关键词: 多人姿态估计, YOLO-Pose, 遮挡场景, 轻量化, NAMAttention

Abstract: Human pose estimation plays a crucial role in various real-world applications， such as sports training，robot behavior training，and intelligent interaction. Due to the shortcomings of complex neural network structures and the low efficiency of most human pose estimation algorithms，a multi-person pose estimation algorithm，YOLO- Pose-GSNS，based on improved YOLO-Pose，was proposed. To reduce the parameters and computational complexity of the module and achieve lightweight by improving computational efficiency，the GSConv convolution module was used instead of the ordinary Conv convolution calculation. Using the NAMAttention module to redesign its feature fusion layer and improve its feature extraction capability，while using four different detection heads to enhance the algorithm′s detection of occluded scenes. Introducing the SIoU loss function to redefine the loss function of bounding box regression and improve the accuracy of localization. Tested on the OC_Human dataset，the improved YOLO-Pose-GSNS model showed a 7. 4% reduction in model size compared to the baseline model，a 3. 4% decrease （19. 5）in GFLOPs，the P-value，R-value，mAP@0.5，and mAP@0.5：0.95 increased by 8. 7%，13. 4%，12.1%，and 17.2%，respectively. The YOLO-Pose-GSNS algorithm proposed in this article not only achieves the model′s lightweight，but also ensures an improvement in the accuracy of multi-person pose estimation in occluded scenes.

Key words: multi-person pose estimation, YOLO-Pose, occluded scene, lightweight, NAMAttention