Towards universal freeway incident detection algorithms

This paper reports the intensive test of the new transport systems centre (TSC) algorithm applied to incident detection on freeways. The TSC algorithm is designed to fulfil the universality expectations of automated incident detection. The algorithm consists of two modules: data processing module and incident detection module. The data processing module is designed to handle specific features of different sites. The Bayesian network based incident detection module is used to store and manage general expert traffic knowledge, and to perform coherent reasoning to detect incidents. The TSC algorithm is tested using 100 field incident data sets obtained from Tullamarine Freeway and South Eastern Freeway in Melbourne, Australia. The performance of the algorithm demonstrates its competitiveness with the best performing neural network algorithm which was developed and tested using the same incident data sets in an early research. Most importantly, both the detection rate and false alarm rate of the TSC algorithm are not sensitive to the incident decision threshold, which greatly improves the stability of incident detection. In addition, a very consistent algorithm performance is achieved when the TSC algorithm is transferred from Southern Expressway of Adelaide to both Tullamarine Freeway and South Eastern Freeway of Melbourne. No substantial algorithm retraining is required. A significant step towards algorithm universality is possible from this research.     

Application of wavelet technique to freeway incident detection

This paper presents an application of the wavelet technique to freeway incident detection because wavelet techniques have demonstrated superior performance in detecting changes in signals in electrical engineering. Unlike the existing wavelet incident detection algorithm, where the wavelet technique is utilized to denoise data before the data is input into an algorithm, this paper presents a different approach in the application of the wavelet technique to incident detection. In this approach, the features that are extracted from traffic measurements by using wavelet transformation are directly utilized in detecting changes in traffic flow. It is shown in the paper that the extracted features from traffic measurements in incident conditions are significantly different from those in normal conditions. This characteristic of the wavelet technique was used in developing the wavelet incident detection algorithm in this study. The algorithm was evaluated in comparison with the multi-layer feed-forward neural network, probabilistic neural network, radial basis function neural network, California and low-pass filtering algorithms. The test results indicate that the wavelet incident detection algorithm performs better than other algorithms, demonstrating its potential for practical application.     

Discrete dynamic models for freeway incident detection systems

This paper considers the problem of freeway incident detection within the general framework of computer‐based freeway surveillance and control. A new approach to the detection of freeway traffic incidents is presented based on a discrete‐time stochastic model of the form ARIMA (0, 1, 3) that describes the dynamics of traffic occupancy observations. This approach utilizes real‐time estimates of the variability in traffic occupancies as detection thresholds, thus eliminating the need for threshold calibration and lessening the problem of false‐alarms. Because the moving average parameters of the ARIMA (0, 1, 3) model change over time, these parameters can be updated occasionally. The performance of the developed detection algorithm has been evaluated in terms of detection rate, false‐alarm rate, and average time‐lag to detection, using a total of 1692 minutes of occupancy observations recorded during 50 representative traffic incidents.     

Development and adaptation of constructive probabilistic neural network in freeway incident detection

This paper investigates the use of constructive probabilistic neural network (CPNN) in freeway incident detection, including model development and adaptation. The CPNN was structured based on mixture Gaussian model and trained by a dynamic decay adjustment algorithm. The model was first trained and evaluated on a simulated incident database in Singapore. The adaptation of CPNN on the I-880 freeway in California was then investigated in both on-line and off-line environments. This paper also compares the performance of the CPNN model with a basic probabilistic neural network (BPNN) model. The results show that CPNN has three main advantages over BPNN: (1) CPNN has clustering ability and therefore could achieve similarly good incident-detection performance with a much smaller network size; (2) each Gaussian component in CPNN has its own smoothing parameter that can be obtained by the dynamic decay adjustment algorithm with a few epochs of training; and (3) the CPNN adaptation methods have the ability to prune obsolete Gaussian components and therefore the size of the network is always within control. CPNN has shown to have better application potentials than BPNN in this research.     

A new methodology for evaluating incident detection algorithms

We present a novel, off-line approach for evaluating incident detection algorithms. Previous evaluations have focused on determining the detection rate versus false alarm rate curve––a process which we argue is inherently fraught with difficulties. Instead, we propose a cost-benefit analysis where cost mimics the real costs of implementing the algorithm and benefit is in terms of reduction in congestion. We argue that these quantities are of more practical interest than the traditional rates. Moreover, these costs, estimated on training data, can be used both as a mechanism to fine tune a single algorithm as well as a meaningful quantity for direct comparisons between different types of incident detection algorithms. We demonstrate our approach with a detailed example.     

