Abstract:
Human pose estimation is a popular research topic in the field of computer vision. With the development of deep learning, human pose estimation models can accurately predict human key points. Aiming at the problems of occlusion of key points, overlapping of key points and complex background, this paper proposes a cascade pyramid model combined with attention mechanism. By integrating attention mechanism into the feature extraction network, this model can obtain richer feature information. With the help of GlobalNet and RefineNet, it can accurately locate the occluded key points. It is shown via the results on public dataset, MPII, MS COCO2017 and 3DOH50K, that this model can attain higher accuracy of human pose estimation in standard and occluded situations than previous models, and has better robustness.