Scholarlink

<p>Existing LiDAR perception frameworks predominantly assume dense point clouds and balanced class distributions, limiting their practical utility when deployed with cost-effective low-channel sensors operating under severe long-tail object frequencies. Single-paradigm architectures exhibit complementary structural weaknesses: attention-based models capture global context but underspecify local geometric relations, while graph models encode spatial neighbourhoods but may fail to integrate long-range dependencies. Motivated by this gap, this study evaluates object-level 3D detection — the joint prediction of semantic class and oriented bounding box from seven-dimensional geometric descriptors extracted from annotated Velodyne VLP-16 frames — across four architectures: a Transformer FCN, an enhanced spatial GNN FCN, and two hybrid variants that integrate both paradigms through different cross-attention fusion orderings. Experiments are conducted on 2,591 real-world frames annotated for six categories (car, cyclist, human, wall, tree, cart) using file-level disjoint splits with verified zero data leakage. Cost-sensitive learning is embedded within a unified multi-task objective combining classification and 3D bounding-box regression. The best-performing hybrid, Transformer→GNN FCN (Architecture D), reaches mAP (AP@0.5) of 0.881 with mean IoU of 0.829 at ≈ 1,845 FPS (model forward-pass inference), while sustaining strong localization for safety-critical minority classes: IoU 0.919 on cart and 0.895 on human. Pairwise Wilcoxon signed-rank tests with Bonferroni correction confirm statistically significant improvements over all single-paradigm baselines (<i>p</i> &lt; 0.001), supporting hybrid attention–relational architectures as a promising direction for sparse LiDAR perception tasks relevant to ADAS and collision avoidance.</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

Cost-Sensitive Hybrid Transformer–Graph Neural Networks for Sparse LiDAR Object Detection in Automated Driving

  • 期刊论文

  • 发表日期：

    2026年5月29日

查看全文
  • Nurul Syahira Amril,
  • Galura Muhammad Suranegara,
  • Arief Suryadi Satyawan,
  • Pamungkas Daud

摘要

Existing LiDAR perception frameworks predominantly assume dense point clouds and balanced class distributions, limiting their practical utility when deployed with cost-effective low-channel sensors operating under severe long-tail object frequencies. Single-paradigm architectures exhibit complementary structural weaknesses: attention-based models capture global context but underspecify local geometric relations, while graph models encode spatial neighbourhoods but may fail to integrate long-range dependencies. Motivated by this gap, this study evaluates object-level 3D detection — the joint prediction of semantic class and oriented bounding box from seven-dimensional geometric descriptors extracted from annotated Velodyne VLP-16 frames — across four architectures: a Transformer FCN, an enhanced spatial GNN FCN, and two hybrid variants that integrate both paradigms through different cross-attention fusion orderings. Experiments are conducted on 2,591 real-world frames annotated for six categories (car, cyclist, human, wall, tree, cart) using file-level disjoint splits with verified zero data leakage. Cost-sensitive learning is embedded within a unified multi-task objective combining classification and 3D bounding-box regression. The best-performing hybrid, Transformer→GNN FCN (Architecture D), reaches mAP (AP@0.5) of 0.881 with mean IoU of 0.829 at ≈ 1,845 FPS (model forward-pass inference), while sustaining strong localization for safety-critical minority classes: IoU 0.919 on cart and 0.895 on human. Pairwise Wilcoxon signed-rank tests with Bonferroni correction confirm statistically significant improvements over all single-paradigm baselines (p < 0.001), supporting hybrid attention–relational architectures as a promising direction for sparse LiDAR perception tasks relevant to ADAS and collision avoidance.