分布式多车协同视觉 SLAM 系统

Reliable localization and navigation are prerequisites for autonomous driving. Single-vehicle visual simultaneous localization and mapping(SLAM)enables vehicle localization in GNSS denial environment. However, the cumulative error will gradually increase with the running time, which leads to a great challenge of continuous and accurate localization. Therefore, localization can be improved by multi-vehicle collaborative visual SLAM. In this paper, a robust and lightweight distributed multi-vehicle collaborative visual SLAM system is proposed. The system uses ORB-SLAM2 as visual odometry, and uses global image descriptors NetVLAD for multi-vehicle place recognition and data association. A method based on data similarity and structural consistency is proposed to solve multi-vehicle loop-closure outlier rejection. Moreover, a distributed pose graph optimization method is proposed, which can enhance the accuracy of multi-vehicle collaborative localization. The system has been tested on our datasets collected by autonomous platform and KITTI datasets. The experiment results show that the proposed system outperforms the existing visual SLAM and collaborative SLAM.