Enhancing the universality and transferability of freeway incident detection using a Bayesian-based neural network

Development of a universal freeway incident detection algorithm is a task that remains unfulfilled despite the promising approaches that have been recently explored. Incident detection researchers are realizing that an operationally successful detection framework needs to fulfill a full set of recognized needs. In this paper we attempt to define one possible set of universality requirements. Among the set of requirements, a freeway incident detection algorithm needs to be operationally accurate and transferable. Guided by the envisioned requirements, we introduce a new algorithm with potential for enhanced performance. The algorithm is a modified form of the Bayesian-based Probabilistic Neural Network (PNN) that utilizes the concept of statistical distance. The paper is divided into three main sections. The first section is a detailed definition of the attributes and capabilities that a potentially universal freeway incident detection framework should possess. The second section discusses the training and testing of the PNN. In the third section, we evaluate the PNN relative to the universality template previously defined. In addition to a large set of simulated incidents, we utilize a fairly large real incident databases from the I-880 freeway in California and the I-35W in Minnesota to comparatively evaluate the performance and transferability of different algorithms, including the PNN. Experimental results indicate that the new PNN-based algorithm is competitive with the Multi Layer Feed Forward (MLF) architecture, which was found in previous studies to yield superior incident detection performance, while being significantly faster to train. In addition, results also point to the possibility of utilizing the real-time learning capability of this new architecture to produce a transferable incident detection algorithm without the need for explicit off-line retraining in the new site. In this respect, and unlike existing algorithms, the PNN has been found to markedly improve in performance with time in service as it retrains itself on captured incident data, verified by the Traffic Management Center (TMC) operator. Moreover, the overall PNN-based framework promises potential enhancements towards the envisioned universality requirements.     

An application of the CUSUM algorithm to freeway incident detection based on two contiguous detectors

The CUSUM (cumulative sum of log‐likelihood ratio) algorithm is a detection algorithm that shows potential for the improvement of incident detection algorithms because it is designed to minimize the mean detect delay for a given false alarm constraint and it can also detect changes with different patterns. In this study, the CUSUM algorithm was applied to freeway incident detection by integrating traffic measurements from two contiguous loop detectors and the non‐stationarity of traffic flows. The developed algorithm was tested based on incident data from the PATH program, with consideration given to the impact of different geometric conditions on algorithm performance. It was also compared with two existing algorithms, in order to address the influence of traffic patterns. The evaluation results show that the CUSUM incident detection algorithm can perform equally well in comparison with the selected algorithms.     

Automated detection of lane-blocking freeway incidents using artificial neural networks

A major source of urban freeway delay in the U.S. is non-recurring congestion caused by incidents. The automated detection of incidents is an important function of a freeway traffic management center. A number of incident detection algorithms, using inductive loop data as input, have been developed over the past several decades, and a few of them are being deployed at urban freeway systems in major cities. These algorithms have shown varying degrees of success in their detection performance. In this paper, we present a new incident detection technique based on artificial neural networks (ANNs). Three types of neural network models, namely the multi-layer feedforward (MLF), the self-organizing feature map (SOFM) and adaptive resonance theory 2 (ART2), were developed to classify traffic surveillance data obtained from loop detectors, with the objective of using the classified output to detect lane-blocking freeway incidents. The models were developed with simulation data from a study site and tested with both simulation and field data at the same site. The MLF was found to have the highest potential, among the three ANNs, to achieve a better incident detection performance. The MLF was also tested with limited field data collected from three other freeway locations to explore its transferability. Our results and analyzes with data from the study site as well as the three test sites have shown that the MLF consistently detected most of the lane-blocking incidents and typically gave a false alarm rate lower than the California, McMaster and Minnesota algorithms currently in use.     

Fuzzy-logic-based incident detection for signalized diamond interchanges

This paper documents a fuzzy-logic-based incident detection algorithm for signalized urban diamond interchanges. The model is capable of detecting lane-blocking incidents whose effects are manifested by patterns of deterioration in traffic conditions that require adjustments in signal control strategies. As a component of a real-time traffic adaptive control system for signalized diamond interchanges, the algorithm feeds an incident report (i.e., the time, location, and severity of the incident) to the system's optimization manager, which uses that information to determine the appropriate signal control strategy.The performance of the model was studied using a simulation of an actual diamond interchange. The simulation study evaluated the model's performance in terms of detection rate, false alarm rate, and mean time to detect. The model's performance was encouraging, and the fuzzy-logic-based approach is considered promising.     

