Research On Long-time Tracking System Of Athletes Based On Machine Vision

A high-quality competition is often inseparable from the tactical arrangement of the coach.Whether the real-time information(such as the speed and exact route)of the athletes can be obtained or not,plays a decisive role in the arrangement of personnel and tactics by the coach.The core function of the athletes' tracking system is to obtain the real-time information of athletes.In outdoor sports competitions,wearable GPS devices are often used to achieve real-time tracking of athletes,while in indoor sports competitions,highprecision multi-sensor systems are generally used to obtain athletes' location information.Both of these approaches make the tracking system overly complex and difficult to maintain.Based on the reality,this paper proposes a tracking system which can be used in mobile devices.The core of the system is a tracking algorithm which can meet the requirement of accuracy and real-time.The research object of this paper is the single target tracking algorithm based on machine vision with deep learning.Through the research of the deep learning tracking algorithm,the improvement of the SiamRPN tracking algorithm is proposed to improve the tracking accuracy and convert the short-term tracking into the long-term tracking.The main work is as follows:1.Based on the research of Siamese network structure,introduced the channel attention mechanism in the template branch.It can select the feature of template image feature which is generated by the feature extraction network,and give the corresponding weights of different channel features so that it can highlight the feature of tracking.It also can let the tracking algorithm to adapt to the variations of target in tracking process.2.In order to make the short-term tracking become long-term tracking,and to make the tracking algorithm rediscover the target when the target disappears and reappears,a simple and efficient local to global search strategy without parameter is introduced.According to the maximum value of the output of the sub-Region Proposal Network classification branch,the current search area can be judged to cover the target.And then the size of the corresponding search area can be changed,so that the transformed search area can cover the target.3.An efficient sampling strategy is proposed: 1)increasing the number of positive sample pairs of different categories can improve the recognition ability and regression accuracy of the tracking algorithm;2)increasing the number of negative sample pairs of different categories can help the tracking model avoid drifting to other arbitrary objects when challenges(such as out-of-view and full occlusion)occur;3)increasing the number of negative sample pairs of the same category can make the tracking model focus on the representation of the current tracking object.Where the positive sample pair is two images containing the same tracking object and the negative sample pair is two images containing different tracking objects.4.The improved SiamRPN single target tracking algorithm is applied to the athletes' long-term tracking system.Because GPU supporting deep learning operation is not generally available in the mobile terminal devices,such as mobile phones,Thrift framework is used to connect the mobile terminal devices with GPU-equipped computers to build the athletes' long-term tracking system.