Object Detection And Human Pose Estimation For Sports Videos

With the development of society,more and more sports videos are available in people's daily life.Analysis of these sports videos can not only bring more exciting content to the audience,but also find out and improve the shortcomings of players.Player detection and pose estimation are key points in sports videos analysis.Existing human detection and pose estimation algorithms have achieved good performance on general human detection tasks.But in sports videos,these methods detect player and audience at the same time,and can not further distinguish between them,which will interfere video analysis.At the same time,sports video data with box annotation is scarce,and it is costly to obtain a model suitable for the field of sports video.Players are a special case of human.If we can use the general human detection model to detect players in sports videos,a lot of costs can be saved.In order to reduce the cost of labeling and training,and to realize the detection and pose estimation of players in sports video,this thesis studies how to transfer the general human detection model to the field of sports video.The main research works are as follows:1.A human detection algorithm based on instance feature metric is proposed.The algorithm first extracts the features based on the human detection results of the general detection algorithm,then initializes a feature vector to represent the player.Subsequently,the algorithm learns the linear transformation module to map the human features to the player feature space,and finally recognizes the player by measuring the similarity between the human features and the player features.Based on the general human detection algorithm,the algorithm can learn a feature map of a general target to player target by multi instance learning without any box annotation,and the detection performance is comparable with dedicated player detection network.2.A player pose estimation algorithm based on local space constraints is proposed.The algorithm is based on the general human pose estimation algorithm.First,the image is locally constrained according to the player detection box,only the content related to the player in the image is retained,and then the processed image is input into the general human pose estimation model for pose detection,and finally the pose is mapped back to the original image to complete the player's pose estimation.Based on the general human pose estimation algorithm and through image preprocessing with player detection box,the algorithm can obtain the player's pose directly without optional pose discrimination module.This algorithm effectively reduces the computational cost without affecting the accuracy of human pose estimation.3.The system of detection and pose estimation for player in sports video is realized.Based on Python programming,the system can be conveniently deployed on Windows,Linux and other system platforms.This system has achieved detection and pose estimation for player in sports video end-to-end.