Multi-person pose estimation based on an improved lightweight YOLO-Pose model
YOLO-Pose as a human pose estimation algorithm model had a good performance in terms of accuracy and speed,which suffered from a large false detection rate in complex and occluded scenes on the other hand.There was still room for optimization of the model complexity.In this paper,these issues were addressed by incorporating the Slim-neck module and Res2Net module to redesign the feature fusion layer,reducing computational and parameter overhead while enhancing the information extraction capability of feature extraction.Furthermore,the EIoU loss function was introduced to accelerate the convergence speed of bounding box detection and to improve localization accuracy.Experimental results on the compressed OC_Human dataset demonstrated that the improved model achieved a 10.6%improvement in P-value,a 3.1%increase in mAP@0.5,and a 2.9%increase in mAP@.5:95 compared to the original YOLO-Pose model,respectively.Moreover,the amount of parameters(Params)and computational complexity(GFLOPs)were reduced by 16.7%and 19.3%,respectively.The improved model showed enhanced accuracy and lightweight characteristics,which was suitable for deployment on resource-constrained edge computing devices.
human pose estimationYOLO-PoselightweightingSlim-neck
NETL
NSTL
万方数据
