Research On A Joint-Based Algorithm Of Motion Recognition And Posture Analysis System

In traditional sports training, training methods are usually based on visual observation, and with the development of computer vision, people began to use the camera to capture and analyze the action of athletes. However, tracking and analyzing process usually requires manual involvement, which leads system to a lack of automation. Therefore, in order to analyze more automatically and precisely, we do research on the motion recognition and posture analysis technology based on human joints. This monocular vision-based method can overcome the traditional manual method and improve intelligence and accuracy of the system, so as to provide instructors with more comprehensive information.Firstly, we use background subtraction and frame differencing method to detect foreground and adopt morphological processing to noise filtering. Secondly, eight-connected-neighborhood method is used to perform image thinning, and multi-feature fusion method is used to perform feature initializing. Next, we adopt Lucas-Kanade optical flow algorithm and Kalman filter to track and predict trajectory of joints. Finally, we apply K-nearest neighbor classifier to evaluate the similarity of human actions, and reconstruct a3D model by a proportional and orthogonal projection model to obtain angles of3D joints, so as to provide some training suggestions. The specific contents of this study are as follows:1. After studying the common detection methods, we propose a threshold-based method to extract moving objects. When the background image is easily captured, the combination of background subtraction and frame differencing can segment a complete foreground, which makes the silhouette image with high integrity and low complexity. On the other hand, When the background image cannot be obtained, we propose a background updating method using flood fill segmentation to construct background and foreground by calculating differential frames, which has a good background modeling results.2. Traditional joints extraction methods are optical marking or hand-labeling, which has large error and low degree of automation. Therefore, this paper presents a multi-feature fusion method, which includes vertical integration projection, horizontal scanning, index lookup tables and length ratio constraint, to initialize sixteen human joints automatically under the circumstance when the person maintains a predetermined posture. This method significantly improves intelligence and accuracy of the system. Next, we adopt Lucas-Kanade optical flow algorithm and Kalman filter to track and predict trajectory of joints, and solved the problem with wrong tracking and tracking lost, due to occlusion or excessive movement.3. In action recognition and analysis process, we do quantitative analysis of motion. We apply K-nearest neighbor classifier to match a testing action with a standard action and calculate the similarity of the two. And we reconstruct a3D skeleton by a proportional and orthogonal projection model, based on the priori information of body length and continuity constraints, which solved the3D information lacking problem caused by monocular vision without camera calibration, and provide training suggestions by visual analysis of joints information.The human motion recognition and posture analysis system we construct is verified in golf swing action. The system can capture human joints accurately from a single camera, recognize golf action and get3D movement information, in order to provide effective assistance to sports training.