基于全局特征信息感知网络的二维人体姿态估计

2D human pose estimation based on global feature information perception network

针对高分辨率网络(high-resolution network,HRNet)在人体姿态估计任务中全局特征信息获取能力不足导致的人体关键点预测不够准确的问题,提出一种基于全局特征信息感知网络的人体姿态估计模型。该模型采用双分支结构,包括HRNet分支和全局特征信息感知分支,其中全局特征信息感知分支中全局特征信息获取模块将图片分割成多个序列块,再通过编码器获取其全局特征,最后通过全局特征信息融合模块将全局特征信息高效地嵌入HRNet分支中。在COCO数据集和MPII数据集上的实验结果表明,与其他传统的人体姿态估计模型相比,改进后模型的精度有明显提升。

To address the problem of inaccurate prediction of human key points caused by the lack of global feature information acquisition ability of the high-resolution network(HRNet)in human pose estimation task,a human pose estimation model based on global feature information perception network was proposed.The model adopted a dual-branch structure,including the HRNet branch and the global feature information sensing branch.In the global feature information sensing branch,the global feature information acquisition module divided the picture into multiple sequence blocks,and then obtained its global features through the encoder.Finally,the global feature information was efficiently embedded in the HRNet branch through the global feature information fusion module.The experimental results on COCO dataset and MPII dataset show that,compared with other traditional human pose estimation models,the accuracy of the improved model is significantly improved.