A Pillar-Based 3D Object Detection Method Using 4D Radar with Radial Velocity Encoding

Le résumé fourni par la source

4D millimeter-wave radar is emerging as a cornerstone of autonomous driving perception due to its all-weather reliability and capability to measure Doppler velocity. However, applying standard LiDAR-based detectors directly to 4D radar data presents challenges due to the sparsity and noise of radar point clouds. In our experiments, we observed that a geometry-only baseline suffers from model collapse, where detection recall drops to zero during training due to gradient instability in feature extraction. To address this, we propose a robust multi-modal 3D object detection framework for sparse radar data. First, we introduce a Geometric Camera Prior module, which utilizes geometric projection constraints to filter ghost targets and background noise outside the camera's field of view. Second, we design an Enhanced Statistical Feature Encoding module. Unlike traditional methods that rely solely on absolute coordinates, our encoder explicitly incorporates radial velocity, local cluster centers, and statistical moments within each pillar. This provides the network with an effectively represent sparse object shapes. Experimental results demonstrate that the proposed method achieves a CenterRecall@2.0m of 67.42%. The ablation study on training dynamics reveals that our statistical encoding eliminates the smooth loss convergence and stable optimization.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.