Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Angle-Equivariant Convolutional Neural Networks for Interference Mitigation in Automotive Radar

Christian Oswald, Mate Toth, Paul Meissner, Franz Pernkopf

TL;DR

This work tackles mutual interference in automotive FMCW radar by introducing CCNN-3D, a fully convolutional, angle-equivariant neural network that operates on Range-Doppler-Angle (RDA) maps using rank-three convolutions. By sharing filters across the angle dimension, CCNN-3D achieves strong AoA generalization with fewer parameters than prior complex-valued CNNs, and it outperforms baselines on object-detection-oriented metrics. The approach demonstrates improved F1-scores and competitive EVM/PPMSE, while highlighting the trade-off between model capacity and generalization, particularly under fixed AoA interference scenarios. Overall, the method offers a robust, parameter-efficient alternative for interference mitigation in multi-antenna automotive radar, with potential for further simplification and transparency improvements in future work.

Abstract

In automotive applications, frequency modulated continuous wave (FMCW) radar is an established technology to determine the distance, velocity and angle of objects in the vicinity of the vehicle. The quality of predictions might be seriously impaired if mutual interference between radar sensors occurs. Previous work processes data from the entire receiver array in parallel to increase interference mitigation quality using neural networks (NNs). However, these architectures do not generalize well across different angles of arrival (AoAs) of interferences and objects. In this paper we introduce fully convolutional neural network (CNN) with rank-three convolutions which is able to transfer learned patterns between different AoAs. Our proposed architecture outperforms previous work while having higher robustness and a lower number of trainable parameters. We evaluate our network on a diverse data set and demonstrate its angle equivariance.

Angle-Equivariant Convolutional Neural Networks for Interference Mitigation in Automotive Radar

TL;DR

This work tackles mutual interference in automotive FMCW radar by introducing CCNN-3D, a fully convolutional, angle-equivariant neural network that operates on Range-Doppler-Angle (RDA) maps using rank-three convolutions. By sharing filters across the angle dimension, CCNN-3D achieves strong AoA generalization with fewer parameters than prior complex-valued CNNs, and it outperforms baselines on object-detection-oriented metrics. The approach demonstrates improved F1-scores and competitive EVM/PPMSE, while highlighting the trade-off between model capacity and generalization, particularly under fixed AoA interference scenarios. Overall, the method offers a robust, parameter-efficient alternative for interference mitigation in multi-antenna automotive radar, with potential for further simplification and transparency improvements in future work.

Abstract

In automotive applications, frequency modulated continuous wave (FMCW) radar is an established technology to determine the distance, velocity and angle of objects in the vicinity of the vehicle. The quality of predictions might be seriously impaired if mutual interference between radar sensors occurs. Previous work processes data from the entire receiver array in parallel to increase interference mitigation quality using neural networks (NNs). However, these architectures do not generalize well across different angles of arrival (AoAs) of interferences and objects. In this paper we introduce fully convolutional neural network (CNN) with rank-three convolutions which is able to transfer learned patterns between different AoAs. Our proposed architecture outperforms previous work while having higher robustness and a lower number of trainable parameters. We evaluate our network on a diverse data set and demonstrate its angle equivariance.
Paper Structure (11 sections, 3 equations, 3 figures, 3 tables)

This paper contains 11 sections, 3 equations, 3 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Angle-Equivariant CNNs
  3. Experimental Setup
  4. Data Set
  5. Evaluation Metrics
  6. F1-score
  7. Error Vector Magnitude (EVM)
  8. Peak Phase Mean Squared Error (PPMSE)
  9. Evaluated Architectures & Training Setup
  10. Experiments
  11. Conclusion

Figures (3)

  • Figure 1: Location of the proposed CCNN-3D in the digital signal processing chain. We denote the interfered multi-channel time domain signal as $s_I$. Its Fourier transform $S_I$ and the NN's prediction $\hat{S}_C$ are rank-three tensors with dimensions $[N_R, N_D, N_{\theta}]$, where $N_R$ is the number of samples per frequency modulation (FM) sweeps, $N_D$ the number of FM sweeps and $N_{\theta}$ the number of angle-bins.
  • Figure 2: Exemplary data sample used to train the proposed CCNN-3D, depicted as range-angle maps. The top map depicts a clean sample, where object locations are visible as peaks. Clean samples are used as optimization targets during training. The middle map shows the same sample corrupted by an interference impending from roughly 45 degrees, masking the objects. Interfered samples are used as input for CCNN-3D. The bottom map shows the CCNN-3D's prediction for the middle map. We have up-sampled the plots' angle resolution for better interpretability; Note that no such up-sampling is needed when feeding RDA-maps to CCNN-3D. Each map is scaled such that its maximum value is zero dB.
  • Figure 3: Overview of the proposed structure. The network's input, output and activations consist of $N_R$ range, $N_D$ doppler and $N_\theta$ angle-bins, and convolution kernels have size $[K_R, K_D, K_{\theta}]$.