CPESS: Lightweight Hemiplegia Recognition Based on Video Pose Estimation and Sparse Spatiotemporal Graph Convolutional Networks
摘要
Hemiplegia recognition is crucial for early diagnosis and rehabilitation. Existing video-based methods suffer from high computational cost and parameter redundancy, hindering mobile deployment. We propose CPESS (Connect Pose Estimation Sparse-ST-GCN), a novel lightweight framework. CPESS uses AlphaPose with an SE-ResNet backbone for robust skeleton extraction, followed by a novel Sparse Spatio-Temporal Graph Convolutional Network (Sparse ST-GCN) for efficient gait analysis. This core network incorporates dynamic graph learning and explicit structural sparsification to minimize FLOPs and parameter count while maintaining accuracy. Experiments show CPESS significantly reduces redundancy compared to traditional ST-GCN, achieving 80.0% accuracy with only 0.75 G FLOPs and 3.74 M parameters. We integrate this model into an intelligent rehabilitation system based on augmented reality, demonstrating its potential for real-time clinical application.