OpenLiDARMap presents a GNSS-free mapping framework that combines sparse public map priors with LiDAR data through scan-to-map and scan-to-scan alignment. This approach achieves georeferenced and drift-free point cloud maps.
Key Highlights
- Dual ICP-Based Matching Strategy – Integrates scan-to-scan and scan-to-map ICP matching using robust kernels (Tukey & Cauchy) for drift mitigation and local consistency across LiDAR frames.
- Pose-Graph Optimization with Map Priors – Solves a nonlinear least squares problem using Ceres Solver, balancing relative and absolute pose constraints without relying on GNSS signals.
- Sparse Reference Map from Open Data – Utilizes publicly available building footprints (e.g., OSM) and digital surface models to create lightweight georeferenced point cloud maps.
- Efficient Submap Construction – Builds and maintains a voxel-based local submap with dynamic memory pruning (100m cutoff) for real-time scan-to-scan alignment.
- ICP Robustness with small-gicp – Employs small-gicp with voxel hashing and iVox data structures for efficient nearest neighbor queries and robust alignment.
- Platform & Sensor Agnostic – Validated across multiple setups (single & multi-LiDAR, vehicles & segways) without hyperparameter tuning, highlighting broad generalization.
- Low-Latency Processing – Entire pipeline executes in ~30ms per frame on a Ryzen 7700 desktop, enabling real-time mapping potential.
- Drift-Free Performance – Demonstrates zero-drift trajectories over kilometers on challenging datasets (KITTI, NCLT, EDGAR) without loop closures or GNSS.
- Outdated Map Resilience – Performs accurately with temporal misalignment between LiDAR scans and reference maps (e.g., using 2000–2023 aerial data for 2011 sequences).
- Quantitative Gains – Achieves 10× lower ATE compared to baseline LiDAR odometry and improves Mean Map Entropy over ground truth trajectories in key benchmarks.
Paper
- Paper: https://arxiv.org/abs/2501.11111
- Opensource Code: https://github.com/TUMFTM/OpenLiDARMap
Related articles from LearnOpenCV:
- Visual SLAM: https://learnopencv.com/monocular-slam-in-python/
- LiDAR SLAM: https://learnopencv.com/lidar-slam-with-ros2/
- Building an autonomous Vehicle using CARLA: https://learnopencv.com/pid-controller-ros-2-carla/
The post OpenLiDARMap: Zero-Drift Point Cloud Mapping using Map Priors appeared first on OpenCV.