DK-SLAM: Monocular Visual SLAM with Deep Keypoint Learning, Tracking and Loop-Closing (2401.09160v2)

Abstract: The performance of visual SLAM in complex, real-world scenarios is often compromised by unreliable feature extraction and matching when using handcrafted features. Although deep learning-based local features excel at capturing high-level information and perform well on matching benchmarks, they struggle with generalization in continuous motion scenes, adversely affecting loop detection accuracy. Our system employs a Model-Agnostic Meta-Learning (MAML) strategy to optimize the training of keypoint extraction networks, enhancing their adaptability to diverse environments. Additionally, we introduce a coarse-to-fine feature tracking mechanism for learned keypoints. It begins with a direct method to approximate the relative pose between consecutive frames, followed by a feature matching method for refined pose estimation. To mitigate cumulative positioning errors, DK-SLAM incorporates a novel online learning module that utilizes binary features for loop closure detection. This module dynamically identifies loop nodes within a sequence, ensuring accurate and efficient localization. Experimental evaluations on publicly available datasets demonstrate that DK-SLAM outperforms leading traditional and learning based SLAM systems, such as ORB-SLAM3 and LIFT-SLAM. These results underscore the efficacy and robustness of our DK-SLAM in varied and challenging real-world environments.

• The paper introduces a deep keypoint integration approach using Model-Agnostic Meta-Learning to adapt visual features for robust SLAM.
• The paper presents a two-stage tracking method that combines direct pose estimation with feature matching for precise localization.
• The paper demonstrates improved loop closure detection and overall performance on KITTI and EuRoC datasets compared to ORB-SLAM3.

Introduction to DK-SLAM

Simultaneous Localization and Mapping (SLAM) technology is critical to the backbone of autonomous systems, including cars, drones, and robots, enabling them to navigate and understand their environments. Recent advances have seen the rise of deep learning techniques that have improved the performance of visual SLAM systems. However, despite significant progress, challenges persist, particularly with handcrafted visual features that falter in dynamically lit or textured environments. This necessitates the development of SLAM systems that can adapt and learn robust features to operate effectively across a range of real-world scenes.

Innovation in DK-SLAM

A key innovation of the proposed DK-SLAM framework is the integration of adaptive deep local features into the SLAM process. The system employs a Meta-Learning strategy, specifically Model-Agnostic Meta-Learning (MAML), to enable fast and efficient training of deep feature extractors. This training schema significantly enhances the system's capacity to adapt to varied scenarios without compromising the learned knowledge. In a two-stage tracking approach, the framework initially employs a direct method to roughly estimate the pose between frames, followed by a feature matching refinement for exact pose estimation. By addressing the issue of cumulative positioning error, DK-SLAM presents an online learning binary feature-based module that ensures accurate detection of loops within sequences, a crucial factor for any SLAM system's efficacy.

Scalability and Performance

Experimentation showcases DK-SLAM's superior performance, particularly when compared against notable frameworks like ORB-SLAM3. The experiments involved publicly available datasets like KITTI and EuRoC, demonstrating DK-SLAM's edge in handling realistic and diverse traffic scenarios in outdoor settings, and the challenges of indoor navigation encountered by micro aerial vehicles (MAVs). Notably, DK-SLAM's online binary BoW consistently displayed an impressive recall rate in loop closure tests, underscoring the system’s robustness.

Future Directions

While DK-SLAM paves the way for more robust and adaptable visual SLAM systems, the interplay between GPU-based online front-end processes and CPU-based back-end tasks does surface efficiency issues. Moving forward, research will focus on improving system efficiency, potentially exploring knowledge distillation techniques for parameter reduction and comprehensive optimizations within the SLAM framework. With continuous enhancements, DK-SLAM is set to advance the frontier of intelligent autonomous navigation in both familiar and unstructured environments.