Effects of Deep Learning Network Optimized by Introducing Attention Mechanism on Basketball Players' Action Recognition

Abstract

Video, as an important carrier of big data storage, can help people to achieve behavioral analysis as well as localization. The study applies it to the training of basketball players as well as the judgment of the game, and regulates their sports posture through action recognition. The study first builds and optimizes a basic network based on the two-stream fusion algorithm. Subsequently, the time domain neural network and spatial neural network are improved using the multi-module attention mechanism as well as the residual network, respectively, in order to realize the extraction of deep spatio-temporal features. The study introduced a novel multi-head self-attentive temporal capture module to capture multiple time points. Additionally, the experiment used a spatial attention mechanism to capture spatial target association information. The study aims to increase the model's performance in action recognition and strengthen its understanding of important sequences. A moderate Sigmoid activation function is used by both the time-domain neural network and the spatial neural network to enhance the linkage between the images in each frame. The two-stream method is then used to merge the images. According to the trial results, the research design model's basketball action recognition accuracy was 95.4% and its recognition speed was 20 frames per second. In comparison with the rest of the models, the design model achieved a 25% improvement in recognition speed and a 47.27% reduction in average recognition time, which gives it the best overall performance. Therefore, the proposed optimized spatio-temporal neural network based on two-stream architecture is able to perform basketball sports action recognition faster and more accurately.