Neuro-fuzzy learning for automated incident detection

Traffic incidents such as vehicle accidents, weather and construction works are a major cause of congestion. Incident detection is thus an important function in freeway and arterial traffic management systems. Most of the large scale and operational incident detection systems make use of data collected from inductive loop detectors. Several new approaches, such as probe vehicles and video image processing tools, have recently been proposed and demonstrated. This research aims at model development for automatic incident detection and travel time estimation employing neuro-fuzzy techniques. As a first step, in this paper we develop an initial model for incident detection using a standard neuro-fuzzy algorithm. In subsequent development we propose a model where the fuzzy rules are themselves extracted from the data using an associative data mining algorithm. The results of the initial experiments suggest that the proposed model has plausible incident detection rates and low false alarm rates. The test results also suggest that the proposed model enhances accuracy of incident detection in an arterial road and can be expected to contribute to formal traffic policy.