Human pose estimation is one of the important challenges in computer vision tasks. In the past three years, various algorithms for pose estimation have emerged one after another and have achieved good results. However, from time to time, there are some keypoints that cannot be accurately predicted due to the ambiguity in images. To address such problems, we make corresponding improvements on RMPE multi-person pose estimator. On the one hand, we introduce a channel-attention mechanism to enhance the predictive ability of heat map on keypoints. On the other hand, we design a geometry constraint strategy during the test phase to control the way of keypoint connection which can compensate insufficient network functions and training samples. We carry out experiments on a standard benchmark for multi-person pose estimation, and the results prove that the proposed method can more effectively distinguish ambiguous keypoints and estimate poses.