Semi-Supervised GAN for Road Structure Recognition of Automotive FMCW Radar Systems

Research in autonomous driving systems technology, which is considered as a leader of the fourth industrial revolution, is defining a new era of mobility. Due to its safety and reliability in real-time traffic environments, radar, one of the most important components utilized in driverless vehicles, is actively carried out. For automotive radar systems on the road, each road environment produces superfluous echoes known as clutter, and the magnitude distribution of received radar signal varies reliance on road structures, leading to an increasing requirement for classifying the road environment and adopting a suitable target detection algorithm for each road environment characteristic. However, the classification of road environments using super-vised algorithms such as feedforward neural networks (FNN) or convolutional neural networks (CNN) requires a massive amount of training data, which is a popular impediment in deep learning. In order to tackle the problem of shortage of labeled data, in this study, we propose a semi-supervised GAN approach to recognize different road environments with auto-motive frequency-modulated continuous-wave (FMCW) radar systems. The proposed model achieves a substantial performance improvement over other existing methods, especially when only a small proportion of the training data are labeled, demonstrating the potential of the proposed Semi-GAN-based method for the challenging task of various road environments recognition.