Human action classification based on combinational features from video

Classifying accurately human action is a fundamental problem in computer vision. The goal of human action classification is to automatically detect and analyse actions from an unknown video for the sake of studying the human behaviors in specific applications. However, it is a challenging task for the computer to achieve robust action classification due to illumination changes, camera motion, background variety, occlusion and the variations of people and actions. Our motivation in studying the problem is to propose an effective and faithful method of human action classification, and use it to build the realtime surveillance system with classification and activity analysis capabilities. We present an approach to represent videos by novel multifeatures. In this framework, features of the human silhouette are extracted from the video sequences, and motion features of the human also are calculated. Furthmore, an optical flow is employed to demonstrate the presence and direction of motion, and then these features are combined. In the final step, the SVMs classifier is used to classify actions of human into pre-defined classes like bending, walking, jacking, jumping, running and skipping. Experimental results show a significant improvement over KIH and WEIZMANN.