A Multi-Robot Collaborative Monocular SLAM Based on Semi-Direct Method

Abstract

Aiming at the problems of low accuracy and poor robustness of current multi-robot cooperative visual simultaneous localization and mapping (SLAM) algorithms in complex environments, a multi-robot cooperative monocular SLAM algorithm based on the semi-direct method is proposed. The algorithm adopts a centralized collaborative framework. In this framework, each robot runs a direct-method based visual odometry, which can both preserves their own autonomy and enables fast and robust pose tracking on local maps. The central server uses the communication module to receive the marginalized keyframes and keypoints of all robots, and utilizes the feature method to further refine the poses of these keyframes and build reusable local sparse feature maps. These maps are fused to build a global map when they are detected to overlap. Experiments are carried out on TUM and EuRoC datasets and the results show that the algorithm in this paper has higher accuracy and robustness in co-localization.