USP-SLAM: Deep Learning Based Visual SLAM with Robust Feature Extraction under Dynamic Environments

The majority of visual SLAM systems that rely on feature point methods assume that scenes are static. Consequently, addressing the interference caused by dynamic objects in reality poses a persistent challenge. This paper proposes a visual semantic SLAM system, USP-SLAM, to address this challenge. It takes advantage of a lightweight feature extraction network, modified based on the Superpoint network, to realize robust feature extraction even with images of low illumination and rapid viewpoint changes. Moreover, a modified version of the Unet network is incorporated into USP-SLAM to achieve accurate semantic segmentation of dynamic objects. Combining these two networks, the dynamic point removal module within USP-SLAM uses the optical flow method to achieve satisfying visual SLAM performance. The experimental results on the publicly available TUM RGB-D dataset show that for dynamic scene sequences, USP-SLAM improves the absolute trajectory error by 96.12% compared to its reference system ORBSLAM2. Besides, USP-SLAM also outperforms in non-dynamic scenarios, demonstrating the superiority of its system design.

I. INTRODUCTION

Simultaneous localization and mapping (SLAM) enable mobile robots incrementally build a map of unknown environment and estimate its own pose with the help of various sensors. Comparing with Light Detection and Ranging (LiDAR) based SLAM, visual SLAM (vSLAM) has become popular in recent years due to the richer texture information obtained by cameras at a lower cost. By leveraging this advantage, vSLAM may be integrated into various applications to enhance the overall performance of SLAM [1].