Convolutional Autoencoder aided loop closure detection for monocular SLAM

Abstract

A correct loop closure detection is an important component of a robust SLAM (simultaneous localization and mapping) system. Loop closing refers to the process of correctly asserting that a mobile robot has returned to a previous visited location. Failing to detect a loop closure does in general not pose a threat to the positioning and mapping system of a robot, but a wrong loop closure can lead to drift of the robot and can therefore jeopardize the robot’s mission. In this paper a robust, highly parallelizable standalone algorithm for globally detecting loop closures is proposed. The algorithm is purposely built with the goal of avoiding false positives, while maintaining reasonable true positives performances. Tests conducted on the KITTI and the Scott Reef 25 dataset show that when bag-of-words approaches perform poorly, our presented approach is able to avoid wrong loop closures.