Abstract
This paper presents a novel framework for accurate exercise posture recognition and health indicator prediction based on improved convolutional neural networks. We propose a multi-scale feature fusion architecture incorporating spatiotemporal attention mechanisms to enhance key point detection precision while maintaining computational efficiency. The system achieves superior posture recognition performance with 78.6% mAP and 91.5% PCK@0.5, outperforming state-of-the-art methods while maintaining real-time inference capabilities (27.3 FPS). For health indicator prediction, we develop a CNN-LSTM model with personalized parameter adaptation that accurately forecasts multiple physiological metrics including cardiorespiratory fitness, muscular strength, and metabolic rate, achieving 86.1-92.6% prediction accuracy across diverse health dimensions. Comprehensive evaluations on both self-collected and public datasets demonstrate the system's robustness across varying exercise types, environmental conditions, and demographic groups. The proposed approach offers significant potential for applications in personal fitness coaching, rehabilitation monitoring, and preventive healthcare by providing automated exercise form evaluation and personalized health insights.