基于神经网络的毫米波雷达与视觉传感器联合标定方法

Joint calibration method of millimeter wave radar and vision sensor based on neural network

为提高毫米波雷达与摄像机联合标定的标定精度并缩短标定时间,提出了一种基于神经网络的毫米波雷达与摄像机的联合标定方法,利用神经网络的非线性映射能力直接建立毫米波雷达坐标系下的目标与其在图像像素坐标系下的映射关系;采用测试样本对所建立的神经网络进行泛化能力评估,并与传统的联合标定方法进行对比。结果表明:与传统标定方法相比,该方法操作简单,标定精度高且标定时间短,在验证集上的总体平均标定误差为0.16089像素,时间为100s,重投影后在X、Y方向上的图像残差方差分别为0.0018、0.0021,表明该方法具有很好的稳定性。

In order to improve the calibration accuracy and reduce the calibration time of millimeter wave radar and camera joint calibration,a joint calibration method was proposed based on the neural network. The relationship between the millimeter wave radar coordinate system and image pixel coordinate system was directly established by the nonlinear mapping of the neural network. The generalization ability of the established neural network was evaluated with test samples,and compared with the traditional joint calibration method. The results show that compared with the traditional calibration method,the new proposed method is easy to operate,and the calibration result is more accurate and the calibration time is shorter. The overall average calibration error on the verification set is 0.160 89 pixels and the calibration time is 100 s. The image residual variances in the X and Y directions after re-projection are 0.001 8 and 0.002 1,respectively. It illustrates that the proposed method has good stability.