Direct Sparse Mapping (1904.06577v2)

Abstract: Photometric bundle adjustment, PBA, accurately estimates geometry from video. However, current PBA systems have a temporary map that cannot manage scene reobservations. We present, DSM, a full monocular visual SLAM based on PBA. Its persistent map handles reobservations, yielding the most accurate results up to date on EuRoC for a direct method.

Review of "Direct Sparse Mapping"

The paper "Direct Sparse Mapping" presents a novel approach to direct monocular Visual Simultaneous Localization and Mapping (VSLAM) by integrating photometric bundle adjustment (PBA) with persistent mapping techniques. The authors introduce Direct Sparse Mapping (DSM) as a VSLAM system that incorporates photometric information consistently into the mapping and optimization process. This work specifically addresses the limitations of conventional VO systems that inherently rely on marginalizing map points, often leading to duplicated points and drift accumulations. In contrast, DSM emphasizes the significance of maintaining and reobserving map points, thereby enhancing robustness against common VO challenges.

Core Contributions and Methodology

1. Persistent Map Utilization: DSM effectively extends PBA approaches by creating and retaining a persistent map structure. Instead of discarding points after they exit the camera's immediate field of view, points are continuously reused, which improves accuracy in reobserved areas by minimizing drift and resolving structure inconsistencies.
2. Local Map Covisibility Window (LMCW): The paper introduces the LMCW strategy, which is a method to select active keyframes not only based on temporal proximity but also on covisibility criteria. Keyframes that remain within the local window are those that maximize map point reobservations, thereby optimizing geometric estimation using significant keyframe parallax data.
3. Robust Photometric Bundle Adjustment: A notable addition to the optimisation process is the introduction of a multiscale, coarse-to-fine PBA capable of handling large convergence radii. Furthermore, the PBA integrates a robust influence function grounded on a t-distribution, enhancing resilience against non-photoconsistent scenarios caused by occlusions and lighting variabilities.
4. T-Distribution and Outlier Management: The authors adopt a statistical framework to model photometric errors using the t-distribution, which facilitates effective outlier management. This approach adjusts to varying noise levels and scene dynamics by fitting error distributions per keyframe.
5. Quantitative Validation: Extensive experimentation on the EuRoC MAV dataset demonstrates DSM’s capabilities and improvements over existing methods. It achieves more accurate trajectory and structure estimations compared to leading state-of-the-art direct VO frameworks like DSO and LDSO. Notably, DSM managed to achieve less than 0.1m RMS ATE in 10 out of 11 sequences without a dedicated loop closure module.

Implications and Future Directions

The persistent map approach presented in this research marks a significant step in bridging the gap between VO and global consistency in SLAM. The robust handling of map point reobservations underscores how critical spatial consistency is in minimizing drift and enhancing environment map fidelity. The DSM framework offers a promising foundation for further development in autonomous navigation and real-time applications, suggesting that future work could explore integrating place recognition modules for loop closure detection, which remains an open challenge in vast, repetitive environments. Additionally, improvements in computational efficiency, such as SIMD optimizations for the PBA process, could facilitate real-time performance in embedded systems.

Conclusion

"Direct Sparse Mapping" innovatively advances direct VSLAM methodologies by leveraging persistent mapping and robust optimization techniques. Through strategic map point management and a resilient photometric error model, DSM enhances both the trajectory accuracy and 3D map reliability, achieving notable accomplishments without heavily relying on heuristic loop closure strategies. This work lays a robust groundwork for future exploration and application in the expanding domain of autonomous navigation and robotic perception.