Human pose estimation and LSTM-based diver heading prediction for AUV navigation guidance

Information transmission in underwater conditions cannot depend on various media, thus an Autonomous Underwater Vehicle (AUV) is imperative to assist divers with conducting underwater operations and information transmission. However, little work has been done to promote the interaction between AUV and divers. To help resolve this issue, this paper proposes a method to help the navigation guidance of AUV by predicting the divers’ headings. In this work, the visual method is chosen to estimate pose of divers since it can adapt to complex underwater environment conditions at a lower cost than other methods. The contour information is used to obtain the keypoints of the diver which are input into the Long-Short Term Memory Network to predict the headings. The dataset which redefines the skeleton of divers in our work is original and open for future researches. Compared with the original method, the accuracy of pose estimation on this dataset is 85 $$\%$$ , an increase of 9 $$\%$$ , and with a small error of diver’s heading prediction, which indicates that our method can obtain competitive prediction results.