Lightweight Convolutional Neural Network-based Drone Detection Using Radar Spectrograms

In this paper, we present a preliminary investigation of a lightweight convolutional neural network (CNN) architecture as a drone detection method using the long window Short-time Fourier transform (STFT) spectrograms of radar signals reflected from various drones as input. The micro-Doppler signatures within these signals manifest as horizontal Helicopter Rotation Modulation (HERM) lines in the spectrograms, providing a distinguishing feature for identifying the presence or absence of drones. The radar data used in this paper was collected in an outdoor setting with a commercially available V-band radar. The proposed network is trained using the collected limited dataset. We compare the detection performance of the proposed lightweight CNN-based detector with that of the pre-trained MobileNetV2 model after applying transfer learning with the collected dataset. Despite training with the limited dataset, the lightweight CNN model provides similar detection performance with only 35% trainable parameters compared to MobileNetV2.