Naslov
Extrinsic and temporal calibration of heterogeneous mobile robot exteroceptive sensor systems

Autori
Peršić, Juraj

Vrsta, podvrsta i kategorija rada
Ocjenski radovi, doktorska disertacija

Fakultet
Fakultet elektrotehnike i računarstva

Mjesto
Zagreb

Datum
16.07

Godina
2021

Stranica
131

Mentor
Petrović, Ivan

Ključne riječi
sensor calibration, extrinsic calibration, temporal calibration, moving object tracking, calibration target, radar, lidar, camera, identifiability, Fisher Information Matrix, Gaussian Processes, Lie groups, on-manifold optimization

Sažetak
Robust environment perception of a mobile robot strongly relies on fusion of multiple heterogeneous sensors. Sensor fusion algorithms aim to harness all the valuable information from different sensor modalities, while circumventing their weaknesses. However, to achieve that goal, proper sensor calibration is essential. It can be achieved with offline (target-based) methods or online by relying on information from the environment. Regardless of the approach, it should provide internal sensor parameters, i.e., intrinsic calibration, accompanied with spatial and temporal relations between the sensors, i.e. extrinsic and temporal calibration. This thesis aims to solve extrinsic and temporal calibration of radar – camera – lidar systems in both offline and online manner. Extrinsic calibration tries to find transform between coordinate frames of two or more sensors. This problem is more complicated with heterogeneous sensors, due to different operating principles of the sensors and subsequently different types of data they produce. It is essential to find correct correspondences which are then used in the next step, estimation of extrinsic parameters. One of the strategies that enables robust correspondence registration is calibration based on a special target. This thesis proposes a novel calibration target suitable for accurate 6 degrees of freedom calibration of radar – camera – lidar systems. Furthermore, measurements of the target enable two-step optimization which leads to accurate extrinsic calibration. While the first step is rather standard reprojection error optimization, a novel second step based on radar cross section (RCS) is proposed. It exploits newly discovered effect of radar’s RCS estimation error related to the elevation angle. Temporal calibration tries to align timestamps of multiple sensors based on comparison of their measurements. It requires motion, either of the sensor systems or an object that the sensor system perceives. This thesis proposes a method for temporal calibration based on moving target tracking thus enabling temporal calibration of radars with other sensors such as cameras and lidars. The backbone of the proposed approach are Gaussian Processes used for continuous-time trajectory representation. It is shown that continuous-time representation is essential for accurate temporal calibration since it enables theoretically grounded temporal correspondence registration between asynchronous sensors with different frame rates. Furthermore, a novel joint spatiotemporal calibration is proposed that owes its efficiency to the Exactly Sparse Gaussian Process Regression and on-manifold optimization. Developed method enables efficient and accurate multisensor calibration that is applicable to a wide range of sensors. Online calibration uses information from the environment to generate correspondences between the sensors, thus avoiding specialized targets. This thesis proposes a novel method for online calibration based on moving object tracking applied to radar – camera – lidar systems. The method builds upon the standard detection and tracking pipeline of any autonomous stack that is performed for each sensor separately. It adds a calibration-agnostic track-to-track association scheme that works well under miscalibration. Furthermore, lightweight online decalibration detection scheme is proposed based on analytical pairwise calibration solution. Lastly, complete recalibration of the system is achieved through graph-based multisensor calibration. Combination of the proposed target-based and targetless methods enables a complete solution to calibration of radar – camera – lidar sensor systems.

Izvorni jezik
Engleski

Znanstvena područja
Elektrotehnika, Računarstvo, Temeljne tehničke znanosti

POVEZANOST RADA