A Density-Aware Complementation and Sparse Fusion Network for 3-D Object Detection
Abstract
In autonomous driving, fusing Light Detection And Ranging (LiDAR) and camera information is increasingly important to achieve accurate and robust three-dimensional object detection. Due to the inherent sparsity of LiDAR point clouds, which reduces detection accuracy, conventional early fusion and region of interest-guided methods cannot effectively complement or fuse features in sparse regions, leading to suboptimal detection performance. In this paper, we propose a new density-aware complementation and sparse fusion network. Specifically, our work better complements point cloud features in sparse regions in bird’s-eye view (BEV) and realizes cross-modal fusion interactions in a unified space, both three-dimensional and two-dimensional space. First, we propose a density-aware complementation module that quantizes point cloud density directly in BEV, enhancing sparse features with geometric and semantic context to generate complementary BEV representations. Second, the sparse cross-bird’s-eye view fusion module is used to fuse the complementary features with the pseudo-point cloud features to realize cross-modal feature aggregation in BEV. Finally, the semantic point cloud rendering module is employed to complete semantic information rendering of point cloud voxel features in three-dimensional space, thereby producing more comprehensive region of interest features. On the Karlsruhe Institute of Technology and Toyota Technological Institute (KITTI) test set, our network achieves three-dimensional Average Precisions of 89.44%, 83.72% and 78.83% for easy, moderate and hard car categories, respectively. On the KITTI validation set, we further achieves a BEV mean Average Precision of 95.22%, with significant gains for heavily occluded and distant objects.