Calibration validation

MSA engineers validate Calibration Anywhere results during platform setup, remote service, and platform hardening.

Calibration Anywhere performs automatic self-validation when deployed on-prem or for online background operation.

Main Street Autonomy engineers validate calibration accuracy by examining various relationships described below. These and other other metrics that characterize the amount and type of motion, the amount, quality, range, and distribution of features, and metrics that relate observations in each sensor relative to other sensors are used for automatic self-validation.

Sensor fusion: colorized lidar point clouds

Lidar and camera sensor fusion - on-road autonomy

Example calibration validation- colorized lidar point cloud.

Lidar points are motion-compensated, time-compensated, and painted the appropriate pixel color from the cameras. Coherent appropriately-painted scenes demonstrate:

correct camera -> lidar extrinsics
correct camera -> camera extrinsics
correct lidar -> lidar extrinsics
correct camera -> lidar time offsets
correct camera intrinsics
correct camera readout time for rolling shutter cameras

Camera/lidar/radar alignment

Example calibration validation- camera/lidar alignment.

Objects in the overlapping FOV of multi-modal sensors are evaluated to determine if they are similarly positioned in the world, which demonstrates:

correct lidar -> lidar extrinsics
correct radar -> radar extrinsics
correct camera -> camera extrinsics
correct camera -> lidar -> radar extrinsics
correct camera intrinsics
correct camera, lidar, and radar time offsets

Localization, mapping, and feature tracking

Example calibration validation- feature tracking.

This visualization shows features tracked in each camera, the corresponding map built from the features, and the robot trajectory overlaid onto the feature map, which demonstrates:

calibration suitability for accurate localization, mapping, and feature detection
correct camera -> lidar extrinsics
correct camera -> camera extrinsics for overlapping cameras
correct camera intrinsics
correct camera -> lidar time offsets
correct camera readout time for rolling shutter cameras

Rectified camera images show straight real-world edges

Example calibration validation- rectified camera images.

Camera images are evaluated to determine if real-world edges (like doors, windows, and sides of buildings) are straight in the rectified image, regardless of camera lens or configuration, which demonstrates:

correct camera intrinsics

Epipolar line measurements (reprojection error)

Example calibration validation- epipolar line measurement.

Example calibration validation- epipolar line measurement, fisheye_right zoom.

Example calibration validation- epipolar line measurement, fisheye_left zoom.

Reprojection error measures that a point in the world seen in multiple cameras has sub-pixel errors between detections in one camera and projections into another camera, which demonstrates:

correct camera -> camera extrinsics
correct camera intrinsics

Trajectory-implied gyro and acceleration vs. IMU measurements

Example calibration validation- trajectory-implied gyro vs imu measurement.

Minimal error between trajectory-implied estimates vs. IMU measurements demonstrates:

correct odometry and trajectory estimation
correct trajectory and IMU coregistration
correct IMU orientation with vehicle
correct IMU time offsets