UAV Systems Engineering · M4-L4 · Hands-on lab (runs on your machine)
Compare visual and LiDAR localization
Put the two estimators side by side on identical data and learn where each wins.
What you will do
Using a dataset that provides both camera and LiDAR (and IMU) with ground truth, you will run a visual-inertial and a LiDAR-inertial system and compare them. The goal is not a winner, but an explanation: which segment degraded which system and why.
Before you start
- Labs “Run ORB-SLAM3”, “Run FAST-LIO2” and “Analyze trajectory drift”.
- A multi-sensor dataset with synchronised camera, LiDAR, IMU and ground truth (for example one of the public multi-sensor sets listed in the FAST-LIVO2 or LIO-SAM repositories).
Steps
- Run both systems on the same sequence with calibrated configurations and save the trajectories in TUM format.
- Align each to the ground truth with the same method; compute ATE and RPE over fixed distances.
- Plot RPE per 10 m segment for both systems on one chart against distance and mark the segments by environment type (corridor, open, dark, textured).
- For the three largest per-system error segments, look at the sensor data: feature counts for vision, the registration residual and degeneracy for LiDAR, and write the cause.
- Propose a fusion or switching strategy based on these findings (lesson M2-T16, M4-T9) and the monitor signals you would use.
Report template
| Segment | Environment | Visual error | LiDAR error | Likely cause | Monitor signal |
|---|---|---|---|---|---|
| … | long corridor | … | … | LiDAR degenerate along axis | registration eigenvalue |
| … | dark area | … | … | too few features | feature count |
| … | open field | … | … | flat ground | registration eigenvalue |
How to know it worked
- ATE and RPE for both systems with the alignment method stated.
- A per-segment comparison with causes backed by logged monitor signals.
- A fusion or switching proposal.
If it does not work
| Symptom | Likely cause | What to check |
|---|---|---|
| Trajectories in different frames | Different conventions | Align with Umeyama and check the first segments visually |
| One system fails early | Calibration or configuration error | Verify intrinsics, extrinsics and IMU noise before comparing |
| Unfair comparison | Different time ranges or poses sampled | Associate with the same timestamps and evaluate over the common interval |
References
Record your progress
Saved in this browser. This lab runs on your own machine, so the site cannot check it for you.