Gesture Detection Based On The Hidden Markov Model

In highly informationalized modern society, video surveillance has been widely used in every corner of life. Currently, surveillance is mostly based on traditional record based methods. Some key areas still need human watchers to stay before monitors. Obviously, this is an inefficient and expensive way for surveillance with the monitoring data growing in a blowout way. Along with the rapid development of computer vision and image processing methods, intelligent video surveillance techniques have become the research hotspot. In intelligent video surveillance, how to recognize some specific gesture actions is an important issue that with both theoretic and practical value. Therefore, gesture recognition is one of the main research directions for intelligent video surveillance. Because the monitored scene is often very complex, an effective gesture detection scheme is crucial to obtain satisfactory gesture detection results.Inspired by the success of hidden Markov model(HMM) in the field of language recognition, in this thesis a HMM based gesture detection approach is developed. In the proposed framework, the entire detection flow includes following main processing steps: foreground detection, foreground subtraction, target tracking, feature extraction and gesture chain recognition.Based on the characteristic that body gestures are continuous in temporal-spatial space, in the proposed approach the temporal-spatial information of the gesture process is selected as features for body gesture analyzing. The body gesture can be effectively described by the temporal-spatial feature analyzing method. For the temporal information of the body gesture between two consecutive frames, the optical flow algorithm is used to extract the body motion information. The distribution of amplitude and phase of the optical flow filed are employed as motion features. On the spatial information of a single frame, i.e., the movement of the body at a specific time, a novel body matching approach is proposed based on a fast gradient response maps based algorithm. The proposed approach exploits the key movement of local regions rather than the whole body gesture, thus reduce the complexity of the matching process. At the same time, in the process of template matching, hierarchical templates of different gestures and angels are built, so that the detection speed can be increased while the detection accuracy can be enhanced. At last, based on the obtained temporal and spatial information, the HMM is employed to estimate the state sequence of the body gesture, which is the detection result.In this thesis, the technique framework of the proposed scheme is introduced, and a prototype system of HMM based photographing gesture detection is realized. Experimental results demonstrate that the proposed approach is very robust and produce real-time results.