Inhaltszusammenfassung

We present a comparative analysis of the Steering-Model, a hybrid DSP–Machine Learning approach for Direction-of-Arrival (DoA) estimation, against the FFT-based spatial-spectrum technique and the Multiple Signal Classification (MUSIC) algorithm. FFT-based methods provide computational efficiency, yet their resolution is constrained by finite aperture and spectral leakage. In contrast, MUSIC achieves super-resolution through eigenvalue decomposition and subspace separation, but incurs high com...We present a comparative analysis of the Steering-Model, a hybrid DSP–Machine Learning approach for Direction-of-Arrival (DoA) estimation, against the FFT-based spatial-spectrum technique and the Multiple Signal Classification (MUSIC) algorithm. FFT-based methods provide computational efficiency, yet their resolution is constrained by finite aperture and spectral leakage. In contrast, MUSIC achieves super-resolution through eigenvalue decomposition and subspace separation, but incurs high computational cost and requires accurate source-count estimation. The proposed Steering-Model introduces a learnable Steering-Layer that replaces fixed subspace separation and steering vectors with trainable real- and complex-valued weights, formulating DoA estimation as a multi-label classification task over a predefined angular grid. The architecture comprises 52,152 parameters (305 KB), employs the Wasserstein distance as the loss function, and eliminates eigenvalue decomposition and source-count estimation requirements. Evaluation on synthetic and real-world datasets demonstrates that the Steering-Model achieves sharper spectral peak resolution than FFT-based methods across various SNR levels and reduces computational runtime by up to 80% compared to MUSIC. Notably, although trained exclusively on single-target synthetic data, the model generalizes to real radar data and multi-target scenarios, resolving sources with angular separations as small as 10° in a 12-receiver array under moderate SNR conditions. The Steering-Model offers a practical alternative that balances computational efficiency and angular resolution, suitable for embedded implementations in sensor networks.» weiterlesen» einklappen

Autoren

Klassifikation

DDC Sachgruppe:
Ingenieurwissenschaften