Centralized multi-robot visual SLAM and scene reconstruction

With the rapid development of visual SLAM(Simultaneous localization and mapping), single-robot visual SLAM has been difficult to meet the efficiency requirements for mapping as well as localization. In this paper, a centralized multi-robot visual SLAM system is proposed, which consists of the following modules: tracking, mapping, communication, hybrid optimization, and loop closure detection. In the tracking module, each agent receives stereo images as well as RGB-D images, uses images to obtain the initial pose, and passes the keyframes to the mapping module. In this module, the depth image is utilized for reconstruction of single frame dense point cloud on the client side. Quantized redundant keyframes removal method will be applied on the server side. When the server receives messages from its mapping module, it will run the hybrid optimization module, which constructs the reprojection 2D error and 3D error for multi-constraint optimization, and uses the optimized pose for incremental scene reconstruction. Finally, loop closure detection module is responsible for similar scene recognition and global map fusion. We evaluate the performance of the system on datasets. The results show that this system has higher accuracy in pose estimation relative to conventional multi-robot systems. In addition, the quantized redundant keyframe removal method enables the system to have better real-time performance while maintaining accuracy. Finally, the scene reconstruction capability provides more possibilities for multi-robot navigation as well as obstacle avoidance.