Targetless calibration
Targetless calibration is a method of determining sensor parameters (extrinsics, intrinsics, and time offsets) from sensor data collected during normal robot motion, without using physical calibration targets or checkerboards placed in the environment.
How traditional target-based calibration works
Most calibration workflows require a physical target: a checkerboard, an AprilGrid, or a lidar-specific reflector board. The workflow is:
- Print or purchase the target
- Manually move the target through the sensor’s field of view
- Detect the target in each frame
- Solve for sensor parameters using the known geometry of the target
The target has precisely known physical dimensions so the solver has a known reference for computing sensor parameters.
How targetless calibration works
Targetless calibration uses geometric features that already exist in the environment: edges, surfaces, and structural features that sensors are able to observe during robot operation.
When a robot moves through an environment, the sensors observe the same physical scene from different poses (positions and orientations) over time. Features in the environment effectively become a large and well-distributed checkerboard enabling calibration at a variety of ranges.
A figure-8 motion pattern is effective because it results in observations of nearby features from many different poses, but many different calibration movements are suitable.
What targetless calibration produces
Targetless calibrations determine:
- Extrinsics: the 6DoF pose of each sensor relative to a common reference frame
- Intrinsics: the parameters determining how sensor data should be modified to account for internal data distortion
- Time offsets: the offset between each sensor’s reported timestamp and its true time of capture
When targetless calibration matters
Field deployments
Printing and mounting a precise checkerboard and/or AprilGrid and moving it through the field of view of every sensor on a large vehicle is not practical outside a controlled lab. Different sensing modalities require different targets and this process is best performed in a controlled environment.
Targetless calibration works wherever the robot operates, indoors or outdoors: in a parking lot, an office building, a warehouse floor, or a workshop.
Thermal cameras
AprilGrids and checkerboards are designed for visible-spectrum cameras. They are not reliably detectable in thermal infrared imagery without significant modification.
Production fleets
Larger fleets of robots will incur higher costs and longer calibration times when using target-based calibration workflows. Calibration results may be operator dependent.
Targetless calibration can be triggered autonomously during normal operations, without interrupting the production schedule.
Sensors without a shared visual modality
A system combining cameras, lidars, radars, IMUs, and wheel encoders has no single target type that all sensors can observe simultaneously. Target-based multi-modal calibration requires separate targets per sensor type, and matching them across modalities introduces additional complexity.
Targetless calibration uses the common geometry of the environment, which all sensors observe regardless of modality.
Limitations of targetless calibration
Targetless calibration requires:
- Sufficient motion: a robot that moves only in a straight line cannot provide enough geometric diversity for calibration. A figure-8, three-point turn, or similar multi-axis motion is required.
- Adequate feature density: an open field with no nearby structure provides few features for sensors to observe. The calibration motion should be performed near walls, buildings, or other feature-rich geometry.