Design And Implementation Of Basketball Players' Action Standard System Based On RGBD Pose Recognition

In basketball training,the analysis of an player can provide the coach the show about players,who can develop appropriate training plan to improve the team's competition results.With the application of video analysis,we can analyze the action of basketball players in training,we introduce the human pose estimation to provide accurate analysis for players.Currently based on the traditional RGB video image analysis of many studies,however,due to the influence of illumination,appearance,background,and the serious occlusion between the players,there is a big error in the results of video analysis.With the appearance of Microsoft Kinect in 2010 makes the depth data more and more easy to obtain.At the same time,the depth image is less affected by illumination and background,so the method of video analysis based on depth image has received attention.Action standard of basketball players include:human pose estimation,action alignment and similarity measurement.Most researchers use joint point data obtained from monocular camera for human pose estimation,and these data need to be improved in accurately recovering 3D human pose.In order to improve the accuracy of 3D human pose estimation,this paper adopts a method of 3D skeleton point extraction method based RGBD images;for action alignment,this paper uses a dynamic time warping(DTW)algorithm;for similarity measurement,this paper proposes a pearsonr correlation coefficient method based on joint feature of human body parts.The main contributions of this paper are as follows:(1)A method of basketball players' action standard based on human pose estimation and dynamic time warping is studied.(2)A method for extracting human skeleton from depth images is used.3D human pose network uses the voxel grid of the depth image and the score image of 2D human keypoints in color image as input,which predict 3D skeleton points of human body.(3)It gives a dynamic time warping method for motion alignment to solve the problem of inconsistent action time of different person.By using joint points of body to represent human body pose feature,the human body similarity measurement experiment is carried out on the basis of pearsonr correlation coefficient.(4)A system of correct the action of basketball player is designed,which collects color video and depth video of the players' training actions through the data acquisition module,then upload the data to action standardize module,and it helps players analysis the difference between training actions and standard actions.