Abstract:To achieve more accurate positioning of human key points, this study proposes a multi-stage selective pass-based human pose estimation network model MGNet. Firstly, by taking HRNet as the backbone network, this model removes the fourth stage in HRNet, and realizes the connection of multi-stage resolution via parallel connection, aiming to achieve the balance between complexity and detection ability of the model. Secondly, a new dynamic upsampling method is developed to replace the bilinear interpolation method, which can make full use of the rich semantic information in the feature map and improve the feature fusion ability of the network. Then, a hybrid efficient channel attention mechanism is designed to enhance the global and local channel information of the model, and a lightweight multi-stage selective pass-based CA attention module (MCAM) is proposed. This module is employed to replace the residual block of HRNet, which ensures the model performance, reduces the model complexity, and ultimately achieves light weight. Finally, a depth-aware feature enhancement module is designed to make the final output feature contain both local depth information and global information. Experimental results show that MGNet yields 72.1% accuracy on the COCO2017 verification set, and compared with the backbone network HRNet, the parameter amount is reduced by 70%, with the accuracy loss of 1.3%. Meanwhile, on the MPII dataset, average accuracy of 89.4% is achieved, with a large amount of computation reduced. By comprehensively considering the performance on the COCO2017 and MPII datasets, MGNet not only yields optimal accuracy in the lightweight model, but also has comparable or better performance compared to the large model.