In order to solve the problem that the impact of speed, and the poor detection performance of the common keypoints detection method caused by the uncertain number of people in the image and the relative size of different human bodies or body parts, an improved DenseNet network structure is proposed for human pose estimation. This network structure is a single-stage and end-to-end network, which uses deep convolutional neural networks for feature extraction. At the end of the convolutional network, it can get 6 different scales of feature maps by using a specific scale-transfer structure. Then the network can integrate different levels of features for multi-scale keypoints detection, which effectively improves the detection accuracy of keypoints. The bottom-up approach is adopted to ensure the processing speed of the multi-person pose estimation task. Experiments show that this method improves the mean average precision of multi-person keypoints detection by 1% compared with other general methods. It provides a new method for balancing the speed and accuracy of attitude estimation.