An Action Evaluating System Based On 3d Human Posture

Research of human action includes action recognition,gesture recognition,human posture estimation and so on.As one of the hot topics in the field of computer vision,the research of action can be widely applied in the fields such as virtual reality,video surveillance,and sports analysis.In this thesis,the study of action is based on three-dimensional posture of human body,and developing an golf action evaluating system is main purpose of the thesis.The existing systems for motion capturing usually need additional auxiliary equipments,which is inconvenient to operate.And some devices can only be used indoors.Methods of monocular depth estimation have low accuracy and poor robustness.Therefore,this paper studies the method of binocular stereo vision and deep learning to get 3D pose of human.First,the two-dimensional joints are extracted by cascade network structures,then we get 3D information of joints based on parallax principle,in the process of golf swing,in order to acquire the posture more correctly,the Kalman filter is used to correct the information of joints,so the 3D posture of human body can be reconstructed.Generally,speeds of action for different people are not the same,therefore,Dynamic Time Warping can be used to solve the problem of unaligned actions.After alignment,we can evaluate the similarity of actions in a more reasonable way.Finally,a system for evaluating action of golf swing is developed in this paper.The system has functions containing camera calibration,data acquisition and action evaluation,and it can be performed as supplementary training system by providing the joint angles data of learner and the standard.This system can be helpful for people who wants to improve their swing level.The main contributions of the thesis are as follows:(1)Due to the shortcomings of existing motion capture methods,this paper proposes a method for capturing 3D posture of human by combining deep learning and binocular stereo vision.This method has high accuracy,what's more,it can be used real-time and widely.(2)For the evaluation of two actions with different speed,this paper proposes Dynamic Time Warping for motion alignment,by choosing good features,two videos with inconsistent length can be aligned.(3)A system for evaluating action of golf swing is developed,which can provide the action evaluation of trainer and the standard,and the difference between the postures of trainer and the standard in the key frames are given.The system can be performed as an auxiliary system for trainer.