Driving Behavior Inference from Traffic Surveillance Data

Autonomous extraction of information from traffic video surveillance has received widespread attention in recent years due to the capability of analyzing massive data persis- tently. In this paper, we present a system for extracting vehicle information from traffic video sequence build a framework for defining driver behavior models. We detect the sedans and split them into two categories – taxi and otherwise, according to the color and size. The labeled cars from two categories form two exemplars of driver behavior – professional and standard. We use these two exemplars to compute the unknown parameters of the Hidden Markov Model (HMM) and demonstrate the quantitative discrimination of these two categories, which can be used for many applications of intelligent transportation system, such as abnormality detection and transportation planning. We also demonstrate applications which utilize the proposed model to detect aggressive drivers and professional drivers. The system is extensible with large surveillance databases for auto- mated detection of driver behavior and states in-situ for a variety of traffic event models. Furthermore, statistical information on driver behavior of different vehicle types, demographics and time periods for various locations of interest can be obtained by off-line processing of the surveillance data using the proposed model. The proposed framework can further be integrated with automated surveillance systems for guiding traffic controllers as well as city planners in better management of traffic.