A stacked hourglass network (SHN) is a representative research result in human pose estimation; however, it ignores the local information of joints. Therefore, this study proposes a human pose estimation model based on an improved hourglass network. First, multiple residual modules and convolution layer with a step size of 2 are used to obtain low- to high-level features. In order to highlight the local detailed feature information, the number of residual modules and channels are adjusted as the number of network layers deepens. Then, an online difficult keypoint mining module is integrated to extract local features such as texture and shape of the occluded part. Finally, deconvolution is used to maximize the restoration of the original local features. The experimental results show that the average accuracy of the proposed model on the COCO data set reaches 74.6%. In addition, the total parameter amount is 1.5×10 ⁷, which is 5.1 percentage points higher than the average accuracy of superimposing eight SHN (8-SHN), and its total parameter amount is only 1/3 of 8-SHN.