Frequency-domain Calibration of LiDARs: An online method for extrinsic calibration of LiDAR sensors on marine vessels using IMU data and frequency analysis

Sammanfattning

LiDAR-based autonomous boat operations rely on accurate extrinsic calibration.Current practice requires manual measurement of each LiDAR’s position and orientation(pose), which is time-consuming and must be repeated if a sensor is displacedduring operation. This thesis presents an online and target-less method for automaticextrinsic calibration of LiDARs mounted at arbitrary positions on a marinevessel.The proposed method uses the vessel’s rigid-body dynamics through a four-steppipeline based on data from the LiDAR’s internal IMU. First, the sensor’s roll andpitch offsets are determined using gravity projection. Second, a straight-line accelerationmaneuver determines the yaw offset. Third, periodic yaw oscillations areanalyzed in the frequency domain to estimate the longitudinal and lateral positions.Fourth, roll oscillations are used to estimate the vertical position. The output is asix degrees of freedom pose estimate that serves as an initial guess for a fine calibrationusing the Generalized Iterative Closest Point (GICP) algorithm.The method was tested in three environments: simulations with synthetic data, acontrolled test rig and a 13 meter vessel. On the test rig, at a known sensor distanceof 27.2 cm from the center of rotation (CoR), the longitudinal, lateral andvertical positions were estimated within 1 cm, 3 cm and 1.5 cm respectively. On thevessel, longitudinal position errors were 15-17% (1.04-1.19 m), lateral errors 3-10%(0.05-0.17 m) and vertical errors 23-29% (0.33-0.42 m), all relative to the distancefrom the CoR. The main limitations were gravity compensation, uncertainty in theestimated CoR and difficulty of producing sinusoidal maneuvers without dedicatedmaneuvering software.The fine calibration using GICP was demonstrated on a sensor pair with overlappingfield of view of a nearby building. A convergence region test showed that thealgorithm converges for initial guess errors below approximately 1.5 m per axis, confirmingthat the pipeline provides a sufficiently accurate starting point for relativecalibration between LiDARs.