ソナー画像と深層学習によるウミガメの位置・上下動の同時推定手法

―AUVによる遊泳生物観測に向けて―

抄録

In order to achieve 3D tracking of swimming organisms by an AUV (Autonomous Underwater Vehicle), we proposed a method to simultaneously estimate the position and vertical motion of sea turtles from multibeam imaging sonar images using deep learning. We trained the CNN based detector using not only the sonar images but also information about vertical motion, obtained by a data logger attached to the sea turtles. Two methods were examined: classification of the sea turtle state by the amount of depth change and regression of the pitch angle. It was also confirmed that there was a correlation between the amount of depth change and pitch angle. The effectiveness of the method was confirmed by data acquired at an aquarium.