KANs-Hybrid Adjacency Graph Transformer for 3D Human Pose Estimation

Research on human posture estimation faces challenges such as low accuracy and poor model generalization capabilities. To resolve these concerns, this paper proposes a novel graph neural network (GNN) framework for human posture estimation. It adaptively models dynamic feature-based connectivity by fusing a dynamic adjacency matrix with a static one, using learnable weighting parameters to balance static geometry and dynamic relationships. The framework further enhances graph convolution by incorporating a Chebyshev polynomial convolution mechanism to capture higher-order neighborhood information. Inspired by Kolmogorov Arnold Networks (KANs), a learnable linear transform layer KAN is integrated to learn cross-layer graph structure representations, effectively capturing both local and global joint dependencies. Experimental results demonstrate that the proposed method achieves strong performance on the benchmark datasets, attaining 41.5 mm in MPJPE(T=81) and 32.6 mm in P-MPJPE on Human3.6M, as well as 98.6% PCK, 83.6% AUC, and 23.4 mm P1 error on MPI-INF-3DHP.