Self-Supervised Learning from FMCW Radar Signals for vehicles detection

Overview & Synopsis

Radar sensing is a promising modality for autonomous perception due to its robustness under adverse weather and lighting conditions. However, most existing radar pipelines rely on heavily preprocessed representations that may discard valuable information from raw FMCW signals. In this work, we introduce TRAME, a self-supervised framework combining ViT-MAE reconstruction and temporal modeling to learn structured radar representations directly from noisy low-level radar observations. We also extend the RaDICaL dataset with a reproducible cross-modal annotation framework based on synchronized RGB-D projections in the radar domain. Our method improves radar patch classification over CNN-based baselines, while temporal modeling further boosts detection on reconstructed radar representations. These results demonstrate the effectiveness of combining self-supervised radar reconstruction with cross-modal supervision for robust radar-only perception.