Measuring trajectory error
seed
When developing localization systems, the output trajectory's error needs to be quantified.
The output global trajectory from the online system is compared against a reference ground-truth trajectory. The reference trajectory would be generated with better sensors + offline post processing.
Common metrics
- Translation and rotation % error
- Median & Max translation and rotation errors
- Absolute Trajectory Error & Relative Pose Errors - commonly used in SLAM papers
Visualization
- Convert the errors to vehicle frame for better interpretation
- 3 line plots of x, y, z, roll, pitch, yaw
- Map overlay of the errors
- Covariance estimate of localization vs actual error
- Compare errors of different localization inputs: wheel odometry, LIDAR registration, GPS
Resources
- https://github.com/MichaelGrupp/evo