Development and Evaluation of Neural Network Model for Incident Detection on Urban Arterials using Simulated Database

Abstract

Incident detection in urban arterial situation is more difficult than the similar job in freeway situation because of the presence of traffic signals and other intersections with associated recurrent queue. Most of the earlier automatic incident detection algorithms address mainly freeway situation. This study aims at development, calibration, validation and testing of an ANN model for incident detection in Kuala Lumpur (KL) arterials using simulated incident database. Database for the study is generated under incident and no- incident condition by simulating the traffic flow through the arterial network of Golden Triangle area of KL. Calibration efforts are divided into different tasks such as the time interval for input neuron, recalculation interval, location of the detector and the threshold values for the model. The calibrated model for optimum location of detector yields 98.5% of detection rate and 2.9% of false alarm rate for normal traffic demand situation. It is found that in case of link longer than 350 m data from two detectors are required for better performance of the ANN model but a single detector data is good enough for link length of around 220 m or less. Testing of the model on other link sites also yields similar results with more accurate detection in case of shorter links. While one cycle time was found to be long enough as a recalculation interval, further sensitivity analysis on this revealed that lower cycle time of around 60 s degrades the performance of the model in terms of false alarm rate. The results from this study provide useful insights for the design of AID system in urban arterials.