Scholarlink

<p>Robust and real-time 3D vehicle detection constitutes a core task in perception systems for highway autonomous driving. While state-of-the-art high-performance point-voxel fusion detectors deliver exceptional accuracy, they still struggle to achieve both high detection precision and weather resilience under stringent latency constraints (&lt; 100&#xa0;ms). To tackle this challenge, this paper proposes FN-DHV-VDHS, a point-voxel fusion network specifically designed for highway scenarios, which addresses the aforementioned issue through innovative algorithmic architecture design. Our key contributions are threefold: (1) We propose a bidirectional sampling strategy that preserves critical foreground points and global context points via dual-path intelligent sampling, while eliminating the time-consuming multi-layer feature propagation module; (2) We design a BEV-Point Context Cross Attention (BEV-PCCA) module, which substitutes high-overhead cross-scale multi-source feature aggregation with structured vector attention mechanisms; (3) We construct a vote-grid RoI pooling module that unifies global semantic aggregation and local detail extraction as parallelizable operations. Validated on KITTI and KITTI-C, FN-DHV-VDHS achieves 88.86%/81.57%/78.40% AP (easy/moderate/hard) at an inference speed of 15.2 FPS (65.2&#xa0;ms end-to-end latency), with only 88.29 G FLOPs and 14.19&#xa0;M parameters. This model outperforms PV-RCNN++ while reducing the computational load by 22.7%. Under adverse weather conditions, it attains a mean Resilience Rate (mRR) of 86.29% and 87.57% on the nuScenes-C and WOD-C robustness benchmarks, respectively, marking an 18.6% improvement over baseline voting-based methods. Ablation studies further verify the individual effectiveness and synergistic effects of each core module. Experimental results confirm that the proposed FN-DHV-VDHS offers a practical and effective algorithmic framework for real-time 3D vehicle perception in highway scenarios under complex weather conditions.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

FN-DHV-VDHS: a point-voxel fusion network with improved deep hough voting for robust real-time 3D vehicle detection in highway scenarios

  • 期刊论文

  • 发表日期：

    2026年3月13日

查看全文
  • Xingpeng Xie,
  • Chihang Zhao,
  • Xuan Li

摘要

Robust and real-time 3D vehicle detection constitutes a core task in perception systems for highway autonomous driving. While state-of-the-art high-performance point-voxel fusion detectors deliver exceptional accuracy, they still struggle to achieve both high detection precision and weather resilience under stringent latency constraints (< 100 ms). To tackle this challenge, this paper proposes FN-DHV-VDHS, a point-voxel fusion network specifically designed for highway scenarios, which addresses the aforementioned issue through innovative algorithmic architecture design. Our key contributions are threefold: (1) We propose a bidirectional sampling strategy that preserves critical foreground points and global context points via dual-path intelligent sampling, while eliminating the time-consuming multi-layer feature propagation module; (2) We design a BEV-Point Context Cross Attention (BEV-PCCA) module, which substitutes high-overhead cross-scale multi-source feature aggregation with structured vector attention mechanisms; (3) We construct a vote-grid RoI pooling module that unifies global semantic aggregation and local detail extraction as parallelizable operations. Validated on KITTI and KITTI-C, FN-DHV-VDHS achieves 88.86%/81.57%/78.40% AP (easy/moderate/hard) at an inference speed of 15.2 FPS (65.2 ms end-to-end latency), with only 88.29 G FLOPs and 14.19 M parameters. This model outperforms PV-RCNN++ while reducing the computational load by 22.7%. Under adverse weather conditions, it attains a mean Resilience Rate (mRR) of 86.29% and 87.57% on the nuScenes-C and WOD-C robustness benchmarks, respectively, marking an 18.6% improvement over baseline voting-based methods. Ablation studies further verify the individual effectiveness and synergistic effects of each core module. Experimental results confirm that the proposed FN-DHV-VDHS offers a practical and effective algorithmic framework for real-time 3D vehicle perception in highway scenarios under complex weather conditions.