Human Action Recognition with Pose Similarity

This paper presents a method for representing and recognizing human actions based on pose similarity. For pose representation, we extend Histogram of Oriented Gradients (HOG) with directional statistics to obtain a HOG based descriptor with a smaller dimension. Then a directional similarity measurement for the proposed descriptor is put forward to provide a measure consistent with human perception. To recognize human actions, each testing frame is classified with Nearest Neighbor classifier using the similarity measurement, and each testing sequence of frames is classified with an equal weight voting scheme. Detailed illustration and analysis on HOG with directional statistics are given to show that the proposed descriptor and similarity measurement are reasonable. Experiments on the WEIZMANN dataset demonstrate that with proper similarity measurement, very simple and direct method of human action recognition can achieve desirable performance.