SP-Pillars: An Efficient LiDAR 3D Objects Detection Framework With Multi-Scale Feature Perception and Optimization

In autonomous driving, achieving rapid detection of target categories and locations is a key technology. However, the data volume of radar point clouds is enormous, and processing efficiency becomes a limiting factor, so the balance between speed and accuracy is crucial. To address this challenge, t...

Full description

Saved in:
Bibliographic Details
Main Authors: Tingshuai Chen, Ye Yuan, Bingyang Yin, Yuanhong Liao
Format: Article
Language:English
Published: IEEE 2025-01-01
Series:IEEE Access
Subjects:
Online Access:https://ieeexplore.ieee.org/document/10978003/
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:In autonomous driving, achieving rapid detection of target categories and locations is a key technology. However, the data volume of radar point clouds is enormous, and processing efficiency becomes a limiting factor, so the balance between speed and accuracy is crucial. To address this challenge, this paper proposes a 3D object detection algorithm SP-Pillars that can effectively learn point cloud features. Firstly, a Pillar Feature Weighted Network (PFWNet) is proposed for processing point cloud information, which divides the point cloud into pillar structures and uses SPCV feature attention network to focus on its multi-level feature information. After feature extraction and dimensionality reduction, pseudo images are generated. Subsequently, before extracting pseudo image features, a multi-core perception network (PKINet) is introduced to further mine local contextual information and reduce computational complexity, enabling the backbone network to effectively learn features. The experimental evaluation results on the KITTI dataset indicate that the proposed algorithm is reliable and effective. Compared with other related algorithms, this algorithm exhibits excellent detection performance, slightly improving detection speed while maintaining high accuracy, meeting the requirements of real-time processing, and has important application value in optimizing autonomous driving technology.
ISSN:2169-3536