DESNet: Real-time human pose estimation for sports applications combining IoT and deep learning

Abstract

With the rapid development of IoT technology, real-time human pose estimation has become increasingly important in sports training feedback systems. However, current methods often fall short in balancing high accuracy with low computational resource requirements, especially in resource-constrained environments. Deep learning has shown significant potential in enhancing computer vision tasks, including human pose estimation. In this study, we propose DESNet, an improved EfficientHRNet model that integrates IoT technology. DESNet combines Dynamic Multi-Scale Context (DMC) modules and Squeeze-and-Excitation (SE) modules, and utilizes IoT for real-time data collection, transmission, and processing. Experimental results show that DESNet achieves an average precision (AP) of 74.8% on the COCO dataset and a PCKh (Percentage of Correct Keypoints with head-normalized) of 90.9% on the MPII dataset, outperforming existing lightweight models. The integration of deep learning and IoT technology not only improves the accuracy and efficiency of human pose estimation but also significantly enhances the timeliness and robustness of feedback in sports training applications. Our findings demonstrate that DESNet is a powerful tool for real-time human pose analysis, offering promising solutions for intelligent sports training and rehabilitation systems.