposed algorithms. Algorithm performance results are presented and compared to a conventional loop detector based AID algorithm. In the last section, conclusions are made and recommendations are provided. DATA FOR EVALUATING THE PROPOSED AVI BASED AID ALGORITHMS This section describes the simulated data and the different parameter values used to test and calibrate the three proposed AVI based AID algorithms. The data for testing the algorithms was generated using a simulation model because no AVI field data were available. The use of a simulation model also provides the following benefits: 1. Complete knowledge of true incident start and end times. 2. Control over the number, location, severity, and duration of incidents within the evaluation data set. 3. Ability to test algorithm performance for a range of level of market penetration (LMP) of AVI equipped vehicles. Network Description The network used in this study is modeled after eight interchanges along a 12-km freeway section of Highway 401 in Toronto, Canada. This facility experiences an average daily traffic flow of approximately 340,000 vehicles, making it one of the most heavily traveled freeways in North America. This freeway section includes an express facility and parallel collector facility. Initial simulation results on this network exhibited unrealistic congestion patterns, which were attributed to limitations in the model's route selection capabilities. Consequently, the network was modified to provide only a single route from any origin to any destination with the result that only the collector facility was modeled for this study. As illustrated in Figure 1, the eastbound and westbound freeway directions are both divided B. Hellinga and G. Knapp 4 into 10 segments approximately 1.2 km in length with AVI roadside antennas at both ends of each segment. The network was simulated using the Integration traffic simulation model (Van Aerde, 1998). The origin-destination traffic demand was constructed to replicate the build up of the AM peak from 5:30 AM to 10:30 AM. A total of 101,142 vehicle trips were simulated during this 5 hour time period. The network experiences severe recurring congestion at several locations during the simulation. This permits the testing of AID during both uncongested and congested traffic conditions. As an AVI equipped vehicle passes a roadside antenna the vehicle is uniquely identified through wireless communication between the vehicle’s transponder and the antenna. Since an AVI equipped vehicle can be uniquely identified, its travel time between antennas can be calculated. If a vehicle is not equipped with a transponder, the roadside antenna can not communicate, and no data can be collected for the vehicle. The simulation model permits tracking of the link travel times of individual vehicles. The model was run assuming travel times could be obtained for all vehicles. A post processor was developed to combine the individual link travel times for each vehicle to produce a travel time associated with each roadway segment between AVI antennas. Travel time reports were not created for vehicles that failed to pass the upstream antenna on a segment (i.e. vehicle entered the segment via an onramp downstream of the antenna) or failed to pass the downstream antenna on a segment (i.e. exited the freeway via an off-ramp upstream of the antenna). The resulting data sets provided individual segment travel times by time of day assuming all vehicles were equipped with AVI transponders. When the AVI AID algorithms were tested, samples of vehicles were selected from this data set according to the LMP assumed. This time-series of data was considered representative of the timeseries of data that would be received by a traffic management centre in real-time. Figure 2 illustrates typical individual vehicle segment travel time data obtained from the simulation for vehicles traversing the westbound direction of Highway 401 between Highway 400 and B. Hellinga and G. Knapp 5 Weston Road. These data also illustrate the impact on vehicle travel times of an incident that occurs on this segment from 7:30 AM to 7:40 AM. The second peak in travel times (between 8 and 9 AM) is a result of recurrent congestion. Incident Data In addition to the base non-incident case, twenty-four separate incident scenarios were simulated, resulting in a total of 125 hours of simulated traffic conditions. All the scenarios used the same network and O-D demand characteristics. However, each incident scenario included the modelling of 5 unique incidents, for a total of 120 simulated incidents. The key characteristics of these 120 incidents were varied, included incident location (20 locations), duration (5, 10, 20, and 30 minutes), time of day (60 during peak and 60 during off-peak), severity (100 single lane closures and 20 two-lane closures on roadways have three lanes), and traffic conditions. PROPOSED AVI BASED ALGORITHMS Three algorithms have been developed for examination in this paper, the Confidence Limit Algorithm, the Speed and Confidence Limit Algorithm, and the Dual Confidence Limit Algorithm. All algorithms are based on travel time data from AVI equipped vehicles. The algorithms can be considered to be statistical time-series models. The premise for all three models is that the travel time experienced by vehicles over a section of roadway increases more rapidly as a result of a change in capacity (i.e. such as the reduction in capacity that results from the occurrence of an incident) than it does as a result of a change in demand. Therefore, each of these algorithms attempts to characterise the mean and variance of the travel times associated with the traffic conditions prior to an incident. When an incident occurs, the traffic situation from which the travel times result, is changed and the statistical characteristics also change. Thus, the travel times resulting from traffic conditions prior to an incident can be thought of as belonging to one population, and those from traffic conditions after an incident has occurred, belonging to another population. The proposed algorithms attempt to determine if reported travel times are outside of the confidence limits associated with the current population, and if so, it is assumed that an incident has occurred. B. Hellinga and G. Knapp 6 For all three algorithms, the individual AVI travel time reports were aggregated over 20second time intervals. Aggregation was carried out to reflect practical implementation requirements. If aggregation were not carried out, the AID algorithm would need to be applied each time an AVI report was received. For the data illustrated in Figure 2, a total of 13,307 AVI reports would be received for
[1]
Xiao Liu,et al.
ARTIFICIAL NEURAL NETWORKS FOR FREEWAY INCIDENT DETECTION
,
1995
.
[2]
Edmond Chin-Ping Chang,et al.
IMPROVED FREEWAY INCIDENT DETECTION USING FUZZY SET THEORY
,
1994
.
[3]
Harold J. Paynet.
Freeway Incident-Detection Algorithms Based on Decision Trees With States
,
2022
.
[4]
J A Lindley,et al.
URBAN FREEWAY CONGESTION: QUANTIFICATION OF THE PROBLEM AND EFFECTIVENESS OF POTENTIAL SOLUTIONS
,
1987
.
[5]
Yong Shi,et al.
ON-LINE TESTING OF THE MCMASTER INCIDENT DETECTION ALGORITHM UNDER RECURRENT CONGESTION
,
1993
.
[6]
Fred L. Hall,et al.
DISTINGUISHING BETWEEN INCIDENT CONGESTION AND RECURRENT CONGESTION: A PROPOSED LOGIC
,
1989
.
[7]
Stephen G. Ritchie,et al.
Simulation Of Freeway Incident Detection Using Artificial Neural Networks
,
1993
.
[8]
Yorgos J. Stephanedes,et al.
SMOOTHING ALGORITHMS FOR INCIDENT DETECTION
,
1993
.