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

Area of Science:

• Radar Signal Processing
• Deep Learning for Sensor Fusion
• Computer Vision Calibration

Background:

• Deep learning models for radar detection struggle with hardware variations, limiting their real-world application.
• Hardware heterogeneity in radar systems leads to poor generalization of detection models.

Purpose of the Study:

• To develop a unified framework for robust deep learning-based radar detection.
• To enhance cross-device generalization capabilities of radar detection systems.

Main Methods:

• Implemented a unified framework combining Perspective-n-Point (PnP) calibration with closed-loop verification for spatial alignment.
• Utilized multi-range bin calibration and chirp-wise Z-score standardization for signal statistics unification.
• Employed a lightweight global-temporal adapter (GTA) with global gating and three-point attention for feature consistency.

Main Results:

• Achieved 86.32% average precision (AP) on the ROD2021 dataset.
• Significantly outperformed the E-RODNet baseline by 22.88 percentage points.
• Demonstrated strong generalization across diverse radar platforms with only a 0.96% parameter increase.

Conclusions:

• The proposed framework effectively addresses hardware heterogeneity in radar detection.
• Combining signal-level standardization and feature-level adaptation is key to improving cross-device generalization.
• The method offers a promising solution for deploying deep learning radar detection in diverse environments.