Fast Calibration of Radar and Camera Images

Abstract

In this study, we developed a simple and efficient method using radar-measured longitudinal distance to transform the coordinates on the 2D image into the 3D camera coordinate system through the neural network, where the inputs are the millimeter-wave radar measurement and the outputs are the target's horizontal and vertical axis coordinates in the camera coordinate system. To validate the proposed method, we used corner reflectors to generate the target's coordinates with a total of 700 data sets. The experimental results show an excellent calibration performance with a mean absolute error (MAE) of target 3D coordinate location between the prediction and actual camera coordinate of 6.8 cm and a variance of 5.5 cm.