摘要
实时、高性能的三维目标检测对于自动驾驶系统至关重要,其高效且准确的感知能力是关键。尽管基于激光雷达的柱体检测器因其紧凑的点云表示和计算效率而被广泛采用,但现有方法通常存在多尺度特征融合不足的问题,这限制了其在复杂场景下的鲁棒性。为解决这一挑战,本文提出了PillarAMF,这是首个集成ConvNeXt V2作为点云特征提取主干框架的模型。此外,本文引入了自适应多尺度特征金字塔网络,以增强层次化特征聚合能力,从而显著提高了检测精度。在nuScenes数据集上进行的大量实验表明,PillarAMF-Large模型实现了64.5%的mAP与70.1%的NDS,同时显著降低了计算开销与参数量。本文所提出的方法在基于柱体的检测模型中确立了新的性能标杆。
Real-time,high-performance 3D object detection is essential for autonomous driving systems,where efficient and accurate perception is critical.While LiDAR-based pillar detectors are widely adopted due to their compact point cloud representation and computational efficiency,existing methods often struggle with inadequate multi-scale feature fusion,which limits their robustness in complex scenarios.To address this challenge,this paper proposed PillarAMF,the first framework to integrate ConvNeXt V2 as the backbone for point cloud feature extraction.Furthermore,an Adaptive Multi-scale Feature Pyramid Network was introduced to improve hierarchical feature aggregation,significantly boosting detection accuracy.Extensive experiments on the nuScenes validation dataset show that the PillarAMF-Large model achieves 64.5%mAP and 70.1%NDS,while substantially reducing computational cost and the number of parameters.The proposed method sets a new state-of-the-art among pillar-based approaches.
作者
王泽梁
蔡砚刚
李仪
凌智蕾
WANG Zeliang;CAI Yangang;LI Yi;LING Zhilei(School of Information and Communication Engineering,Beijing University of Posts and Telecommunications,Beijing 100876,China)
出处
《信息传播研究》
2026年第1期21-29,共9页
Information and Communication Research
