Lightweight neural networks with enhanced feature extraction for suspicious human activity detection using radar micro-Doppler signatures

The proposed models are presented with their respective architectures and computational flow algorithms, demonstrating notable improvements in detection accuracy for highly sensitive suspicious human activities (SHA) detection. Their classification performance is comprehensively evaluated through baseline comparisons, classification-report metrics, ablation studies, loss and accuracy curves, ROC analysis, and t-SNE visualizations to assess clustering effectiveness on both raw and feature embeddings. To further establish generalization and robustness, the models are validated using the University of Glasgow Radar Signatures of Human Activities dataset and through open-field real-time experiments. The experimental results demonstrate that the proposed models maintain high classification performance even when micro-Doppler (m-D) images are collected in real-world environments. The three proposed models achieve classification accuracies of 99.28%, 99.41%, and 99.54%, respectively, with Precision, Recall, and F1-Score values at or very close to 100%. These results are achieved with relatively low computational complexity, having parameter counts of only 0.57M, 0.31M, and 0.75M and computational requirements of 3.96, 1.55, and 8.05 GFLOPS, respectively. Their memory footprints are 2.28, 1.18, and 3.32 MB, while GPU inference throughput reaches 670–697 FPS, demonstrating an effective balance between accuracy, efficiency, and real-time applicability.