Visual Vehicle Tracking Based on Conditional Random Fields

This paper proposes an approach to moving vehicle tracking in surveillance videos based on conditional random fields (CRF). The key idea is to integrate a variety of relevant knowledge about vehicle tracking into a uniform probabilistic framework by using the CRF model. In this work, the CRF model integrates spatial and temporal contextual information of vehicle motion, and the appearance information of the vehicle. An approximate inference algorithm, loopy belief propagation, is used to recursively estimate the vehicle region from the history of observed images. Moreover, the background model is updated adaptively to cope with non-stationary background processes. Experimental results show that the proposed approach is able to accurately track moving vehicles in monocular image sequences. Besides, region-level tracking realizes precise localization of vehicles.