Life-long Learning System of Driving Behaviors from Vehicle Data Streams

Vehicle collision is an important issue of Intelligent Transportation Systems (ITS), specially at the end-of-queue of traffic congestion. To anticipate potential collisions, many researches focus on the prediction of traffic status and vehicle behaviors such as the future vehicle speed profile. The state of the art shows that this direction can encounter some challenges relating to the reliability of long-term predictions and the significance of predictive information meaning whether the provided information is sufficient to represent traffic dynamics. In our study, we propose to ADRIP (Adaptive multi-agent system for DRIving behaviors Prediction), based on multi-agent system to predict driving behaviors using vehicle data streams. In the scope of our study, driving behavior on a road segment is modeled by the Mobility Profile which is deduced from vehicle speed sequences. To allow long-term predictions, ADRIP first implements a life-long learning strategy of driving behaviors which is the focus of this paper. Our strategy investigates dynamic clustering for data stream. Obtained results underline its adequacy to handle the traffic problem.