Predictive and Probabilistic Tracking to Detect Stopped Vehicles

Many techniques and models have been proposed for vehicles surveillance in highways. In the past, tracking algorithms based on Kalman filter have been largely used for their efficiency in the prediction and low computational cost. However, predictive filters cannot solve long-lasting occlusions. In this paper, we propose a new mixed predictive and probabilistic tracking that exploits the advantages of predictive filters for moving vehicles and adopts probabilistic and appearance-based tracking for stopped vehicles. The proposed tracking is part of a complete video surveillance system, oriented to control tunnels and highways from cluttered views, that is implemented in an embedded DSP platform and provides background suppression, a novel shadow detection algorithm, tracking, and scene recognition module. The experimental results are obtained over several hours of videos acquired in pre-existing platforms of CCTV surveillance systems