TL;DR
This paper presents a robust 360° visual-inertial SLAM system designed for the Hilti-Trimble SLAM Challenge 2026. By combining a dual-fisheye camera setup with a point-and-line feature frontend and a factor-graph backend, the system achieves sub-10cm mean position accuracy in difficult indoor environments characterized by low light and repetitive textures.
What Problem Does This Paper Address?
Indoor autonomous navigation remains a "boss level" challenge for robotics. Standard SLAM (Simultaneous Localization and Mapping) systems often fail when faced with:
- Low Light: Dark basements or unlit corridors where traditional feature trackers lose their grip.
- Repetitive Geometry: Long, featureless hallways where every frame looks identical (the "aliasing" problem).
- Dynamic Obstacles: People or machinery moving through the scene that can "pull" the trajectory estimate away from reality.
- Limited Field of View (FOV): Standard cameras can miss loop-closure opportunities simply because they aren't looking in the right direction.
The authors tackle these issues by leveraging a 360° dual-fisheye camera system, providing a wrap-around view that ensures the robot rarely "loses its way."
Methods
The pipeline is divided into three distinct stages:
1. Pre-processing (The "Cleanup" Crew)
Before any geometry happens, the raw images are conditioned. They use Zero-DCE++ for low-light enhancement to pull details out of the shadows and a YOLO-based detector to mask out dynamic objects (like humans), ensuring the SLAM system only tracks static landmarks.
2. The Frontend: $\sqrt{\text{VINS}}$
The "local" brain of the system is based on $\sqrt{\text{VINS}}$, a square-root filter-based visual-inertial odometry (VIO) system. It doesn't just track points; it uses:
- SuperPoint & LightGlue: Deep-learning-based point features and matching for high robustness.
- ELSED Line Segments: Tracking structural lines (like door frames or ceiling edges), which are vital in textureless corridors where points are scarce.
3. The Backend: Factor Graph Optimization
The "global" brain uses GTSAM to build a factor graph. Key features include:
- SALAD Global Descriptors: Used for detecting loop closures (recognizing a place the robot has been before).
- Cross-Camera Constraints: By using the known physical distance (baseline) between the two fisheye lenses, the system can accurately recover and maintain the metric scale of the map.
- Structureless Bundle Adjustment (BA): Jointly optimizes the robot's path and the camera's internal/external calibration parameters.
Datasets
The system was evaluated on the Hilti-Trimble SLAM Challenge 2026 dataset. This benchmark consists of several complex indoor sequences (labeled by floor, e.g., floor_EG, floor_UG1) that include floorplan priors for ground-truth comparison.
Key Insights and Ideas
- 360° Vision is a Cheat Code: Using dual fisheye lenses effectively eliminates the "lost" state. Even if one camera faces a blank wall, the other is likely seeing a feature-rich area.
- Lines + Points = Stability: Most VIO systems rely solely on points. By adding ELSED line segments, the authors provide the system with "structural anchors" that are much more reliable in man-made environments.
- The Power of Pre-processing: The inclusion of low-light enhancement (Zero-DCE++) suggests that the "vision" part of SLAM is now heavily reliant on neural image enhancement to function in real-world, non-ideal lighting.
- Metric Scale via Multi-Cam: Using the inter-camera baseline as a constraint in the factor graph is a clever way to prevent the "scale drift" that often plagues monocular or poorly calibrated stereo systems.
Key Results
The system demonstrated impressive precision across diverse indoor environments:
- Mean RMSE (Root Mean Square Error): 0.1090 meters.
- Mean Position Error: 0.0951 meters (sub-10cm accuracy).
- Coverage: Achieved nearly 100% coverage across almost all test sequences, meaning the system rarely drifted or failed to complete a run.
- Final Challenge Score: 2393.9009.
Limitations and Future Work
- Non-Causal Processing: The authors note that the backend (loop closure and BA) can revise earlier poses based on later observations. While this produces a highly accurate map, it implies the final exported trajectory is "non-causal" (i.e., it's a post-processed result rather than a strictly real-time one).
- Computational Load: While the frontend is described as "ultrafast," the combination of deep-learning features (SuperPoint/LightGlue), line detection, and complex factor graph optimization likely requires significant GPU/CPU resources.
- Manual Alignment: For the qualitative visualizations in the paper, the alignment between the estimated map and the floorplan was performed manually. Automating this global localization within a prior map remains a future step.
Summary updated