Robust Indoor SLAM based on Pedestrian Recognition by Using RGB-D Camera

The scene rigidity assumption limits the use of most feature-based visual simultaneous localization and mapping (SLAM) systems in real world. To relax this assumption and improve the robustness and accuracy, we develop a novel RGB-D visual SLAM system based on ORB-SLAM2 and dynamic object detection in this paper. The developed system is able to detect indoor pedestrians in dynamic scenarios and then deletes the corresponding features. Such a manipulation enables the subsequent visual SLAM module to generate a more accurate camera trajectory since dynamic objects are removed. Experimental results on TUM RGB-D dataset and real world scenarios show that the proposed system reduces the trajectory estimation error by one order of magnitude in comparison with the original ORB-SLAM2. In addition, the results tested on NVIDIA Jetson TX2 show that its processing speed surpasses 24 frames per second.