Statistical and neural classifiers to detect traffic operational problems on urban arterials

This paper proposes using artifical neural networks in a modular architecture to assist in detecting different types of operational problems on signalized urban arterials. A traffic surveillance infrastructure that includes inductive loop detectors on intersection approaches as well as mid-block system loops for traffic monitoring is used. For arterials, problems that require the attention of a traffic management center operator include lane-blocking incidents, special event conditions, and detector malfunctioning. Problem detection depends on factors such as operating conditions, configuration of sensors within the network, and block or link length. The feasibility of training and testing neural network models as components of a modular architecture, with an appropriate model for each sub-problem of pattern recognition, is demonstrated. The performance of this modular architecture exceeded that of any single architecture applied to the detection of different types of operational problems. The paper also reports on the performance of each type of model considered and the effect traffic flow levels and detector configuration have on the performance of the incident detection model. The results show that with the selection of a suitable architecture, the modular neural classifier outperforms alternative discriminant analysis-based classifiers. This is demonstrated using cyclic data from microscopic simulation and field data from Urban Traffic Control System implementations in the Cities of Los Angeles and Anaheim, California.