近年，深度学习下的多人姿态估计研究取得长足进步，但如何应对场景中的尺度变化以及如何高效检测并分组多人姿态关键点仍是当前的巨大挑战。为提升网络对多人姿态的尺度感知能力，权衡模型的速度与精度，本文在关键点检测方面提出了一种尺度感知的多人姿态估计算法，结合高分辨率表征和变形感受野设计多人关键点特征提取模块，并更新迭代网络；同时在网络头部提出热力图指导的特征融合修正策略，丰富表征的多尺度表达。在Associate Embedding上应用自适应检测网络，MSCOCO数据集的定位精度提高了6.0%，表现出对困难姿势和中小尺度关键点的检测优势。

In recent years, the research of multi-person pose estimation based on deep learning has made great progress. However, there remain huge challenges in coping with scale variation as well as efficiently detecting and grouping multi-person pose keypoints. In order to improve the scale sensibility of network and make a trade-off between speed and accuracy, we propose a scale-sensitive multi-person pose estimation algorithm involving keypoint detection network. Combining high-resolution representation and deformable receptive field, we design a multi-person keypoint feature extraction module and updated the network iteratively. Moreover, we propose a heatmap-guided feature fusion and refinement strategy to enrich multi-scale expression for the head of network. Applying adaptive detection network to the classic method Associate Embedding, with 6.0% localization improvement on MSCOCO validation accuracy, shows advantage on difficult poses and small scale keypoints detection.