Lightweight Human Pose Estimation Based on Self-Attention Mechanism

To address the problems of a large number of parameters, high computational complexity and long detection time in many current human pose estimation network models, we adopt an hourglass structure to build a single-way lightweight network model with fewer parameters and shorter computation time. In order to ensure the accuracy of the model, we use window self-attention mechanism with fewer number of parameters, and this self-attention module that can effectively extract local and global information is redesigned on its basis to make the feature information learned by the model richer, and it merges with the inverted residual network architecture to become a separate module of WGNet (Global and Local Networks, a new structure that we construct), and finally WGNet is flexibly embedded into different stages of the model. The results of training and validation on COCO (Common Objects in Context) and MPII (Max Planck Institute for Informatics) datasets show that the model reduces the number of parameters by 25%, the computational complexity by 41%, and the inference time by nearly two times compared with HRFormer, which also uses the windowed self-attention mechanism, at the expense of only 3.5% accuracy.