Human Skeleton-based Action Recognition Algorithm Based on Improved PoseC3D

In response to the limitations of existing skeleton-based action recognition algorithms that rely on 3D convolutional networks for modeling long-range dependencies, this paper proposes an enhanced human skeleton action recognition algorithm based on an improved version of PoseC3D. First, the human skeleton sequence is converted into a 3D heatmap stack as an input to the network; second, a plug-and-play 3D attention mechanism module is introduced to enhance the ability of the network to perceive critical information; and finally, a bi-directional long and short-term memory network is combined to deal with long-term dependencies, the ability of the network to capture temporal features is enhanced, especially the change of complex actions over time in the successive skeleton sequences. The experimental results demonstrate that the proposed method achieves an accuracy of 97.3% (X-View) and 90.4% (X-Set) on the NTU-RGB+D60 and NTU-RGB+D120 datasets, respectively. These results represent an improvement of 1.8% and 0.3<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">%</sup> over the baseline algorithm, PoseC3D, thereby validating the effectiveness of the proposed approach.