Group Activity Recognition Technique And System Combining Local Motion Pattern And Individual Relationships

With the development of artificial intelligence technology,more and more scholars pay attention to the intelligent processing of video information.The state has vigorously issued several policies to develop the sports industry in recent years.Intelligent analysis and processing of sports videos have broad application prospects in assisted training,professional commentary,technical and tactical analysis.Many studies focus on the intelligent behavior recognition task of group sports video.Video-based group activity is completed by multi-person collaboration and includes many individual local motions;simultaneously,there are rich interaction relationships between individuals.There is an excellent correlation between this information and group behavior,which significantly contributes to the realization of the group activity recognition task.The motion in group videos of sports events is a hybrid motion formed by coupling multiple motions such as the global camera motion,individual local motion,and scene-independent motion;meanwhile,the interaction between individuals is random and complex.Therefore,it is very challenging for group activity recognition to analyze the motion pattern of mixed motion to obtain accurate individual local motion and accurately model the interaction between individuals.This thesis studies the intelligent system for group activity recognition of basketball game videos.The main work includes the following.(1)A motion estimation algorithm based on the encoder-decoder structure is proposed.The local motion information of players in basketball game videos contributes to analyzing and identifying group activity.Therefore,it is necessary to use motion pattern analysis to decouple mixed sports accurately and further use player local motion to perform group activity identification tasks.However,existing methods cannot accurately estimate player local motion and camera global motion in mixed motion.Aiming at this problem,this thesis innovatively uses different encoding structures to model low-dimensional camera motion features and high-dimensional player motion features from the mixed motion.It restores the features to the complete camera global motion and local motion through the decoding structure.Since the basketball game video dataset does not have actual camera motion and player motion as supervision signals,this thesis proposes a mask-based network training method to achieve practical network training.Further,this thesis verifies the accuracy of the camera global motion estimation in the homography estimation task,uses the player’s local motion to improve the accuracy of the group activity recognition task and verifies the accuracy of the player’s local motion extraction.(2)A group behavior recognition algorithm based on convolution of individual relation graphs in spatio-temporal domain is proposed.There are many inter-individual interactions in basketball game videos,which contribute to group activity recognition.Therefore,constructing rich individual-level features and further modeling the interaction between individuals in the temporal and spatial domains are vital steps to realize the task of group activity recognition.This thesis first utilizes player local motion and RGB images to model rich individual-level features.Secondly,the inter-individual interaction in the temporal and spatial domains is modeled based on the graph convolutional network.After graph convolution,the player interaction features are transferred in the temporal dimension in a concatenated manner.Finally,the individual-level features are fused into group-level features to recognize group activity.The experimental results on the basketball game dataset show that the proposed group activity recognition algorithm can achieve high performance.(3)A group activity recognition system for basketball game videos is developed.Firstly,the requirements of the system are analyzed.Secondly,the implementation framework of each functional module is designed based on the two algorithms mentioned above.Finally,a complete group activity recognition system is realized.The visual display,functional test,and performance test of the system verify the usability and accuracy of the system.