Simulator evaluation of incident detection by transponder-equipped vehicles

This paper explores the efficacy of the driver-based incident detection using the vehicle-to-roadside communication (VRC) system. The proliferation of vehicle tags in the US for automatic toll collection, traffic monitoring, and vehicle navigation and information systems has created an infrastructure capable of supporting a driver-based incident detection system. The research reported herein investigated the use of "activatable" vehicle tags by drivers to send an incident signal to the Traffic Management Center through VRC reader stations spaced uniformly on a highway. The simulation results showed that good detection performance was achieved even at lower levels of market penetration of vehicle tags. The results further showed that detection performance is significantly affected by the severity of the incident in terms of number of lanes closed, the spacing of the VRC reader stations, traffic volume at the time of the incident, and the reporting propensity of the traveling public.The performance of the VRC-based incident reporting system was compared to the performance of two incident detection algorithms that rely on traffic data collected through the automatic vehicle identification (AVI) system. The comparison showed that the VRC-based incident reporting system attained shorter detection times and higher detection rates under fairly similar simulated conditions. The paper also discusses issues that need further study through simulation and field experimentation of the VRC-based incident reporting system.     

Distributed surveillance on freeways emphasizing incident detection and verification

Efficient freeway management requires continuous decision-making based on conditions on the network and an understanding of the impacts of the decisions made. These conditions are usually measured with fixed-point surveillance systems, most of which are deployed in such a manner as to require communication links that are always connected and are polled at regular intervals. All of the sensor data are typically sent to a Traffic Management Center (TMC) for assessment, yet most of the time no action is taken in response to the data, leading to unnecessarily high communication costs.To reduce communication costs without a significant loss in the quality of information received at the TMC this paper lays the foundation for an event driven communication system by examining the sequence of events at the detector stations in the context of incident detection. Where, following the broadest convention on freeways, an incident is any non-recurring event that causes a temporary bottleneck and restricts flow. Although the focus is incident detection, the proposed communication system could easily support many other applications that use aggregate data, e.g., measuring average annual daily travel (AADT). The methodology is generalizable to most common freeway geometries and care is taken in the paper to specifically address the situation where an incident interacts with a recurring bottleneck.To address the normally high communications costs, a portion of the decision-making process is transferred from the TMC to the field controllers, which would make the initial evaluation of conditions and only send data that might elicit a control response or benefit comparative decisions between detector stations. In other words, rather than relying on the conventional, centrally polled communication system, these events could be used to initiate communication from the field when the potential value outweighs the cost per communication. The process could also lead to better data handling for decision-making or archiving in a conventional, polled communication system as well. We develop the methodology by deconstructing several incidents on a freeway and identify the observable events at a pair of detector stations that may be upstream, downstream or straddle the incident. This analytical process could be repeated for any other condition of interest.     

Examination of factors affecting freeway incident clearance times: a comparison of the generalized F model and several alternative nested models

Traffic incidents are a principal cause of congestion on urban freeways, reducing capacity and creating risks for both involved motorists and incident response personnel. As incident durations increase, the risk of secondary incidents or crashes also becomes problematic. In response to these issues, many road agencies in metropolitan areas have initiated incident management programs aimed at detecting, responding to, and clearing incidents to restore freeways to full capacity as quickly and safely as possible. This study examined those factors that impact the time required by the Michigan Department of Transportation Freeway Courtesy Patrol to clear incidents that occurred on the southeastern Michigan freeway network. These models were developed using traffic flow data, roadway geometry information, and an extensive incident inventory database. A series of parametric hazard duration models were developed, each assuming a different underlying probability distribution for the hazard function. Although each modeling framework provided results that were similar in terms of the direction of factor effects, there was significant variability in terms of the estimated magnitude of these impacts. The generalized F distribution was shown to provide the best fit to the incident clearance time data, and the use of poorer fitting distributions was shown to result in severe over‐estimation or under‐estimation of factor effects. Those factors that were found to impact incident clearance times included the time of day and month when the incident occurred, the geometric and traffic characteristics of the freeway segment, and the characteristics of each incident. Copyright © 2012 John Wiley & Sons, Ltd.     

