Robot Localization Based on Semantic Information in Dynamic Indoor Environments with Similar Layouts

• Published in: IEEE International Conference on Robotics and Biomimetics (Online) pp. 1520 - 1525
• Main Authors: Song, Renjie, Liu, Jin, Bi, Wenfu, Zhang, Ying, Zhang, Mengyang, Zhang, Cui-Hua, Hua, Changchun
• Format: Conference Proceeding
• Language: English
• Published: IEEE 10.12.2024
• Subjects: Cameras; Filtering; Indoor environment; Laser radar; Location awareness; robot localization; Robot vision systems; Robots; Robustness; semantic scene graph; Semantics; Service robot; Service robots; similar scenarios
• ISSN: 2994-3574
• DOI: 10.1109/ROBIO64047.2024.10907541

Accurate localization in indoor spaces is critical for robotic operations. However, the common occurrence of similar scene layouts and dynamic changes in indoor environments, including repetitive furniture, obstacles, and moving objects, presents challenges for indoor localization using LiDAR. Therefore, when a robot locates indoors, it must explore and apply more diverse technological approaches. This paper investigates the potential of using widely distributed and diverse indoor furniture as references for localization in complex and dynamic indoor environments. We propose a novel indoor localization method for service robots that relies solely on a pre-constructed scene graph, a robot camera, and LiDAR. To address the issue of redundancy in image databases during the robot's environment exploration, we introduce a strategy that incorporates semantic information. By filtering similar objects across multiple scenes, this strategy effectively reduces the linear dependence between online retrieval time and database size. During rough relocalization, the robot predicts the localization range by constructing a real-time semantic scene graph, which provides a regional constraint. Subsequently, accurate results are obtained using local laser scans from a 2D LiDAR. Simulation experiments validate the effectiveness of this method, demonstrating significant improvements over probabilistic grid maps in terms of storage requirements and relocalization capabilities. The method enhances the robot's adaptability and robustness in dynamic and similar environmental scenarios.