Histogram of oriented gradients and submanifold applications in object tracking

In order to track objects steadily when illumination or pose changes, a novel approach based on histogram of oriented gradients (HOG) and submanifold was proposed. Firstly, regions of objects were divided into sub-regions to obtain HOG features separately; then features of sub-regions were combined and mapped into submanifold space using locality preserving projection (LPP). Integral histogram was used to accelerate feature extraction. In the process of tracking, firstly, features of samples in submanifold space were trained off line, then, particle filter was used to accomplish tracking, where similarity was measured in submanifold space by distances between features of particles and the mean of training samples. Experimental results show that the proposed approach is effective and robust when illumination, scale and pose of objects change and objects are occluded.