Detection of traffic anomalies using fuzzy logic based techniques

Traffic incident detection involves both the collection and analysis of traffic data. The paper discusses the development of a novel time-indexed traffic anomaly detection algorithm. A unique partition of time into the "type of day", and "time of day" is performed. Using this partition, a novel fuzzy neuromorphic unsupervised learning algorithm is used to calibrate the "normal" and "abnormal" for each descriptor. Fuzzy composition techniques are used, on a per lane basis, to fuse multiple traffic descriptors in order to determine membership in "normal" or "abnormal" lane status. Then, each lane status is fused to determine an over all road segment status. Initial training of the algorithm takes place during the first few weeks after the sensor is installed. Online background training continues thereafter to continually tune and track seasonal changes.