Abstract
Bundle adjustment plays a vital role in feature-based monocular SLAM. In many modern SLAM pipelines, bundle adjustment is performed to estimate the 6DOF camera trajectory and 3D map (3D point cloud) from the input feature tracks. However, two fundamental weaknesses plague SLAM systems based on bundle adjustment. First, the need to carefully initialise bundle adjustment means that all variables, in particular the map, must be estimated as accurately as possible and maintained over time, which makes the overall algorithm cumbersome. Second, since estimating the 3D structure (which requires sufficient baseline) is inherent in bundle adjustment, the SLAM algorithm will encounter difficulties during periods of slow motion or pure rotational motion. We propose a different SLAM optimisation core: instead of bundle adjustment, we conduct rotation averaging to incrementally optimise only camera orientations. Given the orientations, we estimate the camera positions and 3D points via a quasi-convex formulation that can be solved efficiently and globally optimally. Our approach not only obviates the need to estimate and maintain the positions and 3D map at keyframe rate (which enables simpler SLAM systems), it is also more capable of handling slow motions or pure rotational motions.
Introduction
Feature-based monocular SLAM systems, such as ORB-SLAM, typically rely on Bundle Adjustment (BA) as their primary optimization engine. BA simultaneously estimates the 6-degree-of-freedom (6DOF) camera trajectory and the 3D point cloud by minimizing reprojection errors. While powerful, BA is a non-linear and non-convex problem, leading to significant practical challenges. It requires highly accurate initializations to avoid local minima, necessitating complex heuristics for map maintenance. Furthermore, BA struggles with the "SLAM wiggle" problem: it requires a significant translation baseline to triangulate 3D points, causing failure during slow motion or pure rotation.
The L-infinity SLAM Pipeline
This research proposes a novel SLAM optimization core, termed L-infinity SLAM, which decouples orientation estimation from position and map estimation. This separation simplifies the system and enhances robustness against challenging camera motions.
1. Online Tracking via Rotation Averaging
Instead of the simultaneous estimation performed in BA, the real-time tracking thread focuses solely on camera orientations. By utilizing relative rotations derived from essential matrices, the system performs incremental rotation averaging to determine absolute orientations. This process is computationally efficient and entirely independent of the 3D map, allowing for stable tracking even when triangulation is impossible.
2. Background Mapping via Quasi-Convex Optimization
Once orientations are established, the problem of finding camera positions and 3D points is reduced to the "Known Rotation Problem" (KRot). Traditional BA minimizes the sum of squared errors ($L_2$ norm). In contrast, L-infinity SLAM minimizes the maximum reprojection error ($L_\infty$ norm).
The $L_\infty$ formulation is quasi-convex, meaning it can be solved to global optimality without a precise initial guess. This eliminates the need for the fragile web of heuristics—such as point culling and map aggregation—typically required to "babysit" Bundle Adjustment.
Experimental Evaluation
The proposed system was evaluated against standard BA-based SLAM using two distinct datasets: the Maptek truck sequence and a custom smartphone dataset.
Robustness and Accuracy
On the Maptek dataset (1833 frames), standard BA-SLAM failed to complete the sequence due to the failure of outlier removal heuristics. In contrast, L-infinity SLAM successfully tracked the entire trajectory.
Quantitative analysis shows that L-infinity SLAM maintains low error throughout the sequence, with position errors under 0.67m and rotation errors under 0.83°. Notably, the incremental rotation averaging was found to be an order of magnitude faster than incremental BA.
Handling Pure Rotation
The "Smartphone Dataset" highlighted the primary weakness of BA. In a scenario involving slow walking and sharp turns (low baseline, high rotation), ORB-SLAM failed to initialize. L-infinity SLAM, leveraging its decoupled tracking, successfully initialized and produced a coherent 3D reconstruction.
Limitations and Conclusion
While the $L_\infty$ norm provides quasi-convexity and global optimality, it is inherently sensitive to outliers. A single erroneous feature track can significantly impact the global solution. While robust filtering can mitigate this, it remains a trade-off for the mathematical simplicity and robustness to motion types offered by the system.
L-infinity SLAM demonstrates that by moving away from the Bundle Adjustment paradigm toward quasi-convex optimization, we can create SLAM systems that are both simpler to implement and more robust to the varied motions of real-world devices.
Summary updated