DEY-SLAM: Visual Simultaneous Localization and Mapping in Highly Dynamic Environments Using Depth Information, Epipolar Geometry and Deep Learning

Articles in Press

Document Type : Original Article

Authors

Department of Electrical and Computer Engineering, Isfahan University of Technology, Isfahan, 84156-83111, I.R. of Iran

Abstract

Simultaneous Localization and Mapping (SLAM) refers to a process in which a mobile sensor (such as a robot) simultaneously estimates its pose within an unknown environment while constructing a map of that environment. In recent years, traditional SLAM algorithms have achieved a relative level of maturity in static environments; however, the assumption of environmental staticity is often not valid in practical and industrial applications. As a result, these algorithms are affected by dynamic factors in the environment, which leads to increased localization errors. In this paper, we introduce the DEY-SLAM algorithm to address the impact of dynamic objects in the environment. In the first stage, DEY-SLAM employs the semantic network YOLOv11-seg to remove potentially dynamic feature points. Subsequently, a filter based on depth information and an epipolar geometry test is applied to eliminate the remaining suspicious dynamic points from the SLAM estimation process, thereby improving localization accuracy. Experimental evaluations conducted on sequences from the TUM RGB-D dataset demonstrate that DEY-SLAM achieves satisfactory performance compared to recent SLAM algorithms. Furthermore, in complex scenes of this dataset, the proposed method reduces the RMSE of the absolute trajectory error obtained by the classical ORB-SLAM3 algorithm and improves mapping accuracy by up to approximately 90% on these sequences.

Keywords

Main Subjects

Tabriz Journal of Electrical Engineering

Articles in Press, Corrected Proof
Available Online from 29 July 2026

Files

Share

How to cite

Statistics