Mounting Angle Prediction for Automotive Radar Using Complex-Valued Convolutional Neural Network

In advanced driver-assistance systems (ADASs), the misalignment of the mounting angle of the automotive radar significantly affects the accuracy of object detection and tracking, impacting system safety and performance. This paper introduces the Automotive Radar Alignment Detection Network (AutoRAD-Net), a novel model that leverages complex-valued convolutional neural network (CV-CNN) to address azimuth misalignment challenges in automotive radars. By utilizing complex-valued inputs, AutoRAD-Net effectively learns the physical properties of the radar data, enabling precise azimuth alignment. The model was trained and validated using mounting angle offsets ranging from -3° to +3° and exhibited errors no greater than 0.15° across all tested offsets. Moreover, it demonstrated reliable predictions even for unseen offsets, such as -1.7°, showcasing its generalization capability. The predicted offsets can then be used for physical radar alignment or integrated into compensation algorithms to enhance data interpretation accuracy in ADAS applications. This paper presents AutoRAD-Net as a practical solution for azimuth alignment, advancing radar reliability and performance in autonomous driving systems.