Key points are not available for this paper at this time.
Most existing dynamic simultaneous localization and mapping (SLAM) methods integrating neural networks require high computational support and predefined categories of dynamic objects. This article proposes an online visual SLAM system without predefined dynamic labels for dynamic environments, which is built on ORB-SLAM3. We combine semantic information, depth information, and optical flow to distinguish foreground and background and recognize dynamic regions. The foreground points in dynamic areas are regarded as moving points. We also restore static points on moving objects by the average reprojection error between multiple frames to adapt to nonrigid motion and low-dynamic environments. In pose optimization, we define an optimization weight for every point to decrease the negative influence of potential dynamic points. The experiments on TUM RGB-D and Bonn datasets show that OVD-SLAM achieves more accurate and robust localization than other state-of-art dynamic SLAM methods. To further verify our system’s online performance and robustness, we also test it in real dynamic scenes and apply it to an augmented reality (AR) application using a handheld RGB-D camera.
He et al. (Mon,) studied this question.