Research on Personalized Soldier Training Pose Evaluation Algorithm Based on Deep Learning

This study presents Lite-TransPose, a novel lightweight deep learning-based human pose estimation framework designed to address challenges of high computational complexity and limited real-time performance in traditional models. Leveraging depthwise separable convolution, GhostNet modules, and multi-head attention mechanisms, Lite-TransPose significantly reduces resource requirements while maintaining accuracy. Furthermore, it introduces optical flow-based pose tracking to enhance performance in video stream data. Experimental evaluations on COCO, MPII, and custom soldier motion datasets demonstrate the model's ability to achieve competitive accuracy with reduced parameters and computational costs, making it suitable for deployment in resource-constrained and real-time applications. This work offers a practical solution for human behavior analysis and embedded pose detection in domains such as military training and physical fitness evaluation.