Lightweight PoseC3D network based on depthwise separable convolutions and CBAM

To address the challenges of high computational complexity and insufficient real-time performance of action recognition algorithms in the field of motion-assisted analysis, we propose an action recognition method that integrates lightweight network design with an attention mechanism. First, to mitigate the heavy computational burden of traditional 3D convolutions, we construct a lightweight PoseC3D network by replacing standard 3D convolutions with depthwise separable convolutions, thereby substantially reducing model complexity. Second, to compensate for the accuracy degradation introduced by depthwise separable convolutions, we incorporate the Convolutional Block Attention Module (CBAM) to perform feature re-calibration, enabling the network to focus on the most informative spatial and channel features. A large number of experiments were conducted based on the FineGym dataset and the NTU60 dataset demonstrate that the proposed lightweight PoseC3D model achieves nearly a 50% reduction in GFLOPs while maintaining a Top-1 accuracy of 88.9%. Compared to the original PoseC3D network, our method significantly lowers computational cost with only a marginal decrease in recognition accuracy, effectively balancing performance and efficiency.