Inconsistency Correction and Re-localization for Robust Collaborative SLAM

In this project, we solve two important problems in our CoSLAM system --collaborative visual SLAM involving multiple cameras moving independently on different platforms. Firstly, we consider the correction of the inconsistency between 3D maps generated by different camera groups. This issue is generated when two groups of cameras were separated before and come back to have sufficient view overlap again. We adopt a graph-based approach to optimize the camera poses and individual maps together. Each camera pose is at a vertex in the graph and constrained linear least square problems are formulated and solved to obtain the optimized camera poses and a consistent 3D map. The other addressed issue is occasional failures of the SLAM system. Motion blur will be generated by fast motion of the UAV, and video frames might be lost due to problems of the Wi-Fi transmission. Both problems cause feature tracking failures and break the SLAM system. A re-localization mechanism is designed to make the CoSLAM system robust to these unexpected tracking failures, which register the current video frame with the previous cached key-frames.