Examenspresentation av Anton Gunnarsson och Eric Dahl

Examiner: Lars Hammarstrand

Abstract:
Autonomous driving has the potential to enhance safety and efficiency, also in the mining industry. The techniques used for autonomous driving rely on sensors mounted on vehicles for perception. However, it is crucial to calibrate these sensors for accurate measurements.

This thesis addresses the automatic on-site extrinsic calibration of two LiDAR sensors mounted on an autonomous quarry dumper. Due to external factors, the orientation of the LiDAR sensors is known to deviate from its nominal mounting orientation over time, leading to erroneous interpretation of the sensor readings. The goal is therefore to automatically re-calibrate the sensors by estimating the relative transformation between the LiDAR sensors. This allows compensating LiDAR measurements if the physical mounting position deviates from the nominal position.

The LiDAR sensors are mounted such that their fields of view do not overlap. Thus the problem becomes to determine the relation between two LiDAR sensors that can’t see the same things in the environment at the same time. Two solutions were tested to find the relative transformation. First, the hand-eye calibration method proposed by Daniilidis was applied to the problem, then a method based on single-LiDAR mapping followed by the merging of the produced maps was tested.

Both methods were evaluated in an autonomous driving simulator with a digital twin of a real quarry. Five different locations in the quarry were tested to find features in the environment beneficial to the algorithms in producing better calibration results. The analysis also included testing if the calibration methods worked better with artificial landmark objects as augmentation. Two different objects – cylinder, and box shapes - were tested in amounts of 5 and 10.

A motivation as to why the hand-eye calibration method might not be the best suited for this type of application is suggested. The other method is simpler and provides more accurate results of sub-degree accuracy. Calibration accuracy increases slightly with artificial landmarks, with the largest improvements in terms of translations. The simpler method was also shown to work when provided with measurement data from the real dumper.

Welcome!

Anton, Eric och Lars