Physics-Guided Multi-GSO Spectral Filtering for Degradation-Aware Automotive Radar Point-Cloud Detection

Area of Science:

• Computer Vision
• Machine Learning
• Automotive Sensing

Background:

• Automotive millimeter-wave radar generates sparse point clouds with Doppler velocity and radar cross-section (RCS).
• Existing graph detectors often use a unified representation for semantic prediction and box regression, overlooking differing propagation needs.
• This limitation hinders optimal feature extraction and interpretation in complex driving scenarios.

Purpose of the Study:

• To propose and evaluate a novel residual module, multi-GSO spectral filtering (MGSF), for enhancing automotive radar point cloud processing.
• To investigate the impact of filtering radar features across geometry, Doppler, and RCS domains on detection accuracy.
• To assess the performance of MGSF-TD, a variant applying MGSF to semantic prediction and geometry-only refinement to box regression.

Main Methods:

• Developed MGSF, a residual module utilizing geometry-, Doppler-, and RCS-defined graph shift operators.
• Integrated diffusion and residual components with a node-adaptive gate within the MGSF module.
• Implemented MGSF-TD, specializing MGSF for semantic and box regression tasks on radar point clouds.
• Evaluated MGSF-TD on the RadarScenes validation set, comparing against a baseline RadarGNN checkpoint.

Main Results:

• MGSF-TD improved the official RadarGNN checkpoint's mean Average Precision (mAP) from 60.19% to 60.59% and mean foreground F1 (FG-F1) from 74.06% to 75.10%.
• Under varying radar cross-section (RCS) noise levels, MGSF-TD demonstrated increased FG-F1 margins, particularly at 20 dBsm (+2.38).
• Analysis indicated that geometry-only diffusion contributed to performance gains, with the RCS operator being most effective in low-to-moderate noise conditions.

Conclusions:

• MGSF-TD offers a balanced approach to improving automotive radar detection, providing physically interpretable operating points.
• The proposed method enhances performance on the RadarScenes dataset, particularly under specific noise conditions.
• While MGSF-TD shows gains, a widened baseline matched or exceeded its performance under severe noise, suggesting MGSF is not a universal robustness solution.