CA-SLAM: Contour-Aware SLAM System Based on RGB-D Sensors in Dynamic Environment

Abstract

Visual simultaneous localization and mapping (vSLAM) has become an important key factor in modern technology and is universally used due to the affordable and available camera infrastructure. However, vSLAM systems face challenges in dynamic environments where moving objects can reduce the accuracy of both localization and mapping, especially for mobile robots. To overcome this problem, a new vSLAM framework, contour-aware SLAM (CA-SLAM) is proposed. CA-SLAM integrates the YOLOv8-SEG deep learning model for the object detection, segmentation, and then contour-aware method to utilize the minimum zero normalized cross correlation (ZNCC) across the object contour to accurately detect and create a mask of the dynamic objects. These masks are applied in the ORB-SLAM3 framework to remove dynamic features, thereby significantly improving localization accuracy. The dense mapping process utilizes the RGB-D frames, camera trajectory, and dynamic masks. Additionally, depth inpainting is applied to fill missing values and refine depth information for dense map reconstruction. Extensive validation on public datasets demonstrates that CA-SLAM achieves better and comparable trajectory accuracy in both dynamic and static environments. Furthermore, CA-SLAM demonstrates exceptional performance in dense map reconstruction, delivering highly detailed and accurate 3-D mapping even in complex environments.