Mutual Pose Recognition Based on Multiple Cues and Uncertainty Capture in Multi-robot Systems

As multi-robot systems (MRS) are utilized in more complicated environments, it becomes necessary to develop more robust methods of mutual pose recognition for pair-wise robots. Against the limitations of illumination, markers-dependence, and inexactness of partially overlapped measurements with low precision, this paper proposes a method for robust mutual pose recognition based on multiple cues, including semantic maps, depth maps, normal maps, and intensity maps. These multiple cues are fed to a devised convolutional neural network (CNN) to regress 6-DOF mutual poses. Furthermore, uncertainty capture based on error propagation through CNN is leveraged to filter out uncertain estimations. Finally, the proposed method is utilized in multi-robot SLAM (MR-SLAM) to demonstrate its feasibility and robustness. The experimental results show that the proposed method enhances the robustness of mutual pose recognition and helps to reject uncertain estimations for more accurate data fusion.