Algorithm And Implementation Of Real-Time Human Action Evaluation In Tennis Training Robots

The tennis training robot can identify the tennis ball by the camera and calculate the trajectory,and the robot predicts the landing position of the ball and hits it back.The current tennis training robot lacks the ability to guide and regulate the user's actions.With human action analysis technologies,the robot can evaluate the user's actions and feedback the results,which is of high practical value for the unmanned sports teaching.However,currently in the field of computer vision,the analysis of human actions mostly relies on depth cameras or binocular cameras to get three dimensional spatial information.There are many shortcomings,such as high cost,large environmental interference in data acquisition,and poor real-time performance of three-dimensional information reconstruction.In this thesis,we study the real-time human action evaluation algorithm of tennis training robot with the monocular camera.First of all,in order to solve the problem of human action direction,we propose the Counter Clockwise Rotation Angle of Limbs(CCRAL)as the action evaluation feature.Secondly,in order to evaluate the single frame action,a weighted CCRAL similarity measure is proposed in this thesis.Then,in order to evaluate the action of the whole video and solve the speed inconsistency problem between the template video and the input video,we improve the traditional Dynamic Time Warping(DTW)algorithm to reduce the redundant computation and enhance the real-time by limiting the range of the dynamic search path in the DTW algorithm.Finally,we design and implement a prototype system of tennis action evaluation based on a monocular camera.We conduct the experiments in the real scene.The results show that the system achieves good effects in efficiency and accuracy.