DADAN: Dynamic-Augmented and Density-Aware Network for Accurate 3-D Object Detection With 4-D Radar

The cost-effective automotive 4-D radar is emerging as a feasible solution for 3-D object detection in autonomous driving scenarios. However, point clouds generated by 4-D radar are typically sparse. Current methods encounter challenges in resolving the foreground-background confusion in sparse point clouds due to: 1) the underutilization of the unique dynamic information (velocity and intensity) of 4-D radar points; and 2) the density variations of point clouds where they are near-dense-far-sparse, exacerbating the confusion. Therefore, we propose a dynamic-augmented and density-aware network (DADAN) to address these challenges by: 1) assigning high weights to foreground points based on dynamic information and enhancing their feature encoding to reduce confusion effectively; and 2) utilizing attention mechanisms along with density information to adaptively address exacerbated confusion caused by density variations. The DADAN comprises a dynamic-augmented encoding network, a density-aware backbone, and a detection head. The encoding network incorporates a dynamic-augmented encoding module (DAEM) that assigns high weights to foreground points based on velocity and intensity data to improve their feature encoding. Notably, foreground points typically exhibit distinct dynamic information compared to background points, distinguishing them from the background. The backbone network incorporates an attention mechanism along with density information to generate density-aware features. The detection head is responsible for 3-D bounding boxes and object categories. The experimental results on various public datasets demonstrate the superior performance of the proposed method compared to existing approaches. These results illustrate that the DADAN effectively addresses the issue of confusion, accurately detecting objects in complex scenes.