Basketball Action Recognition Based on Inv-HRNet-BiLSTM

Abstract

With the development of computer vision technology, the application of action recognition in sports is becoming increasingly widespread, and basketball has become a research focus due to its fast pace, drastic limb changes, and frequent occlusion. To address the issues of insufficient spatial modeling capabilities and limited temporal expression effects in existing models, this paper proposes a novel spatiotemporal joint modeling basketball action recognition model called Inv-HRNet-BiLSTM. This model adopts the core methodology of "dynamic spatial feature extraction, temporal dependency modeling, feature enhancement optimization", innovatively integrating the Involution spatial adaptation mechanism with High-Resolution Networks (HRNet) to construct the Inv-HRNet human pose estimation network. By dynamically generating exclusive convolution kernels for each spatial position, the modeling capabilities of joint displacement and motion deformation are enhanced. At the same time, the multi-scale parallel structure of HRNet is used to maintain high-resolution feature transmission and improve detail capture accuracy. On this basis, the model integrates a Bidirectional Long Short-Term Memory Network (BiLSTM) and introduces a dual encoding mechanism, combing position encoding and motion amplitude, to achieve bidirectional deep modeling of action temporal features. This enables the accurate capture of continuity and dynamic changes in action stages. The experimental results show that the accuracy of the Inv-HRNet pose recognition network reaches 94.94%, with a recall rate of 92.23%. The average recognition accuracy of the Inv-HRNet-BiLSTM model in complex scenes, such as occlusion and blur, is 78.58%, and the highest recognition accuracy for eight typical basketball movements, such as dribbling, shooting, and passing, is 95.28%. At the same time, the inference speed is 20.56 ms/frame, and the memory usage is only 723.90 MB, which is significantly better than existing comparative models. The model achieves efficient integration of recognition accuracy, robustness, and real-time performance through organic integration and collaborative optimization of multiple modules, providing a lightweight and high-performance system architecture reference for the field of sports intelligent recognition