Latent Space Metric Learning For Sidescan Sonar Place Recognition

Place recognition is an essential part of a Simultaneous Localization and Mapping, SLAM, framework. By recognizing that the AUV has returned to a previously visited place, constraints on the trajectory can be imposed leading to a reduction in navigation drift during a mission. The aim here is to recognize previously visited places solely from sidescan data using a neural network. Image patches are fed to a ResNet type architecture to extract place descriptors. Triplet loss metric learning is used for training. Results on actual sidescan data are shown.