An Improved RODNet for Object Detection Based on Radar and Camera Fusion

Deep learning-based radar detection often suffers from poor cross-device generalization due to hardware heterogeneity. To address this, we propose a unified framework that combines rigorous calibration with adaptive temporal modeling. The method integrates three coordinated steps: (1) ensuring precise spatial alignment via improved Perspective-n-Point (PnP) calibration with closed-loop verification; (2) unifying signal statistics through multi-range bin calibration and chirp-wise Z-score standardization; and (3) enhancing feature consistency using a lightweight global-temporal adapter (GTA) driven by global gating and three-point attention. By combining signal-level standardization with feature-level adaptation, our framework achieves 86.32% average precision (AP) on the ROD2021 dataset. It outperforms the E-RODNet baseline by 22.88 percentage points with a 0.96% parameter increase, showing strong generalization across diverse radar platforms.