Improvements to the human pose estimation network based on YOLOpose

Human pose estimation is of critical significance in areas such as workplace safety and intelligent interaction. This paper presents a lightweight improved algorithm based on the YOLOpose model to tackle the challenges of high computational demands and slow detection speeds in existing human pose estimation models. Firstly, the more efficient GSConv convolution module is utilised, which significantly decreases the model's computational burden and complexity. The second of these is the integration of the Res2Net module, which enhances the network's feature extraction capabilities and serves to expand the receptive field of its layers, thereby improving accuracy. Furthermore, the loss function is modified to use EloU, which accelerates the convergence speed and enhances the precision of bounding box regression. The experiment yielded results that demonstrate the efficacy of the proposed lightweight enhancements to the YOLOpose model. These enhancements result in a significant reduction of model parameters (86.8%) and computational load (71.2%) while preserving a satisfactory level of accuracy. This effectively mitigates the model's computational complexity, ensuring that the enhanced model maintains recognition accuracy while implementing a lightweight detection algorithm.