Autonomous robots require accurate localization in GPS-denied environments like indoors or urban canyons.GNSS-INS systems are prone to failure in these conditions, while real-time SLAM often drift without loop closures Map-based localization offers a stable and accurate alternative, but it faces several key challenges:
-
Real-time performance and Scalability: Handling high-resolution 3D maps and computing scan-to-map registration efficiently.
-
Drift correction: Fusing local motion estimation with global map constraints while preserving consistency.
-
Dynamic environments: Removing or mitigating the effect of moving objects during scan matching.
-
Localization failures in feature-sparse or unmapped transition zones.
This thesis presents a robust and real-time localization framework for GNSS-denied environments by fusing LiDAR-Inertial Odometry (FAST-LIO2) with multithreaded NDT-based map matching using a sliding-window factor graph. It introduces a scalable submap management strategy and integrates dynamic object removal via deep learning, enabling consistent pose estimation even in dynamic, degraded, or feature-sparse areas. The system achieves centimeter- to decimeter-level accuracy across diverse datasets, maintaining low-latency performance suitable for real- world autonomous navigation. Extensive evaluations show that the proposed method not only surpasses standalone odometry and SLAM baselines but also outperforms recent state-of-the-art map-based localization approaches in accuracy, robustness, and scalability.
Figure 1: Complete Diagram of The Localization System
- Accurate & Drift-Free:Achieves sub-decimeter accuracy by fusing FAST-LIO2 and NDT with a sliding-window factor graph, effectively reducing drift without loop closures.
- Real-Time & Scalable: Maintains less than 23 ms latency using multithreaded NDT and dynamic submap loading. Sliding window factor graph optimization remains bounded regardless of trajectory length.
- Robust to Challenges: Dynamic object removal improves convergence, and fused graph keeps localization stable even when scan matching fails.
- Limitations & Future Work: Assumes a known initial pose and a static map. Future directions include global re-localization, adaptive
cd ~/ros2_ws/src
git clone [email protected]:eliyaskidnae/Robust-LiDAR-Inertial-Re-Localization.git
cd ~/ros2_ws
colcon build --symlink-install --cmake-args -DCMAKE_BUILD_TYPE=Release
follow this documentation to install autoware necessary packages( https://autowarefoundation.github.io/autoware-documentation/main/installation/autoware/source-installation/)
source /install/setup.bash
ros2 launch fast_lio mapping.launch.py config_file:='ouster64.yaml'
ros2 autoware_ndt_scan_matcher publish_init_pose
ros2 launch autoware_map_loader saxion_loader_super.launch.xml
ros2 launch autoware_ndt_scan_matcher ndt_scan_matcher_saxion1.launch.xml
ros2 run map_based_localization fusion_node
Click the image above or watch the demo on YouTube.