Motion segmentation based robust RGB-D SLAM

Publication Type:
Issue Date:
Full metadata record
While research on simultaneous localisation and mapping (SLAM) in static environments can be regarded as a significant success due to intensive work during the last several decades, conducting SLAM, especially vision-based SLAM, in dynamic scenarios is still at its early stage. Although it seems like just one step further, the dynamic elements have brought in many unanticipated challenges, including motion detection, segmentation, tracking and 3D reconstruction of both the static environments and the moving objects, in addition to the handling of motion blur. Solely based on RGB-D data with no prior knowledge available, this work centres upon proposing new practical solution frameworks for conducting SLAM in dynamic environments with efficient and robust motion segmentation methods serving as the basis. After a detailed review of the related achievements for SLAM in static environments as well as dynamic ones, and an analysis of the unaddressed challenges, four different motion segmentation methods, which include two 2-view sparse feature based motion segmentation algorithms, a 2-view semi-dense motion segmentation algorithm and an extended n-view dense moving object segmentation algorithm, are firstly proposed and their advantages, disadvantages and feasibility for different practical SLAM application scenarios are evaluated. Based on the proposed motion segmentation methods, two kinds of solution frameworks for performing SLAM in dynamic scenarios are then put forward: the first one is formulated by integrating our sparse feature based motion segmentation techniques with the available pose-graph SLAM framework; and the other one is built upon dense moving object segmentation and tailored for dense SLAM. Related simulation and experimental results have demonstrated the effectiveness of our approaches.
Please use this identifier to cite or link to this item: