Mining traffic data for road incidents detection

Tackling urban road congestion by means of ITS technologies, involves a number of key challenges. One such challenge is related to the accurate detection of traffic incidents in urban networks for more efficient traffic management. This paper introduces a classification approach that achieves accurate detection of road traffic incidents, based on data retrieved from inductive-loop detectors. In the core of the proposed approach lies a more efficient feature extraction technique, based on the dynamic characteristics of data corresponding to those vehicles that are involved in incidents. Our work observes how dynamic aspects of measured data can be exploited for extracting features that result in measurable improvement of the incident detection rate by the application of a Support Vector Machines classification approach. The latter is known to be one of the most precise solutions that have been widely applied up to now for dealing with relevant incident detection problems. In this paper we conduct appropriate experimental evaluation, including comparison to a set of well known techniques, for assessing the impact of the proposed feature selection technique on the accuracy of the detection rate. The evaluation results show that our approach manages to realize a more accurate and faster incident detection mechanism, although it has a low impact on the improvement of the false alarm rate. We also describe how the proposed approach can be applied to a generic context of ITS for efficient urban traffic management.