Evaluation of DSP-driven and ML methods for radar-based classification of human, vehicles, and UAVs
TL;DR
The results indicate that DSP-driven ML frameworks constitute a viable and interpretable alternative to deep learning for radar target classification, particularly in settings where deep learning is constrained by limited data or deployment requirements.
Abstract
This paper investigates Frequency-Modulated Continuous-Wave (FMCW) radar target classification using an interpretable digital signal processing (DSP) feature-engineering pipeline combined with classical machine learning (ML) models. Range--Doppler patches corresponding to three target classes—car, drone, and human—are extracted from the Real-Doppler RADDAR dataset and localized via CFAR-based detection. A compact feature set encompassing texture, spectral, time-frequency, wavelet, and shape descriptors is derived from each patch and standardized prior to learning. Five supervised classifiers, namely Random Forest, XGBoost, support vector machine with RBF kernel, logistic regression, and Gaussian Naive Bayes, are evaluated using a 70/30 train-test split with five-fold cross-validation. Among the considered models, XGBoost achieves the highest test accuracy of 83%. Feature-importance analysis reveals discriminative attributes aligned with micro-Doppler phenomenology, including spectral bandwidth, rotor-harmonic energy, and time-frequency modulation patterns. These results indicate that DSP-driven ML frameworks constitute a viable and interpretable alternative to deep learning for radar target classification, particularly in settings where deep learning is constrained by limited data or deployment requirements.