Autonomous 3D Mapping Using VINS-Fusion and RTAB-Map

A simulation-based study of visual-inertial SLAM performance under variable lighting for agricultural drone deployment. Built on ROS 2 Humble, PX4 SITL, and Gazebo Harmonic; targets the Holybro X500 V2 with an Intel RealSense D435i depth sensor.

Final Year Project — BEng Electrical Engineering, University of West London (2025–26)

Flight A (noon, sun elevation 60°) — 2.18 million-point dense reconstruction produced by RTAB-Map against the VINS-Fusion pose estimate.

---

Key findings

Three matched flights at sun elevations of 60°, 20°, and 10° demonstrated:

• 5.3× VINS-Fusion drift amplification at dawn (34.45 m endpoint drift) compared to noon (6.47 m) — a super-linear response to lighting degradation.
• Lighting-invariant RTAB-Map dense reconstruction at 2.0–2.2 million points across all three flights (≈5% coefficient of variation).
• Monocular scale shrinkage of 7–12% independent of lighting, attributable to insufficient IMU excitation under the smooth lawn-mower motion profile.

Endpoint drift: 6.47 m at noon, 7.86 m at dusk, 34.45 m at dawn — a 5.3× amplification concentrated in the final 10° of sun-elevation loss.

RGB input (top row) degrades across the three lighting conditions; RTAB-Map dense point clouds (bottom row) remain stable — direct empirical validation of the combined-architecture decoupling.

Full methodology, results, and engineering record are in the dissertation: dissertation/Final_year_report.pdf.

---

Architecture

The pipeline combines two SLAM systems with complementary strengths:

• VINS-Fusion — tightly-coupled visual-inertial odometry provides high-rate pose estimation
• RTAB-Map — graph-based SLAM with ICP registration produces dense 3D point clouds
• scan_cloud subscription path — RTAB-Map consumes the Gazebo depth stream directly, bypassing the RGB-depth synchronised pipeline and decoupling dense reconstruction from photometric conditions

Tech stack

ROS 2 Humble · PX4 SITL · Gazebo Harmonic · VINS-Fusion · RTAB-Map · MAVSDK · pymavlink · Ubuntu 22.04

Repository structure

├── configs/ VINS-Fusion and RTAB-Map YAML configuration ├── launch/ Pipeline launch scripts and integration nodes │ ├── run_flight.sh Top-level flight launcher (ROS 2 + PX4 + SLAM) │ ├── fly_mission.py MAVSDK lawn-mower mission │ ├── resize_node.py Image resize bridge │ ├── odom_tf_bridge.py VINS odometry → TF │ ├── save_cloud.py RTAB-Map cloud serialisation │ └── make_forest_world.py / trim_forest.py Procedural world generation ├── scripts/ Ground-truth pymavlink logger (20 Hz) ├── configs/ VINS-Fusion and RTAB-Map YAML configuration ├── data/ Raw flight data (trajectories, point clouds, metrics) │ ├── flight_A_noon/ Sun elevation 60° │ ├── flight_B_dusk/ Sun elevation 20° │ └── flight_C_dawn/ Sun elevation 10° ├── figures/ Polished dissertation figures ├── analysis/ Per-flight plots + cross-flight comparison ├── videos/ Flight footage ├── docs/ Setup, changelog, design-decisions record └── dissertation/ Full PhD-style technical report (PDF)

Running a flight

./launch/run_flight.sh A 60 0.9 0.45

See docs/SETUP.md for full prerequisites and the per-flight reproduction commands.

Data

All three flights are fully reproducible from the code in this repo. The raw outputs (ground-truth CSV, dense PCD point clouds, throughput metrics, and video) are in data/. The derived ATE/RPE analysis and cross-flight plots are in analysis/.

Author

Mustafa Yaqoobi — BEng Electrical Engineering, University of West London GitHub: @mustafayaqoobi78

License

MIT — see LICENSE.

Acknowledgements

Supervisor: Manan Abdul Khan (University of West London). The pipeline builds on open-source work from the ROS 2, PX4, VINS-Fusion, RTAB-Map, and Gazebo communities.