ARC: Automotive Radar Consistency Regularization for Semi-Supervised Learning

Abstract

In recent years, radar has become a crucial part of road scene perception. In particular, radar increases sensing reliability in poor weather and lighting conditions. State-of-theart deep learning methods require training, but the labeling of radar data needed to generate the required "ground truth" is time-consuming and requires expert knowledge, due to the confusing nature of radar signals. This limits the development of accurate radar models. To address the difficulty of annotating radar datasets, we propose a novel semi-supervised learning framework we call Automotive Radar Consistency (ARC), which enables the inclusion of unlabeled radar frames during training. This improves performance without adding manual labor. Our model learns object features based on micro-Doppler signatures from a time series of radar frames. We posit that the same object in both the forward and reverse temporal directions will share similar features, and we use this as a consistency regularization to encourage the model to learn how to distinguish targets. In addition, we propose Focal JS-divergence to alleviate over-fitting and address the class imbalance problem in semisupervised learning, by focusing on the hard classes and samples in the unlabeled data. Our proposed method is a general training strategy that can be incorporated in most existing deep learning frameworks. Our experimental results on two public datasets with limited annotations show that our method significantly improves the performance of existing supervised methods. Under the same network architecture, our method outperforms the fully supervised method in object detection by as much as 36%.