Robust tracking with spatial pyramid histogram

This paper presents a new method for object tracking based on global spatial correspondence with the geometric distribution of visual words. “Spatial Pyramid Histogram” - SPH is produced by partitioning the image into increasing sub-blocks and computing histograms of features found inside each sub-block. SIFT descriptors are extracted to represent the object to construct a visual dictionary. A classifier is applied to discriminate the target from a number of candidates generated by randomly sampling. Our method also provides a solution to update the dictionary and the spatial information of visual words through selecting most distinctive samples to retrain the classifier. The experiments demonstrate that our method can track objects accurately and robustly even with scaling, rotation, especially partial or severe occlusion.