Design And Implementation Of Intelligent Game Analysis System For Basketball

With the promotion of NBA and CBA basketball leagues,the rise of online short videos,more and more basketball videos are available on the Internet.At the same time,people need to edit footage manually in real time for broadcast of basketball games,and also need to edit highlights motion manually for short web videos.This means a large demand for the manpower from these tedious tasks.In this thesis,we locate and classify the action in each time period of basketball sports videos as a whole,and separately classify the sub-actions occurring at each moment.This method is used to analyse the game and assist people to edit the basketball sport clips of interest intelligently.The existed solutions for locating the video clips of interest are mainly applied to simple scenarios,which directly use single-frame classification or some pose estimation methods.Because they didn't take into account the continuity of action,they couldn't be directly applied to the basketball sport,which action with complex motions,rapid change and fuzzy boundary.This thesis presents an intelligent analysis system,which is mainly based on image classification,video classification and action detection technologies.Specifically,it mainly includes:(1)A highly efficient neural network is designed to extract features.Spatial and temporal attention mechanisms are added to the neural network,which drive the network to learn more distinguishing features for complex actions.(2)The method of action detection is used as the main method to locate the action segments of interest,and the clustering method is used to replace the artificial setting of the anchor point of the candidate segments.Gaussian distribution and attention mechanism are introduced to learn the distributions of different basketball actions,and aimed to improve the ability of neural network to distinguish similar actions.(3)A single-frame classification method is used to assist the action detection result,which corrects the action localization of interest.Weighted losses are added to classification network to resolve the problem of fuzzy motion boundary,which makes it more suitable for action detection tasks.In this thesis,We build a basketball action dataset suitable for action detection tasks,design an algorithm for positioning basketball,implement a prototype software system that can carry the above algorithm and support user interaction.The experiments show that our system has good real-time performance,robustness and accuracy.