Collaborative SLAM with Convolutional Neural Network-based Descriptor for Inter-Map Loop Closure Detection

Abstract

This paper introduces a novel Collaborative Si-multaneous Localization and Mapping (CSLAM) framework, enhanced with a Histogram of Oriented Gradients (HOG) de-scriptor, to improve Inter-Map Loop Closure Detection. Our framework stands out by integrating a convolutional neural network-based loop closure detection, employing the HOG de-scriptor for enhanced illumination robustness, and utilizing collaborative mapping from multiple robotic agents for refined pose estimations and mapping precision. Tested in diverse real-world fields, particularly for landmine detection, the framework demonstrates superior robustness and accuracy, outperforming the existing CCM-SLAM model. Additionally, it incorporates a transformation matrix from visual SLAM for LiDAR Point Clouds correction, showcasing its efficacy in 3D mapping and localization in GNSS-denied settings. Our results indicate that incorporating the CALC descriptor within a CSLAM system significantly enhances loop closure detection and mapping precision, marking a significant step forward in autonomous cooperative navigation and mapping in challenging environments