Application of Reinforcement Learning Algorithms in Intelligent Sports Training Video Tracking

With the advent of the intelligent era, computer-aided motion video analysis has become an important branch of computer science application field. In line with this research trend, this study proposes an intelligent sports training video tracking algorithm based on reinforcement learning, aiming to improve the tracking accuracy of athlete movements in sports videos. First, in order to achieve efficient video tracking from a theoretical perspective, this study first introduced a skeleton key point detection model to maximize the extraction of the main features of human actions in the video. At the same time, considering the instantaneous and explosive nature of sports videos, a motion recognition model with adaptive learning capabilities was constructed by introducing small sample modeling and posture estimation technology. Second, the high-resolution network (HRNet) structure is combined to perform parallel processing of multi-resolution feature maps, so that the high-resolution branch can maintain feature consistency throughout the network depth, thereby accurately identifying targets in complex motion scenes. Third, to further optimize the tracking performance, the reinforcement learning part of this study adopted an algorithm based on the Markov decision process, using a linear weighted template update strategy to adapt to the morphological changes of the target in dynamic scenes. In the experimental part, we first collected relevant data and tested it on a video database containing basketball, baseball, swimming and football, and the tracking accuracy reached 99.23%. The comparative experimental results show that the accuracy of this study is higher.