Introduction
The proliferation of autonomous and semi-autonomous vehicles has led to increasingly crowded radio frequency environments. Most automotive radars utilize Frequency Modulated Continuous Wave (FMCW) signals, which are susceptible to Radio Frequency Interference (RFI) from neighboring vehicles. This interference manifests as a "wall of noise," significantly raising the noise floor and potentially masking critical targets such as pedestrians or other vehicles.
While classical mitigation techniques like "zeroing" (discarding affected signal segments) are common, they often result in substantial data loss. Recent deep learning approaches have improved signal recovery but frequently struggle to preserve the phase information of the radar signal. In radar processing, phase is indispensable for accurate velocity estimation, a critical requirement for safe autonomous navigation.
Methodology: Unfolded RPCA and ROC-AE Blocks
This research proposes a novel architecture that integrates the mathematical rigor of classical signal processing with the adaptive power of deep learning. The approach centers on two primary innovations:
- Unfolded Robust PCA (RPCA): Robust PCA is traditionally used to decompose a signal into a low-rank component (structured interference) and a sparse component (radar targets). By "unfolding" the iterative optimization algorithm into neural network layers, the model learns optimal decomposition parameters directly from data, rather than relying on manual tuning.
- Residual Overcomplete Auto-Encoder (ROC-AE) Blocks: The authors replace standard convolutional layers within the unfolded RPCA framework with specialized ROC-AE blocks. These blocks are "overcomplete," utilizing 32 filters to process the 2-channel complex-valued input (real and imaginary parts). This expansion into a higher-dimensional latent space better captures intricate signal features. Furthermore, "residual" skip connections are employed to facilitate the training of deeper architectures by mitigating gradient issues.
By treating the real and imaginary components as separate input channels, the network maintains the complex-valued nature of the radar data, enabling the precise recovery of phase information.
Experimental Results
The model was evaluated using the ARIM dataset—a synthetic database of 48,000 samples simulating diverse automotive scenarios—and validated with real-world data from an NXP 77 GHz radar transceiver.
The proposed "Unfolded RPCA ROC-AE" model demonstrates superior performance across all key metrics. It achieved a Mean Absolute Error (MAE) of 0.55 dB for amplitude and 2.55 degrees for phase on the test set, significantly outperforming both classical zeroing and baseline unfolded RPCA models.
Beyond accuracy, the model is highly efficient. On a GPU, it processes signals in 40ms, compared to 66ms for a standard Deep Fully Convolutional Network (FCN). Its performance on embedded hardware, specifically the NVIDIA Jetson Xavier NX, confirms its suitability for real-time automotive deployment.
Conclusion
By combining physics-inspired "unfolding" with overcomplete residual blocks, this work provides a robust solution for radar interference mitigation that preserves vital phase information. Future research will focus on extending the model to handle multiple simultaneous interference sources and further optimizing for low-latency embedded execution.
Summary updated