A Robust Object Tracking Approach using Mean Shift

Background pixels in object model will increase localization error of object tracking, but in order to let the object contained in object model, it is inevitable to introduce some background pixels in object model. For reducing the localization error of object tracking, a straightforward approach is to omit the background pixels when the kernel histogram of object model is being computed, but there are many knotty problems for it. A weight parameter integrating background features is used in object model in this paper. The weight parameter indicates the similarity between background feature and object feature and can reduce localization error of object tracking. The experimental results show our approach has good localization precision of object tracking, and is robust against occlusion.