DOR-SLAM: A Visual SLAM Based on Dynamic Object Removal for Dynamic Environments

Abstract

vSALM is a system that enables precise positioning of a mobile robot in an unknown environment and generates corresponding 3D maps. However, the feature matching process of the visual odometry in the SLAM system is easily disturbed by dynamic objects in real-world scenarios, such as moving pedestri-ans and vehicles, which will reduce the accuracy of positioning and mapping. To address this issue., a visual SLAM system with dyna-mic object removal., called as Dynamic Object Removal SLAM (DOR-SLAM)., is proposed based on the ORB-SLAM2 system. Firstly., a dynamic probability calculation formula of pixels based on optical flow and semantic information has been proposed to determine the motion state of object. Then., a process of instance segmentation is used to segment the recognized dynamic objects in the image frames. Next., these dynamic objects will be removed by using a flow-guided video inpainting network., while inpainting the background. To improve system efficiency., a simplified bidirecti-onal optical flow calculation method is proposed and applied to the video inpainting network. The effectiveness of the proposed meth-od is evaluated with dynamic RGB-D sequences from the TUM dataset. The results of the experiments indicate that the proposed DOR-SLAM can improve the accuracy of localization in dynamic environments while maintaining high computational efficiency.