Graph Neural Networks for Object Type Classification Based on Automotive Radar Point Clouds and Spectra

Loveneet Saini; Axel Acosta; Gor Hakobyan

ICASSP 2023accepted0 citations

Graph Neural Networks for Object Type Classification Based on Automotive Radar Point Clouds and Spectra

Loveneet Saini, Axel Acosta, Gor Hakobyan

Abstract

Object type classification (OTC) is an integral part of automotive radar perception. In this work, we propose graph neural networks (GNN) for radar OTC, which jointly process the radar reflection list and spectra. Combining the full set of features available in reflections and rich object representation in radar spectra in a graph structure allows a notable performance improvement, reducing the misclassification rate from 8.2% to 2.87%. With respect to implementation efficiency, we propose an approach that appends the learned spectral features directly into the reflection list, enabling a reduction of number of model parameters.

BibTeX

@inproceedings{icassp2023_graphneuralnetwo,
  title = {Graph Neural Networks for Object Type Classification Based on Automotive Radar Point Clouds and Spectra},
  author = {Loveneet Saini and Axel Acosta and Gor Hakobyan},
  booktitle = {ICASSP 2023},
  year = {2023}
}