CS-SLAM: A Lightweight Semantic SLAM Method for Dynamic Scenarios

Abstract

SLAM systems typically rely on the assumption of scene rigidity. However, in real-world applications, robots often need to operate in dynamic environments, presenting unique challenges to the stability of SLAM systems. Efficient and lightweight SLAM systems play an important role in enabling interactions between robots and their environments. To enhance their applicability in dynamic environments, a lightweight semantic dynamic SLAM framework CS-SLAM has been proposed. First, the article designs a lightweight semantic segmentation network, Cross-SegNet, to remove dynamic feature points. This network includes a lightweight feature learning module, Cross Block, which effectively detects dynamic objects while maintaining a lightweight design, thereby improving the processing efficiency and accuracy of the SLAM system. Second, a spatiotemporal consistency-based auxiliary mask algorithm has been proposed, which compares the mask mapped from the previous frame to the current frame and the mask from the Cross-SegNet segmentation. By calculating the intersection over union (IoU), segmentation results are analyzed and supplemented to enhance the efficiency of removing dynamic feature points. Qualitative and quantitative evaluations on public datasets and real-world scenarios demonstrate the robustness and effectiveness of the proposed approach comparing to existing methods compared to existing methods.