An Unsupervised Neural Network for Loop Detection in Underwater Visual SLAM [preprint]

Abstract

[eng] This paper presents a Neural Network aimed at robust and fast visual loop detection in underwater environments. The proposal is based on an autoencoder architecture, the decoder part being replaced by three fully connected layers. In order to help the proposed network to learn the features that define loop closings, two global image descriptors to be targeted during training are proposed. Also, a method allowing unsupervised training is presented. The experiments, performed in coastal areas of Mallorca (Spain), show the validity of our proposal and compares it favourably to previously existing methods, both based on pre-engineered and learned descriptors.