Bayesian Filtering and Integral Image for Visual Tracking

This paper describes contributions to two problems related to visual tracking: control model design and observation process design. We describe the use of kernel-based Bayesian filtering for the tracking control procedure, and feature-based tracking to improve the observation process of tracking. In the kernelbased Bayesian filtering framework, the analytical representation of density functions by density interpolation and density approximation for the likelihood and the posterior contributes to efficient sampling. Feature-based tracking combines rectangular features with edge oriented histogram so that the combined features are robust to illumination changes, partial occlusion, and clutter while capturing the spatial information of the target. The use of integral image allows the features to be efficiently evaluated. The effectiveness of both algorithms are demonstrated by object tracking results on real videos.