Abstract
Modern sparse Visual SLAM systems typically represent maps as collections of 3D points associated with keyframes. This representation scales poorly and lacks the geometric reasoning necessary to handle occlusions effectively. This paper introduces a voxel-map representation utilizing spatial hashing for constant-time retrieval and raycasting for visibility estimation. Integrated into the SVO pipeline, the proposed method achieves a 46% average improvement in localization accuracy on the EuRoC dataset while maintaining high computational efficiency.
Introduction
In sparse Visual SLAM (Simultaneous Localization and Mapping), the map is traditionally a set of 3D landmarks linked to specific historical camera poses, known as keyframes. To track motion, the system retrieves points from nearby keyframes to find correspondences in the current image. However, this approach faces two critical challenges:
- Scalability: As the explored environment grows, searching through an increasing number of overlapping keyframes becomes computationally prohibitive.
- Geometric Reasoning: Keyframes do not account for physical occlusions. A point might be "nearby" in the database but hidden behind a wall from the current viewpoint, leading to false matches and tracking drift.
The researchers from the University of Zurich and ETH Zurich address these issues by decoupling mapping from tracking through a voxel-based representation.
Methodology
The core of the proposed system is a Voxel Map where 3D landmarks are organized into a regular grid.
Voxel Hashing
To ensure high performance, the system employs a hash table to store only the voxels containing points. This spatial hashing allows for constant-time ($O(1)$) access to any part of the map, ensuring that query speeds do not degrade as the robot explores larger areas.
Raycasting and Occlusion Handling
To determine which landmarks are visible to the camera, the system performs raycasting. Rays are sampled from the camera center through image pixels into the voxel grid. By sampling from "near to far," the algorithm can terminate as soon as it encounters a voxel with sufficient points. This naturally filters out occluded landmarks that a keyframe-based system would erroneously include, significantly reducing "false positive" matches.
Evaluation and Results
The approach was evaluated using both simulated environments (a U-shaped corridor for scalability testing) and the real-world EuRoC MAV dataset.
Efficiency and Scalability
Simulation results demonstrate the scalability of the voxel-hashing approach. While keyframe-based query times increase linearly with the length of the environment, the voxel-map query time remains constant regardless of map size.
Localization Accuracy
On the EuRoC dataset, the voxel-map approach yielded a 46% average improvement in Root Mean Square Error (RMSE) compared to a standard 5-keyframe map setup. The system proved robust in scenarios with significant occlusions where traditional methods failed by attempting to track hidden points.
Conclusion
By replacing keyframes with a voxel-map representation, this work provides a "geometry-aware" backend for sparse SLAM. The decoupling of spatial retrieval from optimization allows for constant-time performance and superior accuracy through better occlusion handling. Future work includes optimizing voxel size sensitivity and extending the representation to support higher-level geometric structures like meshes for robot path planning.
Reference: Muglikar, M., Zhang, Z., & Scaramuzza, D. (2020). Voxel Map for Visual SLAM. IEEE International Conference on Robotics and Automation (ICRA).
Summary updated