Processing radar signals with neural networks has shown promising results in classification and regression tasks. While processed radar data is intrinsically complex-valued, most architectures using neural networks are comprised of real-values and their arithmetic. Previous work has found that keeping the complex-valued number system and extending it into the domain of neural networks can be beneficial. In this paper, we demonstrate that in two-dimensional direction-of-arrival (DoA) estimation, complex-valued neural networks (CVNNs) show better results than real-valued neural networks (RVNNs). Real-world recordings of ten different FMCW radar devices were used to train numerous models, varying in the computational complexity and varying in data properties. Over all models trained, the best CVNN surpassed the best RVNN by 14%. In terms of model complexity, CVNNs also showed better results, both per trainable parameter and per floating point operation (FLOP). Similarly, CVNNs surpass RVNNs, both when trained with decreased data quantity and decreased data quality.     «

Processing radar signals with neural networks has shown promising results in classification and regression tasks. While processed radar data is intrinsically complex-valued, most architectures using neural networks are comprised of real-values and their arithmetic. Previous work has found that keeping the complex-valued number system and extending it into the domain of neural networks can be beneficial. In this paper, we demonstrate that in two-dimensional direction-of-arrival (DoA) estimation,...     »