A SVM kernel for classifying partially occluded images

We propose a novel SVM (Support Vector Machine) kernel for classifying partially occluded images in the process of video tracking. The SVM kernel (called Bhattacharyya kernel) is derived from Bhattacharyya coefficient. In our study, the validity of Bhattacharyya kernel is proven. We use kernel density estimation of histogram as SVM´s feature space. Experiments show the SVM based on Bhattacharyya kernel can keep high classification accuracy when occlusion or clutter of peripheral pixels appears. Bhattacharyya kernel can be generalized easily when using other features.