Real-time data fusion for arterial street incident detection using neural networks

This research contributes to the development of an automatic incident detection system for detecting traffic-delaying events on arterial street networks. Data describing current traffic conditions will be gathered in real-time from two distinct sources: inductive loop detectors and specially-equipped vehicles which measure and report their travel times on roadway links. Two approaches are considered for data fusion, the combination of information from these sources to produce a single decision about the presence or absence of incidents on each link. In one approach, integrated fusion, observed traffic data are combined directly using a neural network. In the other, algorithm output fusion, separate incident detection algorithms individually pre-process data from each source, reporting outputs which are combined using a neural network.; Data for calibrating these system components were generated using computer simulation. Several alternative neural network configurations were investigated for each approach, which varied input variety, use of data from previous time intervals and output format. The results were compared with each other and with approaches calibrated using discriminant analysis. The algorithm output fusion networks performed better than the other approaches. The best network, which used input from nearby links and from recent time periods, detected over ninety-three percent of the incidents with no false alarms, including many that were missed by both of the incident detection algorithms. Fusing algorithm outputs using neural networks was thus found to improve the capability provided by separate source incident detection algorithms operating alone. The importance of validating these results through calibration and testing with field data, as well as improving performance through introduction of an additional data source is discussed.