Study On Model-based Human Motion Tracking And Gesture Analysis Technology

Tracking human body in video and estimating pose is a hot field of computer vision, and has broad application prospects. However, there are some shortcomings in existing methods, such as inaccurate and low pose estimation, therefore it is necessary to research a good method of tracking and pose estimation.There are so many varieties of person features, it is the key issue for human motion analysis to choose and extract features. This paper introduces histogram of oriented gradients feature in human detection, because human detection is the premise of tracking and poses estimation, but it is difficult for the original methods to be real-time, the paper introduces the integral histogram to the calculation of HOG features which reduce the computation time.The tracking method is selected on the basis of different application environments and the characteristics of tracking targets. This paper requires real-time, fast implementation and the observation of prospect is more accurate, so we adopt the tracking method based on Kalman filter. The innovation is combining the HOG with Kalman filter, which not only can track the position of targets, but also can determine the body size laying a good basis for pose estimation.This paper takes advantage of approach which is based on graph structure to estimate human pose, and improve it. We carry out some processing before pose estimation for reducing the search space. First of all, detecting the human body, the location and scale information of human can be determined, and the detecting window not the whole image is the input of pose estimation, which can reduces the search space and improve the speed of pose estimation. According to prior knowledge, we add some restrictions in the detection window, for example, the body's head and torso is usually in the middle of the detection window, the head in the middle of the trunk, the trunk just below the head, thus reducing the search space and enhancing its speed. This paper uses not only the apparent continuity among the video frames, also uses the geometric continuity in the human body pose estimation of video sequence, so as to enhance accuracy and speed.