Real-Time Detection and Classification of Drones, Vehicles, and Humans from Radar Data Using Deep Learning

This paper proposes and systematically compares four deep-learning architectures for real-time detection and classification of drones, vehicles, and humans using range-Doppler radar data from the RAD-DAR dataset. The proposed methods are (i) a lightweight Convolutional Neural Network (CNN) baseline, (ii) a temporally aware CNN-LSTM network augmented with attention, (iii) an adapted YOLOv8 object detector, and (iv) RT-DETR-Large, an end-to-end Transformer detector tuned for real-time radar streams. All models share an identical preprocessing pipeline-power normalisation and clutter suppression so that performance differences arise solely from network design. On the held-out RAD-DAR test split, the attention-enhanced CNN raises the macro F₁ by 0.7 pp over the static CNN (95.0 % vs. 94.3 %), demonstrating the value of temporal context. Moving to detection, YOLOv8 delivers high localisation accuracy with a macro F₁ of 98.9 % 99.6% precision, 98.2% recall), while RT-DETR sets a new benchmark: 99.3 % macro F₁, 99.7 % precision, and 98.8 % recall-consistently above 97 % for each class-at >30 FPS on a single GPU. These results show that Transformer-based detectors can match or exceed convolutional counterparts across all object categories, offering a robust, real-time solution for security-critical radar-surveillance applications.