YD-SLAM: An Visual SLAM Based on Object Detection in Dynamic Environment

Visual Simultaneous Localization and Mapping (V -SLAM) is widely utilized in robot localization and navigation with the rich image information. Although remarkable progress has been made in static environments during the previous decade, research in dynamic environments is a formidable research challenge. Aiming to enhance the resilience and real-time capabilities of visual SLAM in dynamic environments, the study introduces a novel visual SLAM system that leverages the YOLOv5 object detection and ORB-SLAM2 framework. The feature points of dynamic objects are eliminated by using the bidirectional pre-matching mechanism and geometric constraints. This approach adequately diminished the influence exerted by dynamic objects on pose estimation accuracy. Meanwhile, an octomap is created for advanced tasks. The YD-SLAM conducted experiments on a publicly available TUM RGB-D dataset to evaluate its performance, and the absolute trajectory error of YD-SLAM is one order of magnitude higher than that of ORB-SLAM2. In contrast to the associated SLAM algorithms, the presented approach demonstrates enhanced operational velocity and mapping precision.