| In recent years,with the increasing hardware computing power,deep neural networks have also developed rapidly as a branch of machine learning.Nowadays,deep neural networks have obtained comparable performance or even surpassed that of humans in various fields such as computer vision,natural language processing,and bioinformatics,and have become a popular direction of research in the field of computer science.Human pose estimation has become an important and active research direction in the field of computer vision due to the wide range of potential application scenarios.Using deep neural networks and the large-scale datasets collected by researchers,human pose estimation has already performed satisfactorily in 2D pose estimation.Therefore,it is an important research topic to transform the mature network into a lightweight network that can run on more devices with guaranteed high accuracy.In addition,Due to the complexity of 3D pose detection tasks themselves,there is still room for improvement in their performance.,so further improving the 3D pose detection accuracy is still the main direction of current research.In this paper,relevant research and improvements are carried out on the currently available 2D pose detection and 3D pose detection methods combined with attention mechanism respectively,and the main contributions are as follows:(1)To address the problem that the number of parameters and detection performance of 2D single-person pose estimation network HRNet,cannot be taken into account simultaneously,this paper propose Condense HRNet.Design a lightweight module based on depth-separable convolution to replace the original residual module composed of standard convolution,which greatly reduces the number of parameters and computation of the network and improves the efficiency of using convolutional kernel parameters.In addition,the channel attention mechanism is added on this basis,and the one-dimensional convolutional is used to achieve the information exchange between neighboring channels without compressing the number of channels.Furthermore,the multi-resolution fusion method of HRNet is improved and the multistage residual connection is added.Finally,a quantitative comparison with HRNet and other lightweight methods is conducted experimentally.Compared with HRNet-W32 on the COCO dataset,Condense HRNet reduces the number of parameters by 91% and the computation by 75%,26.1M and 7.1GFLOPs,respectively,the average AP metric is only reduced by 2.2% compared with HRNet-W32,which is also not pre-trained on Image Net.(2)To address the problem of restricted perceptual fields for convolutional operations in Voxel Pose,which is a 3D multi-person pose estimation network,this paper study the application of Transformer to human pose estimation and propose the VTP network,which designed a network structure by combining 3D convolution and Transformer and employing Sinkhorn attention to reduce the computational complexity caused by voxel features.It is shown through extensive experiments that VTP achieve superior detection performance compared to multi-view multi-person 3D pose detection methods such as Voxel Pose.Compared with Voxel Pose,VTP has a 1mm decrease in MPJPE error on the Shelf dataset,a 0.3% improvement in the average PCP3 D metric,and a 0.06 mm decrease in MPJPE error on the CMU Panoptic dataset. |
