3D-MLKDNet: Synergistic pose estimation via 3D mesh reconstruction and multi-level keypoint detection for physically disabled individuals

Human activity recognition (HAR) and motion analysis play an important role in rehabilitation, sports science, healthcare monitoring, and human–computer interaction. However, conventional RGB-based vision systems often suffer from limitations such as occlusions, illumination variations, background complexity, and privacy concerns, reducing their reliability in practical healthcare and rehabilitation environments. To address these challenges, this study proposes a privacy-preserving and low-cost multimodal depth-guided exercise classification framework that integrates semantic segmentation, adaptive keypoint localization, lightweight motion representation, and 3D mesh reconstruction for robust rehabilitation-oriented motion analysis. The proposed framework introduces a Multi-Level Keypoint Detection (MLKD-24) method capable of adaptively extracting 24 anatomically consistent 3D joints based on varying upper-limb configurations, enabling precise SMPL-based 3D human mesh reconstruction from depth data. By operating entirely on depth information, the proposed system preserves structural and geometric body characteristics while avoiding appearance-dependent identity cues such as facial texture and color information, thereby improving privacy preservation in healthcare and telerehabilitation applications. Furthermore, ORB and BRISK descriptors are incorporated as lightweight and computationally efficient rotation-invariant motion representations to improve robustness against posture variations and challenging rehabilitation movements without relying on computationally expensive deep learning architectures. Finally, silhouette-based, keypoint-based, descriptor-based, and 3D mesh representations are fused through a unified multimodal framework and classified using a machine learning model for robust rehabilitation exercise recognition. Experimental evaluation on the KIMORE, K3DA, and MEx datasets achieves classification accuracies of 93.9 %, 92.58 %, and 87.51 %, respectively, demonstrating the effectiveness of the proposed framework for fine-grained rehabilitation motion analysis under privacy-sensitive and computationally constrained healthcare environments.