کنفرانس مهندسی برق ایران

صفحه اصلی / سی و چهارمین کنفرانس بین المللی مهندسی برق

Automotive radar target classification using NMF-decomposed micro-Doppler signatures

نویسندگان :

Amin Aghatabar Roodbary¹ Sara Changizi² Mohammad Hassan Bastani³ Fereidoon Behnia⁴

1- دانشگاه صنعتی شریف 2- دانشگاه صنعتی اصفهان 3- دانشگاه صنعتی شریف 4- دانشگاه صنعتی شریف

کلمات کلیدی :

Automotive radar،Advanced Driver Assistance Systems،Micro-Doppler،Non-negative Matrix Factorization

چکیده :

Micro-Doppler phenomena arise in automotive radar when small, rapid motions of a target’s subcompo- nents—such as the swinging arms and legs of pedestrians, the rotating wheels and pedaling motion of bicyclists, or vibration and wheel rotation in vehicles—produce time-varying Doppler modulations superimposed on the bulk Doppler shift. These micro-motion–induced frequency fluctuations form distinctive time–frequency signatures that are highly informative for target classification. In this paper, high-resolution 77-GHz FMCW radar signals are simulated to generate micro-Doppler spectro- grams for three automotive target classes: pedestrians, vehicles, and bicyclists. To enhance robustness and extract discriminative motion patterns, the spectrograms are decomposed using non- negative matrix factorization (NMF) with both standard multi- plicative updates and an EM–ML multiplicative algorithm, evalu- ated for ranks k = 50 and k = 100. The reconstructed signatures are classified using LeNet-5 and a deeper CNN architecture (DeepLeNet), and the results are compared with classifications based on raw spectrograms. Simulation results across SNR levels from 10 to 30 dB show that NMF-based preprocessing significantly improves classification accuracy, especially at low SNR. EM–ML NMF and higher-rank decompositions provide the most discriminative features, while DeepLeNet achieves the highest overall accuracy. These findings demonstrate that coupling NMF-based low-rank decomposition with deep CNNs yields an effective and noise-resilient automotive radar target classification framework.