SF-MDS: A Sparse Fusion of Mixed-Distance Scales for 3D object detection

The uneven density distribution of multimodal data caused by spatial distance variations often degrades feature-matching quality. To address this issue, this study proposes a sparse fusion of mixed-distance scales (SF-MDS) algorithm for three-dimensional (3D) object detection. First, point clouds are divided into near-field and far-field subsets according to object depth using an image-frustum separation module. Second, mixed-distance voxel features are extracted using a mixed-distance scales normalization module after applying different downsampling strategies to the separated point clouds. Next, the sparsity of voxels is exploited in combination with a sparse multi-axis window transformer to effectively capture sparse voxel features. Finally, a multi-weight dynamic fusion module is used to perform adaptive weighted matching of cross-modal features through modal correlation evaluation, thereby achieving accurate 3D object detection. Extensive experiments on the KITTI, nuScenes, and Waymo Open datasets demonstrate that SF-MDS achieves competitive detection performance across different autonomous driving benchmarks. Notably, on the Waymo Open dataset LEVEL_1 ”50m-Inf” long-distance detection task, SF-MDS achieves 71.57% mAP and 70.18% mAPH. In addition, real-world road-scene testing on a vehicle platform demonstrates the deployment feasibility of the proposed algorithm in complex traffic scenarios.