Dual‐View Radar Reconstruction for Enhanced 3D Object Detection With 4D Imaging Radar Point Clouds

ABSTRACT The 4D mmWave radar point clouds are extremely sparse and noisy, which limits their geometric expressiveness for radar‐based 3D object detection. To solve the problem, a dual‐view radar reconstruction framework is proposed to simultaneously generate a bird's‐eye‐view occupancy map and a perspective depth map from raw radar measurements. The two complementary views are back‐projected and fused into a dense pseudo point cloud under cross‐view geometric consistency. A multi‐cloud fusion network is further designed to jointly encode the original radar points and the reconstructed pseudo points, producing enhanced geometric features for 3D detection. Experiments on the View‐of‐Delft dataset demonstrate that the proposed method significantly improves radar‐based 3D object detection performance, with particularly notable gains for near‐range targets. These results validate the effectiveness of dual‐view reconstruction and multi‐source point cloud fusion in enhancing radar 3D perception.