Cross-viewpoint robust object association and mapping for semantic scene understanding

Purpose Semantic Simultaneous Localization and Mapping (SLAM) plays a key role in robotic navigation within complex environments. While semantic objects can improve map readability and localization accuracy, challenges remain, such as viewpoint changes and occlusions that can reduce mapping accuracy and system reliability. To this end, this paper aims to introduce a comprehensive semantic RGB-D SLAM system for enhancing association and mapping. Design/methodology/approach The proposed system incorporates a multi-constraint object association mechanism that integrates semantic and geometric information to achieve robust object association across viewpoints. To further improve accuracy, the authors introduce an outlier detection method that refines object scale and pose estimation. Moreover, a semantic fusion approach based on cluster analysis is used to robustly associate and fuse objects across different perspectives, ensuring a consistent and accurate mapping. Findings The proposed method is thoroughly evaluated on the real indoor sequences, the TUM data set and simulated outdoor sequences. Experimental results demonstrate that our system significantly improves association accuracy and mapping robustness. Originality/value This paper proposes a novel semantic visual SLAM system that enhances localization accuracy and mapping robustness in complex environments. The system combines a multi-constraint object association module, a local outlier factor-based outlier removal module and an object fusion module. These modules work together to improve object association across viewpoints, filter out outliers and ensure consistent mapping through object fusion.