Preprocessor Feature Extractor and Postprocessor Probabilistic Output Interpreter for Improved Freeway Incident Detection

Abstract

Automated incident detection algorithms usually utilize direct measurements of traffic volumes and occupancies as input and output binary indications of the traffic status for every time step. The transferability of such algorithms is often poor because of the extensive overlap between the incident-related cluster and the incident-free cluster of patterns in the input space. The observed overlap is not only significant but also location specific. The extent of the overlap is shown, its causes are discussed, and an input preprocessor is proposed that relies on modifying the input feature space in order to segregate the two clusters of patterns and reduce their overlap. The result is a more readily separable set of classes; easier classification of new patterns, which leads to better performance; and most important, less site-dependent performance and hence significantly better direct transferability of the detection algorithm with less need for retraining at new sites. On the output side, isolated binary algorithm outputs also have several drawbacks that are identified. An output postprocessor is proposed that links the isolated outputs using a Bayesian update process and converts the isolated outputs into a continuous probabilistic measure that is updated for every time step. The output at one interval is therefore used as a prior probability for the next. The preprocessor and the postprocessor proposed in this paper were developed as integrated components of a universally transferable freeway incident detection framework, UNITID, but are generally applicable to any similar incident detection system.