PV-SVRCNN: a point-voxel-based 3D object detection method for sparse LiDAR point clouds
Applied Optics · 10.07.2026 · 65 (20) · 6930
DOI: 10.1364/ao.597242 ↗Kurz gefasst
Fachartikel in der Zeitschrift Applied Optics, online erschienen am 08.07.2026. Verfasst von Chunyang Wang und 3 weiteren Autor:innen. Die Arbeit nennt 20 Literaturangaben.
Abstract (Originalsprache)
3D imaging LiDAR, with its ability to actively emit and receive laser pulses to acquire high-precision 3D geometric information of scenes, has become an indispensable key sensing device in tasks such as autonomous driving and simultaneous localization and mapping. However, due to its inherent characteristics, the point cloud data acquired by 3D imaging LiDAR under long-distance detection conditions are sparse, posing significant challenges for object detection. Existing methods typically struggle to improve detection accuracy under sparse point cloud conditions. To address this problem, this paper proposes a point-voxel sparse voxel region convolutional network (PV-SVRCNN) for 3D LiDAR point cloud object detection. The method introduces a proposal-guided sector sampling (PGSS) strategy, which leverages the radial distribution characteristics of LiDAR point clouds and proposal guidance to decompose the global keypoint sampling task into parallel sector sub-tasks, significantly improving sampling efficiency while maintaining uniform keypoint distribution. Meanwhile, through spatial perceptual attention feature encoding (SPAFE), the neighborhood of keypoints is divided into sub-voxels and encoded into position-sensitive hyper feature vectors according to spatial order, enhancing the representation capability of local geometric structures in sparse point clouds. Experimental results on the KITTI dataset show that the proposed method achieves average precision (AP) of 92.94%, 79.33%, and 84.79% for car, pedestrian, and cyclist categories, respectively, significantly outperforming mainstream comparison methods and effectively improving object detection accuracy and robustness under sparse point cloud conditions, and thereby providing a reliable technical reference for the practical application of LiDAR perception systems in complex environments.
Zitieren
APA
Wang, C., Li, F., Xi, G., & Liu, X. (2026). PV-SVRCNN: a point-voxel-based 3D object detection method for sparse LiDAR point clouds. Applied Optics, 65(20), 6930. https://doi.org/10.1364/ao.597242
Vancouver
Wang C, Li F, Xi G, Liu X. PV-SVRCNN: a point-voxel-based 3D object detection method for sparse LiDAR point clouds. Applied Optics. 2026;65(20):6930. doi:10.1364/ao.597242
BibTeX
@article{Wang2026,
author = {Wang, Chunyang and Li, Fujia and Xi, Guan and Liu, Xuelian},
title = {PV-SVRCNN: a point-voxel-based 3D object detection method for sparse LiDAR point clouds},
journal = {Applied Optics},
year = {2026},
volume = {65},
number = {20},
pages = {6930},
doi = {10.1364/ao.597242}
}
EUVIMED