Robust 3D Pose estimation and Parkinson's Disease classification via Dual-Stage Adaptive Temporal Perception and graph topology modeling network.

Abstract

This paper proposes a unified skeleton-based framework for 3D human pose estimation and Parkinson's disease classification, integrating a Dual-Stage Adaptive Temporal Perception (DATP) strategy and an Adaptive Graph Topology Modeling Network (AGTM-Net). DATP enhances robustness to joint occlusion and sequence degradation through occlusion-aware interpolation, trend-extrapolated frame padding, and multi-scale spatiotemporal modeling. On the MPI-INF-3DHP dataset with 16 missing joints, DATP achieves 77.72 PCK and 43.57 AUC, outperforming state-of-the-art methods. On Human3.6M, DATP also shows strong generalization with MPJPE reduced to 32.68 mm. For clinical classification, AGTM-Net dynamically models skeletal structure variations and achieves an F1-score of 0.898 and accuracy of 0.881 in distinguishing healthy individuals from Parkinson's patients with a score of 0 based on the "3.9 Arising from Chair" task. Interpretability analyses-based on gradient and perturbation methods-highlight the spine, chest, and hips as decisive joints, aligning with clinical understanding of gait disorders and enhancing the model's transparency and clinical reliability.