Lightweight Human Pose Estimation Model for Industrial Scenarios

The application of human pose estimation models in industrial environments is of paramount importance, and considering that industrial environments are often constrained by computational resources and real-time requirements, the development of accurate and efficient models is crucial. Although lightweight posture estimation models have been explored by researchers, we believe there is still room for further optimization. For this reason, this thesis is dedicated to the development of a lightweight pose estimation model based on the YOLOv8 architecture. By replacing the bottleneck module of C2f with the RepViTEMA Block module, the application of the attention mechanism and the deep convolution reduces the number of parameters and the computational complexity of the model. In addition, we take full advantage of the idea of shared convolution to drastically reduce the number of branches in the head structure and make it more concise. Finally, our improved approach, YOLOv8 LWSP, reduces the number of parameters by about 47% and the computational complexity by 40%, with almost no loss in detection accuracy.