【摘要】 人体姿态估计任务依赖于视觉线索和关节间的解剖关系来定位关键点，现有的大多数基于卷积神经网络的方法能非常好地捕获视觉线索，但由于卷积操作感受野受限，往往难于关注远程上下文线索.针对上述问题，提出一种基于注意力机制的人体姿态细化网络，打破卷积神经网络固有的局部操作，扩大网络感受野，对远距离关节依赖关系进行建模.同时，由于网络在训练过程中会弱化不可见关键点，为解决此问题，使用焦点损失函数，使网络更关注于复杂关键点，如膝盖、脚踝等，加速模型收敛速度，提升网络识别能力.在同等实验条件下，分别使用目前精度最高的特征提取网络HRNet与经典特征提取网络ResNet做主干网络进行实验，以证明细化网络的鲁棒性.在人体姿态估计基准数据集MPII上表明，细化网络可以提升人体姿态估计网络性能.更多还原

【Abstract】 The human pose estimation task relied on visual clues and anatomical relationships between joints to locate key points. Most existing methods based on convolutional neural networks could capture visual clues well, but were often limited by the receptive field of the convolutional operation and struggled to focus on remote contextual clues. To address the above issues, a human pose refinement network based on attention mechanism was proposed, breaking the inherent local operation of convolutional neural networks, expanding the network′s receptive field, and modeling long-distance joint dependencies. Moreover, to address the issue of weakened invisible key points during training, a focal loss function was used to make the network pay more attention to complex key points such as knees and ankles, accelerating the convergence speed of the model and improving network recognition ability. Under the same experimental conditions, experiments were conducted using the currently most accurate feature extraction network HRNet and the classical feature extraction network ResNet as the backbone networks to demonstrate the robustness of the refinement network. On the MPII human pose estimation benchmark dataset, the refinement network was shown to improve the performance of the human pose estimation network.更多还原