This repository presents the implementation of my Master's thesis project — a hybrid SLAM system that integrates the complementary strengths of ORB-SLAM3 and FAST-LIO2. The fusion ensures robust state estimation and accurate mapping using camera, LiDAR, and IMU data.

SLAM systems based solely on either visual or LiDAR data face critical limitations:

• ORB-SLAM3 struggles under:
  • High-speed motion (visual tracking loss)
  • Motion blur
  • Low-texture environments
  • Very sparse point map
• FAST-LIO2 suffers when:
  • The environment lacks rich geometric features
  • LiDAR returns become degenerate or sparse

Figure 1: FAST-LIO2 failed due to LiDAR degeneration in long corridors.

To overcome these challenges, this hybrid framework combines the visual robustness of ORB-SLAM3 with the geometric precision of FAST-LIO2, enabling accurate and reliable localization in a wide range of environments.

1. System-Level Fusion (Not Raw Measurement Fusion)

   • Two independent subsystems:
     • FAST-LIO2: LiDAR-Inertial Odometry (LIO)
     • ORB-SLAM3: Visual-Inertial SLAM (VISLAM)
   • Estimation results from each subsystem are fused, not raw sensor data.

2. Short-Term Mutual Assistance

   • When the LiDAR environment is non-degenerated, FAST-LIO2 enhances ORB-SLAM3’s mapping.
   • When LiDAR scans are degenerated, ORB-SLAM3 helps update FAST-LIO2’s IESEKF state.

3. Mid- and Long-Term Data Association

   • Mid-Term: Additional LiDAR association logic improves map continuity.
   • Long-Term: ORB-SLAM3's loop closing ensures global consistency.

Figure 2: System Overview.

Figure 3: Degeneracy based sensor fusion. Green indicates non-degeneracy, red indicates LiDAR degeneracy.

Figure 4: The mid term LiDAR association is carried out using BALM to eliminate accumulated error.

Figure 5: After a long journey, the system is able to perform loop closing and deliver precise mapping result.

Figure 6: This is the consistent and precise mapping result of a challenging dataset where the agent started from a office and transversed long corridors at ground floor and the second floor and visited production halls.

• ROS (Melodic / Noetic)
• C++14
• Eigen, OpenCV, Pangolin, PCL
• Livox SDK (for FAST-LIO2)
• DBoW2, ORB-SLAM3 dependencies

cd ~/catkin_HYBRID_SLAM
catkin_make
source devel/setup.bash
roslaunch hybrid_fusion hybrid_slam.launch

This code is released for academic use only. Refer to the individual licenses of ORB-SLAM3 and FAST-LIO2.

Yumao Liu Master Thesis – TU Darmstadt Email: liuyumao_sid@outlook.com GitHub: LiuYMUNI