A scene-adaptive descriptor for visual SLAM-based locating applications in built environments.

Due to their independence from environment instrumentation, Simultaneous Localization and Mapping (SLAM) based localization and navigation have received increasing attention and been widely applied for applications in the built environment and on construction sites. Compared with Lidar-based SLAM, the main concern with visual SLAM (vSLAM) is its effectiveness and robustness in challenging environments. As a major type of vSLAM algorithm, feature-based methods, including the state-of-the-art ORB-SLAM, rely on rich image features and a robust descriptor for matching feature correspondences across different image frames, which suffers from performance loss in environments with low-texture, low-structure areas (e.g., building corridors) or motion blur that are pretty common in practical applications. Regardless of being traditionally handcrafted or learned from recent data-driven methods such as convolutional neural networks (CNN), previous methods try to obtain an optimal fixed feature transform that works for any scenes. With the aim of improving tracking robustness in challenging environments, as opposed to such fixed feature presentation, this research proposes and explores a learning-based dynamic feature transform that is self-adaptive towards recently observed scenes, which we termed as Deep SAFT. This paper also presents the design details of an implementation of Deep SAFT working with ORB-SLAM and evaluates the modified algorithm on fifteen popular public dataset sequences. The valuation results prove the feasibility and effectiveness of SAFT for improving the matching performance of learning-based descriptors. The proposed SAFT can be integrated with existing feature-based vSLAM algorithms to provide more robust locating service for applications either in the built environment (such as facility management) or on construction sites (such as construction site safety, construction progress monitoring, and infrastructure inspection). • A scene adaptive feature transform (SAFT) is proposed to improve feature matching. • SAFT is trained online and self-adaptive to optimized description of observed scenes. • A SAFT learning network is designed, trained and tested offline on a public dataset. • An integration strategy is proposed to integrate Deep SAFT into a visual SLAM. • A Deep SAFT embedded visual SLAM is evaluated quantitatively and qualitatively. [ABSTRACT FROM AUTHOR]