A Matching-Based Approach for Human Motion Analysis

This paper presents a novel approach to implement estimation and recognition of human motion from uncalibrated monocular video sequences. As it is difficult to find a good motion description for humans, we propose a matching scheme based on a local descriptor and a global descriptor, to detect individual body parts and analyze the shape of the whole body as well. In a frame-by-frame process, both descriptors are combined to implement the matching of the motion pattern and the body orientation. Moreover, we have added a novel spatial-temporal cost factor in the matching scheme which aims at increasing the temporal consistency and reliability of the description. We tested the algorithms on the CMU MoBo database with promising results. The method achieves the motion-type recognition and body-orientation classification at the accuracy of 95% and 98%, respectively. The system can be utilized for an effective human-motion analysis from a monocular video.