Abstract
Over the last few decades, the item recognition problem has been mostly addressed through radar techniques or computer vision algorithms. While signal/image processing has mainly fueled the recognition process in the past, machine/deep learning methods have recently stepped in, to the extent that they nowadays represent the state-of-the-art methodology. In particular, Convolutional Neural Networks are spreading worldwide as effective tools for image-based object recognition. Nevertheless, the images used to feed vision-based algorithms may not be available in some cases, and/or may have poor quality. Furthermore, they can also pose privacy issues. For these reasons, this paper investigates a novel machine learning object recognition approach based on electromagnetic backscattering in the frequency domain. In particular, a 1D Convolutional Neural Network is employed to map the collected, backscattered signals onto two classes of objects. The experimental framework is aimed at data collection through backscattering measurements in the mmWave band with signal generators and spectrum analyzers in controlled environments to ensure data reliability. Results show that the proposed method achieves 100% accuracy in object detection and 84% accuracy in object recognition. This performance makes electromagnetic-based object recognition systems a possible solution to complement vision-based techniques, or even to replace them when they turn out impractical. The findings also reveal a trade-off between accuracy and processing speed when varying signal bandwidths and frequency steps, making this approach flexible and possibly suitable for real-time applications.