A fuzzy-based system for incident detection in urban street networks

Detecting incidents on urban streets or arterials using loop detector data is quite challenging. The pattern of the incident could be quite similar to non-incident cases as intersections get congested. This paper describes the development of a fuzzy logic for incident detection. An Integrated System for Incident Management (

-sim) was developed. An integral component of such system is a microscopic simulator,

-sim-s, an object-oriented model that allows for virtual detector installations at different locations, modeling different intersection layouts, traffic control types and timing, and link characteristics.

-sim-s was utilized to generate various incident scenarios and extracting associated detectors’ accumulative counts. A data clustering technique was utilized to consolidate the various incident scenarios into a single data set for the development of the Fuzzy Logic for incident detection at intersections (

-sim-fl). The

-sim-fl uses the detector data as well as other link properties in flagging detecting incidents.The

-sim-fl can be used to indicate the possibility of an incident, a stalled vehicle, or a sort of traffic disturbance. The devised logic was validated using separate simulation-based incident scenarios.     

Simulation assessment of incident detection by cellular phone call-in programs

This research study was designed to assess by simulation the efficacy of incident detection by cellular phone call-in programs. The assessment was conducted by varying the proportion of drivers with cellular phones on the highway so as to mirror the cellular industry statistics that show a continued growth of ownership of cellular phones in the United States. An analytical model, which combined simulation and the limited field data available in the literature, was used to determine measures of effectiveness of the cellular phone-based detection system. The results showed that a cellular phone detection system offers fast incident detection times and higher detection rates for both shoulder and lane blocking incidents. For example, in moderate traffic flow (i.e. 1,550 vehicles per hour per lane), 90 percent of incidents blocking two lanes were detected in 1.5 minutes when the proportion of drivers with cellular phones was one out of 10 drivers, even with only 20 percent of them willing to report incidents. When the current proportion of cellular ownership, i.e. 1 out of 3, was used in the simulation, the detection time improved to 0.8 minutes. The simulation analysis of incident detection by cellular phones also showed that there is a direct relationship between the probability of detection and the detection time; that is, the specification of a higher detection rate resulted in slower detection times. This is in sharp contrast with the results of field study of automatic incident detection (AID) systems which demonstrated an inverse relationship between probability of detection and detection time. This revised version was published online in June 2006 with corrections to the Cover Date.     

Effective arterial road incident detection: A Bayesian network based algorithm

Timely and accurate incident detection is an essential part of any successful advanced traffic management system. The complex nature of arterial road traffic makes automated incident detection a real challenge. Stable performance and strong transferability remain major issues concerning the existing incident detection algorithms. A new arterial road incident detection algorithm TSC_ar is presented in this paper. In this algorithm, Bayesian networks are used to quantitatively model the causal dependencies between traffic events (e.g. incident) and traffic parameters. Using real time traffic data as evidence, the Bayesian networks update the incident probability at each detection interval through two-way inference. An incident alarm is issued when the estimated incident probability exceeds the predefined decision threshold. The Bayesian networks allow us to subjectively build existing traffic knowledge into their conditional probability tables, which makes the knowledge base for incident detection robust and dynamic. Meanwhile, we incorporate intersection traffic signals into traffic data processing. A total of 40 different types of arterial road incidents are simulated to test the performance of the algorithm. The high detection rate of 88% is obtained while the false alarm rate of the algorithm is maintained as low as 0.62%. Most importantly, it is found that both the detection rate and false alarm rate are not sensitive to the incident decision thresholds. This is the unique feature of the TSC_ar algorithm, which suggests that the Bayesian network approach is advanced in enabling effective arterial road incident detection.     

A new methodology for incident detection and characterization on surface streets

In this paper, a new methodology is presented for real-time detection and characterization of incidents on surface streets. The proposed automatic incident detection approach is capable of detecting incidents promptly as well as characterizing incidents in terms of time-varying lane-changing fractions and queue lengths in blocked lanes, lanes blocked due to incidents, and incident duration. The architecture of the proposed incident detection approach consists of three sequential procedures: (1) Symptom Identification for identification of incident symptoms, (2) Signal Processing for real-time prediction of incident-related lane traffic characteristics and (3) Pattern Recognition for incident recognition. Lane traffic counts and occupancy are the only two major types of input data, which can be readily collected from point detectors. The primary techniques utilized in this paper include: (1) a discrete-time, nonlinear, stochastic system with boundary constraints to predict real-time lane-changing fractions and queue lengths and (2) a pattern-recognition-based algorithm employing modified sequential probability ratio tests (MSPRT) to detect incidents. Off-line tests based on simulated as well as video-based real data were conducted to assess the performance of the proposed algorithm. The test results have indicated the feasibility of achieving real-time incident detection using the proposed methodology.     

