SAVF-net: sparse attention and multi-scale voxel fusion network for high-accuracy 3D measurement with LiDAR point clouds

Abstract To address the challenges of insufficient multi-scale feature representation and limited local geometric modeling in sparse point cloud 3D object detection, this paper proposes an efficient sparse voxel detection network named SAVF-Net (Sparse Attention and Voxel Fusion Network). Built upon a fully sparse detection framework, SAVF-Net introduces two core components: a Neighborhood Sparse Attention (NSA) module and a Parallel Dual-Tower Multi-Scale Voxel Fusion (MSVF-SSM) module. The NSA module leverages a local self-attention mechanism based on sparse neighborhood indices to effectively capture geometric correlations, thereby enhancing the representation of small or boundary objects. Simultaneously, the MSVF-SSM module utilizes parallel sparse convolutions with varying dilation rates, integrated with a channel-spatial dual-attention mechanism, to adaptively fuse global semantic and local geometric features. Experimental results on the KITTI dataset demonstrate that SAVF-Net significantly outperforms mainstream methods, achieving improvements of approximately 2.20% in 3D mAP and 2.14% in BEV mAP. Notably, the model yields a 2%–4% accuracy gain in small-object categories such as pedestrians and cyclists. Furthermore, real-world field tests validate the model’s generalization and stability in complex dynamic scenes. SAVF-Net achieves superior accuracy and robustness while maintaining computational efficiency, providing strong support for autonomous driving perception systems.