Complex-Valued and Quantized Neural Networks for In-Car Occupancy Detection Using IR-UWB Radar

In this paper we investigate per-seat occupancy detection in vehicles using ultra-wideband radar technology combined with convolutional neural networks. We aim to detect the presence of occupants and identify their seating positions. Complex-valued neural networks (CVNNs) are compared with real-valued neural networks to assess their effectiveness in handling complex-valued input data. To keep the required memory as small as possible, quantization-aware training (QAT) is employed, with fixed and trainable bit-widths for model weights and activations. Experimental results demonstrate that our approach achieves high accuracy in occupancy detection, with CVNNs offering potential benefits in scenarios unseen during the training process. Additionally, we show that, in combination with QAT, our models achieve F1-scores above 0.95 on the test set while keeping the memory required for weights and activations as low as 8.98 kB.