Algorithm fusion method to enhance automatic incident detection on Melbourne freeways

This paper addresses the transferability issue faced by many practitioners in developing an effective and efficient automatic incident detection algorithm for different freeways. An algorithm fusion procedure developed for the Central Expressway in Singapore is evaluated to demonstrate its transferability potential in detecting lane-blocking incidents along freeways in Melbourne, Australia. This study observes that the flow-based algorithm fusion options that use a set of different detection threshold values for various pre-incident traffic flow conditions possess promising transferability potential. They give a reasonably high detection rate of above 80% with false alarm rate levels below 0.2% with mean-time-to-detect values less than 150 seconds. These flow-based algorithm fusion options significantly outperform a model specifically developed for traffic conditions on freeways in Melbourne. In conclusion, this method is capable of providing an alternative to the commonly practiced methods in detecting incidents along different sites.     

From Twitter to detector: Real-time traffic incident detection using social media data

The effectiveness of traditional incident detection is often limited by sparse sensor coverage, and reporting incidents to emergency response systems is labor-intensive. We propose to mine tweet texts to extract incident information on both highways and arterials as an efficient and cost-effective alternative to existing data sources. This paper presents a methodology to crawl, process and filter tweets that are accessible by the public for free. Tweets are acquired from Twitter using the REST API in real time. The process of adaptive data acquisition establishes a dictionary of important keywords and their combinations that can imply traffic incidents (TI). A tweet is then mapped into a high dimensional binary vector in a feature space formed by the dictionary, and classified into either TI related or not. All the TI tweets are then geocoded to determine their locations, and further classified into one of the five incident categories.We apply the methodology in two regions, the Pittsburgh and Philadelphia Metropolitan Areas. Overall, mining tweets holds great potentials to complement existing traffic incident data in a very cheap way. A small sample of tweets acquired from the Twitter API cover most of the incidents reported in the existing data set, and additional incidents can be identified through analyzing tweets text. Twitter also provides ample additional information with a reasonable coverage on arterials. A tweet that is related to TI and geocodable accounts for approximately 5% of all the acquired tweets. Of those geocodable TI tweets, 60–70% are posted by influential users (IU), namely public Twitter accounts mostly owned by public agencies and media, while the rest is contributed by individual users. There is more incident information provided by Twitter on weekends than on weekdays. Within the same day, both individuals and IUs tend to report incidents more frequently during the day time than at night, especially during traffic peak hours. Individual tweets are more likely to report incidents near the center of a city, and the volume of information significantly decays outwards from the center.     

Efficient multiple model particle filtering for joint traffic state estimation and incident detection

This article proposes an efficient multiple model particle filter (EMMPF) to solve the problems of traffic state estimation and incident detection, which requires significantly less computation time compared to existing multiple model nonlinear filters. To incorporate the on ramps and off ramps on the highway, junction solvers for a traffic flow model with incident dynamics are developed. The effectiveness of the proposed EMMPF is assessed using a benchmark hybrid state estimation problem, and using synthetic traffic data generated by a micro-simulation software. Then, the traffic estimation framework is implemented using field data collected on Interstate 880 in California. The results show the EMMPF is capable of estimating the traffic state and detecting incidents and requires an order of magnitude less computation time compared to existing algorithms, especially when the hybrid system has a large number of rare models.     

Safety aspects of freeway weaving sections

One source of vehicle conflict is the freeway weaving section, where a merge and diverge in close proximity require vehicles either entering or exiting the freeway to execute one or more lane changes. Using accident data for a portion of Southern California, we examined accidents that occurred on three types of weaving sections defined in traffic engineering: Type A, where every merging or diverging vehicle must execute one lane change, Type B, where either merging or diverging can be done without changing lanes, and Type C, where one maneuver requires at least two lane changes. We found no difference among these three types in terms of overall accident rates for 55 weaving sections over one year (1998). However, there were significant differences in terms of the types of accidents that occur within these types in terms of severity, and location of the primary collision, the factors causing the accident, and the time period in which the accident is most likely to occur. These differences in aspects of safety lead to implications for traffic engineering improvements.