Abstract

A major source of urban freeway delay in the United States is non-recurring congestion caused by incidents such as accidents, disabled vehicles, spilled loads, temporary maintenance and construction activities, signal and detector malfunctions, and other special and unusual events that disrupt the normal flow of traffic. The automated detection of freeway incidents is an important function of a freeway traffic management center. Early detection of incidents is vital for formulating effective response strategies such as timely dispatch of emergency services and incident removal crews, control and routing of traffic around the incident location, and provision of real-time traffic information to motorists. A number of incident detection algorithms, based on conventional approaches, have been developed over the past several decades, and a few of them are being deployed at urban freeway systems in major cities. These conventional algorithms have met with varying degree of success in their detection performance. In this research, a new incident detection technique based on an artificial neural network approach has been proposed. The objective of this research was to demonstrate the use of artificial neural network models for automated detection of lane-blocking incidents on urban freeways. The study focused on the application of neural network models in classifying traffic surveillance data obtained from inductive loop detectors, and the use of the classified output to detect an incident. Three types of neural network models were developed to detect lane-blocking incidents: the multi-layer feed-forward neural network, self-organizing feature map and adaptive resonance theory 2. The models were developed with simulation data from a study site and tested with both simulation and field data at the study site and other locations. The multi-layer feed-forward neural network was found to have the highest potential among the four models to achieve a better incident detection performance. This network consistently detected most of the lane-blocking incidents and gave a false alarm rate lower than the conventional algorithms currently in use. The results have demonstrated the potential of artificial neural network models in improving incident detection performance over currently available techniques.