Foreground guided dual branch learning for sparse 4D radar 3D object detection.

4D imaging radar offers long-range perception and Doppler velocity measurement capabilities, while maintaining strong robustness in nighttime scenes and adverse weather conditions such as rain and fog. These advantages make it a promising sensor for 3D perception in autonomous driving. However, due to radar imaging mechanisms, 4D radar point clouds are usually highly sparse and unevenly distributed, and are often contaminated by background clutter and isolated noise points. As a result, real objects are represented by only a few valid points and can be easily overwhelmed by complex background responses. Existing methods mainly focus on global feature enhancement or local denoising, while paying insufficient attention to explicit modeling of point-wise foreground characteristics. To address these issues, this paper proposes a foreground guided dual branch learning framework for 3D object detection in sparse 4D radar scenarios. Specifically, a differentiable learning-based query generator jointly models local density and motion consistency to estimate point-wise foreground importance and generate foreground aware cues. Then, a foreground guided gated cross branch fusion mechanism selectively injects auxiliary information into potential object regions, thereby enhancing object responses and suppressing background interference. Finally, at the proposal stage, cylindrical geometric constraints and a foreground prioritized keypoint selection strategy are combined to improve local feature representations. Experimental results on VoD and TJ4DRadSet demonstrate that the proposed radar-only method achieves competitive performance, even comparable to several radar-camera fusion methods, while maintaining a favorable balance between detection accuracy and inference efficiency.