A DEEP LEARNING APPROACH FOR LOOP CLOSURE DETECTION IN VISUAL SIMULTANEOUS LOCALIZATION AND MAPPING SYSTEMS

Abstract

Mobile service robots and autonomous vehicles have been recently used in a wide spectrum of real-world applications to perform navigation tasks based on Simultaneous Localization and Mapping (SLAM), which stands for estimating sensor motion and reconstructing map structure of an unknown environment simultaneously. A fundamental task in visual-based SLAM systems, which rely solely on visual sensors, is loop closure detection. Loop closure detection reduces estimation uncertainty via identifying previously visited locations along the path of the robot. This work addresses the loop closure detection problem by proposing two deep learning-based approaches. We introduce PlaceNet, a novel architecture for feature extraction using deep auto-encoders based on fusing semantic information and multi-scale spatial features. Additionally, we present LoopNet, a novel multi-scale attention- based convolutional Siamese network as an end-to-end solution for comparing scene similarity. Finally, we perform experimental analysis on the perfor- mance of PlaceNet and LoopNet workflows in various environments under challenging conditions.