Sensor calibration is the process of determining and correcting systematic measurement errors in sensing devices by comparing their outputs against known reference values or through geometric constraint solving, yielding intrinsic parameter models and extrinsic transformation matrices that map raw sensor readings to physically meaningful quantities. It is a prerequisite for sensor fusion, perception pipelines, and any application requiring quantitatively accurate environmental measurements.
Content
- Sensor calibration has been practised since the earliest scientific instrumentation, but its modern algorithmic form in robotics and computer vision was shaped by foundational work in the 1990s and 2000s. Zhang’s planar calibration method (1998–2000) democratised camera calibration by replacing precision calibration rigs with printable planar patterns, enabling accurate intrinsic estimation from as few as three images at different orientations. Simultaneously, inertial sensor calibration benefited from improved MEMS manufacturing and the availability of high-precision turntable references, enabling consumer-grade IMUs to achieve navigation-grade performance through software correction.
- Technically, camera calibration solves for the projection matrix parameters (intrinsics) and distortion polynomial coefficients by minimising reprojection error across a set of point correspondences between known 3D target geometry and observed 2D image coordinates, typically using the Levenberg-Marquardt algorithm. IMU calibration involves separating deterministic biases and scale factor errors (estimated via multi-position static procedures) from stochastic noise terms characterised by Allan variance analysis. LiDAR-camera extrinsic calibration uses edge correspondences or mutual information measures between depth and colour modalities to estimate the rigid-body transformation between sensor coordinate frames.
- The robotics ecosystem provides mature calibration toolboxes: Kalibr for camera-IMU multi-sensor calibration, MATLAB Camera Calibrator, ROS camera_calibration, and lidar_camera_calibration. Automated, target-free calibration methods using natural scene structure are increasingly deployed on mobile platforms where manual target placement is impractical. Self-calibration routines running onboard autonomous vehicles periodically re-estimate extrinsic parameters to compensate for mechanical drift from vibration and thermal expansion.
- In 2024–2025, calibration has been automated within continuous integration pipelines for autonomous vehicle fleets: vehicles perform drive-by calibration checks against roadside infrastructure, and factory calibration data is refined via fleet-scale online learning. Neural calibration approaches use self-supervised losses to jointly estimate calibration parameters and scene geometry without explicit targets. Calibration quality monitoring has emerged as a safety-critical subsystem, detecting parameter drift that could compromise perception reliability before it propagates into planning and control errors.