基于HRNet的高效人体姿态估计算法

An Efficient Human Pose Estimation Algorithm Based on HRNet

高分辨率网络(High-Resolution Network,HRNet)因并行连接高分辨率卷积且在并行卷积中重复进行多尺度融合来维持高分辨率表示,弥补了重复上采样和下采样过程造成的信息损耗等问题而受到广泛研究。但该网络只选取最高分辨率特征表示作为输出,忽略了其他分辨率分支的特征,且对遮挡等原因产生的困难点的检测精度较低。为了提升网络对关节点定位的精度,提出了一种改进HRNet的网络模型H-HRNet(High Precision-HRNet)。提出了一种结合注意力机制的特征融合方法,对各通道提取到的信息进行融合,提高了网络对关键点的提取精度;为了解决困难点检测精度不高等问题,在特征提取网络后添加调优模块并设计多级监督机制进行监督;为了减少坐标编解码过程中的误差损失,使用一种新的解码策略。实验结果表明,模型在COCO和MPII两个数据集上的精度分别达到了76.1%和90.6%,比基线网络HRNet分别提高了1.7%和0.4%,验证了H-HRNet模型的有效性。

HRNet(High-Resolution Network)is widely studied for its parallel connection of high-resolution convolutions and repeated multi-scale fusion in parallel convolutions to maintain high-resolution representation,which makes up for the information loss caused by repeated up-sampling and down-sampling processes.However,the network only selects the highest resolution feature representation as the output,ignoring the features of other resolutions,and the detection accuracy of difficult points caused by occlusion and other reasons is low.In order to improve the accuracy of the network for joint point positioning,an improved HRNet network model,or H-HRNet(High Precision-HRNet)is proposed.First,a feature fusion method combined with the attention mechanism is proposed.This method fuses the information extracted from each channel,which improves the accuracy of key points extracted by the network;Then,in order to solve the problem of low detection accuracy of difficult points,the optimization module is added after the feature extraction network and a multi-level supervision mechanism is designed for supervision;Finally,in order to reduce the error loss in the process of coordinate encoding and decoding,a new decoding strategy is used.The experimental results show that the accuracy of this model on the COCO and MPII data set is 76.1%and 90.6%respectively,which is 1.7%and 0.4%higher than the baseline HRNet respectively,which verifies the effectiveness of the H-HRNet model.