EffiPoseNet: Integrating Ghost Modules and CoordAttention for Efficient Real-Time Human Pose Estimation on Mobile Devices

With the popularization of mobile devices in daily life, the increasing demand for real-time human pose estimation applications puts forward higher requirements for computational efficiency and accuracy of models. Although existing attitude estimation models perform well in standard benchmarks, their deployment on mobile devices has been greatly challenged, especially due to the large number of model parameters and the large amount of computation. In this paper, we present EffiPoseNet, a lightweight human pose estimation network tailored for mobile devices, which combines the Ghost module and the CoordAttention mechanism to significantly reduce the computational burden and improve model performance. By introducing GhostCAneck module and using sparse feature maps and derived feature maps generated by Ghost module, the parameters and computational complexity of the model are effectively reduced. Besides, we designed the GhostCAblock module, which incorporates the coordinate attention mechanism to enable the model to focus on key areas in the image more precisely. Our extensive experimental evaluation of EffiPoseN et on the COCO dataset shows that EffiPoseNet remains competitive with advanced models on key performance metrics such as average accuracy (AP), with significant reductions in the number of parameters and calculated operands. EffiPoseN et demonstrates the possibility of achieving efficient and accurate human pose estimation on resource-constrained mobile devices.