Dynamic Region Detection and Removal Based on Deep Learning and Optical Flow

Simultaneous Localization And Mapping (SLAM) refers to the process in which a robot interacts with its surrounding environment to achieve self-positioning and construct an environmental map simultaneously. However, traditional SLAM algorithms assume scene rigidity, often leading to inaccurate positioning and mapping in dynamic environments. To address this limitation, we propose a visual RGB-D SLAM system that integrates ORB features, YOLOv5, Lucas-Kanade (LK) optical flow, and geometric information. During the ORB feature extraction stage, dynamic points are identified by utilizing LK optical flow to assess whether the Region of Interest (ROI) generated by the neural network is a dynamic area. Subsequently, dynamic feature points within the dynamic ROI are removed using a combination of optical flow and depth information. For dynamic points that are either missed or unrecognized by the neural network, we employ multi-view geometry for their removal. We tested the system’s performance on the public TUM dataset, demonstrating its capability to achieve SLAM functionality in dynamic environments.