An RGB-D semantic SLAM system based on object detection in real environments

An RGB-D semantic SLAM system based on object detection in real environments

연구 분야: Strategies

학회: Signal, Image and Video Processing

In order to address the additional problems caused by the integration of object detection models in vision SLAM systems, this paper proposed an object detection-based semantic SLAM system that achieves performance comparable to that of semantic segmentation models while significantly accelerating processing. Specifically, we first proposed a dynamic object compensation method based on object detection results, leveraging a constant velocity model and multi-view geometry techniques to enhance system robustness. Second, we refined the selection of dynamic features by combining the advantages of epipolar constraints, vector constraints, and depth constraints. Finally, a region growing algorithm based on an adaptive threshold is proposed to ensure the completeness of environmental information in the dense map construction. In the experimental section, the proposed method is integrated into ORB-SLAM3 and tested using the RGB-D datasets from TUM. The results demonstrate that the proposed method effectively improves the overall performance of the original system. Furthermore, comparisons with state-of-the-art methods show that the proposed method exhibits superior robustness and performance.