3D Human Pose Estimation Method Based on Improved High Resolution Network
In order to solve the problems of occlusion and human key point movement under complex background in real scenes,a top-down two-stage human pose estimation method is proposed.Firstly,the improved YOLO and SORT are used for two-dimensional human de-tection and tracking,and the YOLOv3 network structure,loss function and prior frame size are improved to enhance the network detection ability and feature expression ability,and improve the applicability and accuracy of human target detection.Secondly,HRNet integrated with attention mechanism is used for 2D attitude estimation,and the original residual module is featured to enhance the cross-channel in-formation exchange between multi-layer features at different scales and improve the recognition effect of blocked key points.Finally,GAST-NET is used to generate 3D posture.The experimental results show that mean per joint position error and procrustes analysis mean per joint position error are 45.0 mm and 35.4 mm respectively.Under the condition of serious occlusion,the accurate position of human key points can still be obtained.
国家科技期刊平台
NSTL
万方数据
