Design And Impementation Of Sports Event Analysis System Based On Deep Learning

Deep learning is widely used in computer vision,natural language processing and other fields,which has greatly promoted the all-round progress in the field of artificial intelligence,making more and more researchers devote themselves to the work of landing deep learning into practical scenarios.Along with the booming development of sports industry,its combination with artificial intelligence has great space for exploration and practical value.In this thesis,we provide three different scenarios with practical value for the implementation of algorithms based on computer vision related technology of deep learning for current sports videos for create more possibilities in the future.Specifically,the main contributions made in this thesis are as follows.First,visible watermarks(occlusions)are a common interference problem in processing visual tasks,so this thesis first investigates the watermark removal task.Most of the watermark removal efforts of existing methods treat it as a task of image style migration,ignoring the potential information contained in the watermark itself,and thus are ineffective.As a starting point,this thesis disassembles the watermarked image into three elements:the original image,the watermark and the opacity,and proposes a WFCNet network to estimate the watermark and the opacity,which is then cut off from the original image,thus improving its process interpretability.In the experimental stage,this method is proved to be advanced by comparing with related results.Then,this thesis proposes a position correction system for soccer games,whose goal is to map the positions of players in the game to a twodimensional plane equivalent to a "vertical projection" of the stadium,to enhance the visual viewing effect while visualizing the tactical layout of the players.Based on the idea of perspective transformation,this thesis breaks it down into two sub-tasks of field key point extraction and coordinate mapping.Then,this thesis analyzes a feasible solution for soccer field feature extraction,and implements the algorithm through a Gaussian heat map-based target point detection method,and finally obtains the perspective transformation matrix through the mapping relationship with the 2D plan,and then further maps each player’s position according to the transformation matrix to achieve the expected effect.Finally,this paper also proposes a slow playback clip extraction system for the Olympics.The subject comes from the actual business requirements and is applied to the video scenes of the then-uncoming Tokyo Olympics.Slow playback clips often contain more information elements,such as highlight clips,time point characteristics,etc.,which have more important reference significance for intelligent video editing.In this paper,we first demonstrate the limitations of existing video-related algorithms in dealing with this task,and then propose a personalized solution to the problem,which transforms the task into two parts:extraction of transition frames and interval merging.In the implementation,a SECNet network is proposed for the task to ensure the generalization capability and accuracy of the transitions extraction.Then,a greedy algorithm is proposed for interval merging to further improve the interference resistance of the algorithm system.Finally,the results demonstrated by this method at the Tokyo Olympics are excellent.