Complex-Valued Convolutional Neural Networks for Automotive Scene Classification Based on Range-Beam-Doppler Tensors

In this work, we solve the task of automotive traffic scene classification using a deep learning approach on low-level radar data. In contrast to existing approaches using 2D camera images, the input are complex-valued 3D range-beam-doppler tensors outputted by an automotive radar. We design, train and evaluate three different networks and analyze the effects of different nuances in processing this type of data in deep neural networks. Particular attention is paid to complex-valued convolutions and their usage on complex-valued 3D tensors. The resulting networks achieve a classification accuracy above 95% on split test datasets and show the usability of automotive radar data for traffic scene classification, which can be integrated into weather robust multi-sensor systems.