동일 물체 판별을 위한 실시간 다중 센서 좌표계 보정 방법

Abstract

Driver Assistance Systems (DAS) are widely used for the convenience and safety of automobiles. Moving object information is an important factor for determining the speed control and lane change timing in the DAS. To obtain reliable moving object information, multi-sensor object convergence is widely used. In multi-sensor object convergence, it is important to identify the same object among the object information obtained from multi-sensor. To identify the same object, a sensor extrinsic calibration is necessary to match the different coordinate systems and setup positions of multi-sensor to the same coordinate system. Although sensor calibration is performed, the multi-sensor coordinate systems still have a slight mismatch due to sensor noise, sensor mount deformations, and vehicle dynamics changes. The coordinate system mismatch causes a small error for a nearby object, but the error increases as the distance increases. A large error causes difficulty in identifying the same object. In previous research, a method of multi-sensor coordinate system correction was used by collecting object information obtained from the sensor and estimating the position error of the object. Since the method had to process much object information at one time, it had a computational burden. Therefore, the method could not solve the mismatch of the coordinate system occurring in real time. In this paper, we propose real-time multi-sensor coordinate correction to improve the reliability of object identification. The sensor position compensation model consists of dominant factors (x_err,y_err,θ_err) that have a large influence on the longitudinal and lateral positional change of the object information among the six degrees-of-freedom parameters. The projection approach and recursive least squares are proposed to estimate the model parameters with simpler operations. The computational complexity of the proposed method is lower than that of the previous methods, and the noise effect can be reduced because the parameters are estimated by modeling the information uncertainty stochastically. In the experimental results, the proposed method reduced the computation time by up to 99.7% (at least 90.8%) compared with the Iterative Closest Point (ICP). Additionally, the rate of unidentified measurements was reduced by 94.8 % for the radar and 46.9% for the camera