Toward Training-Free Automatic Detection of Freeway Incidents

Abstract

Automatic incident detection (AID) on freeways has been investigated extensively in the past four decades. Various algorithms covering a broad range of types in regard to complexity, data requirements, and efficiency have been published in the literature. However, a recent nationwide survey concluded that the implementation of AID algorithms in traffic management centers was still limited. The main reasons for this discrepancy were high rates of false alarms and calibration complexity. The main objective of this research was to develop a simple, transferable algorithm that might dispense with training on a preexisting data set. The dynamic thresholds of the proposed algorithm were based on historical traffic data and thus accounted for typical variations of traffic throughout the day. Therefore, this approach could recognize recurrent congestion and therefore reduce the incidence of false alarms. In addition, the proposed method required no human intervention; that factor certainly encouraged its implementation. The presented model was evaluated in a newly developed incident database, which contained 40 incidents. The model performed better than existing algorithms in the literature.