ARTIFICIAL NEURAL NETWORKS FOR FREEWAY INCIDENT DETECTION

Abstract

A freeway incident detection algorithm is developed using back propagation neural networks. Based on real-time occupancy and volume counts from pairs of adjacent loop detector stations, the network is trained with actual data, including 31 incidents from a typical freeway in the Twin Cities Metropolitan Area over the afternoon peak period during 72 days. Results indicate that the neural network, with about 1,000 connections, can learn the main characteristics of a variety of incidents. Algorithm performance, in terms of detection and false alarm rates, is superior to most of the best algorithms that have been tested with this data set.