Multirobot collaborative visual SLAM and real-time map fusion method based on deep learning

Abstract

Multi-robot cooperative visual SLAM schemes often suffer from narrow data transmission bandwidth, unsatisfactory runtime computation efficiency, poor anti-interference performance when dynamic objects exist, and unreliable alignment between multiple sub-maps. Against these practical engineering defects, this study puts forward a deep learning supported cooperative positioning and mapping scheme with online map merging capability. The whole system adopts a semi-distributed framework built with graph neural networks, which only transmits compressed scene descriptors rather than raw image data to lower transmission pressure and realize joint spatial constraint solving for robot groups. A self-learning ADR-Fusion module is further proposed to adjust fusion weights according to point cloud credibility and generate unified global semantic maps. Tests deployed on the Gazebo simulation platform prove that the absolute trajectory error is reduced by 25% against the classic EKF collaborative algorithm under bandwidth restrictions, with a map matching pass rate reaching 96%. The proposed scheme effectively improves mapping precision, scene adaptability and running speed for multi-agent exploration tasks.