Scholarlink

<p>Flying cars in low-altitude intelligent transportation require reliable 3D scene understanding under dense occlusion, rapidly changing viewpoints, and unstable communication links, making collaborative visual computing essential for safe navigation, obstacle avoidance, and traffic coordination. Existing V2X and aerial–ground collaborative perception methods mainly align, compress, and fuse external features, but they seldom determine which ego 3D regions are both weakly observed and task-critical; this leads to redundant transmission, misaligned supplementation, and unclear blind-spot compensation objectives. We propose VGRR-Net, a reproducible visibility-gap-driven framework that first builds ego 3D features from multi-view images, estimates cell-level visibility and task importance, and re-indexes the 3D representation around task-relevant gaps. It then aligns collaborative features from ground vehicles, roadside devices, and UAVs into the flying-car coordinate system, extracts residual evidence relative to gap-aware ego features, and aggregates the residuals according to link reliability. On the self-built FCCP-3D dataset, VGRR-Net is evaluated through scene-independent testing, component ablation, hyperparameter analysis, repeated-run significance testing, communication/occlusion subset analysis, complexity comparison, and qualitative visualization, achieving superior performance in 3D object detection, semantic occupancy, and traversable-space estimation. Our code is available at <a href="https://github.com/Linlin1109/VGRR-Net.git">https://github.com/Linlin1109/VGRR-Net.git</a>.</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

VGRR-Net: a reproducible visual computing framework for visibility-gap-driven collaborative 3D perception in low-altitude flying-car traffic

  • 期刊论文

  • 发表日期：

    2026年7月4日

查看全文
  • Linlin Liu,
  • Shidong Huang,
  • Chunguan Xia,
  • Chuangchuang Chen,
  • Lipeng Zhang

摘要

Flying cars in low-altitude intelligent transportation require reliable 3D scene understanding under dense occlusion, rapidly changing viewpoints, and unstable communication links, making collaborative visual computing essential for safe navigation, obstacle avoidance, and traffic coordination. Existing V2X and aerial–ground collaborative perception methods mainly align, compress, and fuse external features, but they seldom determine which ego 3D regions are both weakly observed and task-critical; this leads to redundant transmission, misaligned supplementation, and unclear blind-spot compensation objectives. We propose VGRR-Net, a reproducible visibility-gap-driven framework that first builds ego 3D features from multi-view images, estimates cell-level visibility and task importance, and re-indexes the 3D representation around task-relevant gaps. It then aligns collaborative features from ground vehicles, roadside devices, and UAVs into the flying-car coordinate system, extracts residual evidence relative to gap-aware ego features, and aggregates the residuals according to link reliability. On the self-built FCCP-3D dataset, VGRR-Net is evaluated through scene-independent testing, component ablation, hyperparameter analysis, repeated-run significance testing, communication/occlusion subset analysis, complexity comparison, and qualitative visualization, achieving superior performance in 3D object detection, semantic occupancy, and traversable-space estimation. Our code is available at https://github.com/Linlin1109/VGRR-Net.git.