Research On Some Human Behavior Recognition On Video

Recognition of human action is one of the important topics in computer vision. It has attracted many researcher's attention in recent years, and can be widely used in intelligent video surveillance, intelligent robots, motion analysis and other areas. However, due to non-rigid motion of human motion and differences in appearance, physique, exercise habits, the human behavior recognition becomes more difficult, so far, there doesn't exist a universal behavior recognition model, most of the studies carries under a particular scene.Our purpose is to recognize such behavior like walk, run, jump and bend, and study some algorithms that related. Since moving target detection is the first step in human behavior recognition, its results have a direct impact on behavior recognition results, so we first studied human movement detection algorithm. We have used temporal difference method, statistical average algorithm and gaussian mixture model to detect moving target, analyzed these algorithms and finally chosed gaussian mixture model. For the detection results, we then removed the noise and juded whether the moving target is human based on the features of human body.For the moving body we detected, we carried out feature extraction, and studied some feature extraction algorithms. We used two methods for feature extraction, one is based on Hu moments, and the other is based on Fourier descriptors. Based on this two methods, we then established two human model: one model based on Hu moments and the other based on improved Fourier descriptors.The final step is human behavior recognition, we first defined some key frames for the four behaviors we intended to recognize, then used equal interval sampling method to get the key frames. For the key frames we established two human body libraries through feature extraction. At last we implemented the recognition of human behavior, compared the results and analysed them.