<scp>KANs</scp> ‐ <scp>GCN</scp> Transformer With Hybrid Adjacency Graphs for Accurate <scp>3D</scp> Human Pose Estimation

ABSTRACT Human pose estimation is widely used in various fields such as sports and fitness, gesture control, unmanned supermarkets, entertainment, and gaming. However, current research on human pose estimation faces challenges such as low accuracy and poor model generalization. This paper proposes a novel human pose estimation framework based on graph neural networks. First, a dynamic adjacency matrix is combined with a static adjacency matrix to adaptively model dynamic, feature‐based connectivity relationships. By introducing learnable weighting parameters, the model can flexibly balance static geometries with feature‐driven dynamic relationships, capturing complex skeletal features more accurately. To further enhance the expressive power of graph convolution, this paper introduces the Chebyshev polynomial convolution mechanism to capture higher‐order neighborhood information. Inspired by Kolmogorov Arnold Networks (KANs), a learnable linear transform layer (KANLinear) is integrated into the Chebyshev graph convolution to learn graph structure feature representations across different layers. This design allows the model to effectively capture both local and global joint point dependencies, demonstrating superior performance in complex skeleton modeling tasks. Experimental results show that the proposed method achieves strong performance on the benchmark dataset Human3.6M, with a P1 error of 41.5 mm and a P2 error of 32.6 mm over 81 frames.