Human posture recognition based on lightweight OpenPose model

Abstract

To address the problem of slow inference in the original OpenPose pose estimation model and lower the computing power of the model, this paper first uses MobilenetV3 as backbone to make a lightweight improvement for OpenPose's network, followed by using label fusion correction to further improve the accuracy of the model. These steps make a real-time pose recognition system built on embedded devices on robots possible. The performance of the improved model is verified on the COCO dataset, and the results show that the accuracy of the improved model is not much different from the original OpenPose model, but the detection speed is improved by a factor of 4. Finally, a pose recognition model was trained on the self-built dataset using the skeleton map output from the improved model and validated on the test set, and the experiments indicated that the accuracy of the pose recognition model was 92.5%, which was real-time and suitable for various application scenarios.