Design And Implementation Of Sports Video Analysis Service System Based On Container Technology

When doing action specification guidance in the sports field,the method using high-speed photography equipment to record slow shots or using motion capture equipment to collect 3D motion information,combined with professional analysis is difficult to meet the growing user groups and needs in the sports guidance market in terms of cost and efficiency.As the research achievements of deep learning in the field of computer vision gradually penetrate into many industries such as sports,medical care,and transportation,sports video analysis based on deep learning has gradually entered the mainstream of the market.In addition to using theoretical knowledge to build a neural network,a mature and available deep learning algorithm also needs to go through a series of steps such as data acquisition,data cleaning,model training,model evaluation,and model optimization.For use the model in an online environment,it is another challenging problem about how to deploy the algorithm in engineering environment efficiently and reliably.The deployment method of the model has evolved from deploying to physical machines directly,to deploying to virtual machines,to deploying to containers in physical machines,and it has developed into a deployment method that uses cloud servers combined with containers today.Taking the deployment of deep learning model as the starting point,this thesis author designed and implemented a sports video analysis service system that integrates video collection,video analysis,video annotation,and model deployment.The system uses container technology for model deployment and algorithm packaging,and implements a scheduling algorithm based on Kubernetes for GPU resource allocation during algorithm runtime.The thesis first introduces the research background and significance of the subject,then describes the system requirements and the research and comparison results of existing GPU sharing algorithms.Next,the overall design of the system is described,and then the architecture design,detailed design and implementation are introduced in the form of subsystems.Finally,the system test is explained,and the work of the thesis is summarized and prospected.The system will call several deep learning models deployed in containers for sports video analysis.Then,the analyzed video will be relabeled by the labeling subsystem to become new data for algorithm developers to optimize the model,so as to obtain a new model with better generalization and better effect.The new model can replace the old one for the video analysis business,forming a closed loop from application to data to algorithm.