Neural network representations for arterial street incident detection data fusion

Abstract

This research considers four neural network representations for detecting incidents on signalized arterials using multiple data sources. Two incident detection algorithms process unique data sources separately: inductive loop detectors, and travel times collected from vehicle probes travelling through the street network. The networks then combine the algorithm inferences about traffic conditions to identify highway links on which incidents are occurring. The four networks consider the following input and structure representations, added incrementally: (1) the two algorithm output values alone; (2) a weighted geometric sum of previous network output values; (3) algorithm scores from links immediately upstream and downstream of the subject link; and (4) weighted geometric sums of previous input values. The four representations were trained as feed-forward networks using error back propagation. Time series inputs were represented with extra processing units and fixed weight connections. The networks were trained with data generated by traffic simulation permitting deliberate control of traffic demand, operation and incident conditions. Each network was trained until performance began to degrade on a reserved data set not used for training. Adding the output time series permitted two of the 24 incidents to be detected sooner than with the network that did not include this input. Similarly, using information from adjacent links in time series permitted all of the incidents to be detected by at least the third time period.