Research On Motion Pose Analysis System Based On Machine Vision

In the past ten years,people's life has been getting better and more and more attention has been paid to sports competitions.In the current information age,the data of sports competition and athlete are very important,especially for basketball,football,volleyball and other team-oriented big ball games.Through these data,the coach can well analyze the situation of his own player and the situation of the other team,so as to better specify the corresponding tactics to get the final success.However,at present,most of the data statistic need to be recorded and counted manually,either on site or after the game.All kinds of sports are composed of various movements,and movements are composed of a continuous sequence of body posture.Therefore,based on machine vision combined with pose estimation and action recognition methods in deep learning,this paper attempts to build a system for basketball to record and count the above-mentioned tedious but very important data instead of people.In this system,the key nodes of each basketball player in the video sequence are extracted and constructed into a fixed-length pose sequence by using the pose estimation method.After the pose sequence information with fixed length is obtained,it is fed into the motion recognition method based on graph convolution to obtain the motion classification.Finally,the system is used to record the actions of each player in each frame and make statistics according to the meaning in basketball.The main work of this paper is as follows:1.This research proposes motion posture analysis system for basketball movement,which is composed of two parts.The first part is the analysis and research of bottom-up pose estimation algorithm,which is used to extract the posture sequence of video.The second part is the analysis and research of the motion recognition algorithm based on the convolution of spatial temporal graph,which is use for recognizing the defined basketball actions in the line with the extracted posture sequence.2.The modified posture estimation method is combined with the motion recognition network trained with new data.Pyqt5 and OpenCV in python were used to establish a front-end human-computer interface for the statistics and display of network output results,and finally a statistical analysis system of basketball action was formed,and its validity was verified by basketball video.3.According to requirements of the system,OpenCV was used to make the acquisition and calibration tool for the basketball video movement and corresponding basketball image,and the related basketball game action classification video data set and body posture data set were established.The corresponding training set and test set were obtained by reasonable segmentation,which were used for the system construction and testing.4.In the first part of the experiment,it was found that using the classic pose estimation model Openpose for the standard basketball game pose estimation would affect the pose recognition effect due to motion blur.According to the principle of motion ambiguity,this paper innovatively proposes a new data enhancement method to increase the data of model training.By combining with transfer learning,the new model is obtained through training to improve the recognition errors caused by motion ambiguity in the data.