Pose estimation for six-axis industrial robots based on pose distillation

Abstract

Industrial robot posture monitoring can be used to detect abnormal movements of industrial robots and thus avoid safety accidents. Most of the current research on industrial robot pose estimation has focused on improving the accuracy aspect, and there is a lack of research on its recognition efficiency. In this work, we build a model for recognizing the pose of industrial robots based on the human pose estimation model DensePose, combined with a strategy of pose distillation. Individual industrial robot images are fed into a highly accurate pose estimation network model and the model is compressed by using distillation with a fitted loss function to achieve accurate and efficient industrial robot pose recognition, laying the foundation for safe monitoring of industrial robots. The evaluation shows that compared to the original model without the FDPD method compression, the model compressed by the FDPD method is 6.5% more efficient while maintaining high accuracy.