YER-SLAM: A Dynamic Visual SLAM Based on Object Detection with Region Growing Algorithm

Abstract

Traditional vision SLAM, which assumes a static environment, is unable to handle dynamic objects in dynamic environments. The presence of dynamic feature points decreases the localization accuracy and robustness of SLAM. In light of this, we propose YER-SLAM, a dynamic SLAM system employing object detection and region-growing algorithm, which fully utilizes the semantic and geometric information contained in images to identify and eliminate dynamic features in the environment. Firstly, we use YOLOv5 object detection to classify the image into static, dynamic, and potentially dynamic regions to generate a semantic mask that incorporates a priori knowledge. Secondly, we utilize a dynamic object detection algorithm that tightly couples object detection with epipolar constraints to initially identify dynamic features in the scene. Subsequently, we propose a novel strategy for the elimination of dynamic feature points, which integrates the acquired dynamic point with a region-growing algorithm to generate a mask for the dynamic region, thereby enabling the exclusion of the region’s feature points in the tracking process. Finally, the experimental results from the TUM dataset reveal that our algorithm achieves a 94.71% reduction in the average absolute trajectory error in highly dynamic environments, compared to ORBSLAM2. YER-SLAM presented in this paper effectively improves localization accuracy in dynamic environments.