The main obstacle to the robustness of car tracking is large shadows of vehicles. Even with a good foreground model the tracking process is liable to be disrupted by the shadows. This paper proposes a new probabilistic background model which is capable of modeling shadow as well as foreground and background regions. Unlike many other background models it is no longer necessary to select the training data since the distributions for different regions can be learnt from a mixed video sequence. This background model functions as a low level process for a car tracker. An observation density based on the output of the model is presented. The use of a particle filter in the car tracker makes it possible to feed the information of the low level process into a high-level process via importance sampling. The effectiveness of both the low level process and the observation likelihood are demonstrated.
展开▼