Scholarlink

<p>Unmanned aerial vehicles (UAVs) drive demand for advanced 3D perception. However, existing approaches face critical limitations: dense Bird’s-Eye View (BEV) methods incur high computational costs and poor long-range performance, while sparse query methods exhibit deficiencies in dynamic detection.​ To address these, we propose <i>Depth-aware Position Embedding TRansformation</i> (DPETR), a novel 3D detection paradigm. The DPETR integrates three core modules: a hybrid depth estimation network that combines probability buckets with direct regression for precise absolute depth; a 3D adaptive query generation module enhanced through auxiliary 2D proposals; and Geometry Self-Attention module, which leverages depth maps as priors to guide the fusion of RGB and depth features.​ Experiments on real-world (nuScenes) and simulated UAV (Carla Drone) datasets validate its superiority. The DPETR boosts mAP by 3.2% on nuScenes validation and 8.2% on CDrone test versus baselines. It shows significant improvements in long-range detection accuracy and cross-scenario generalization.</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

DPETR: a new image-based depth-aware position embedding transformation model for UAV 3D object detection

  • 期刊论文

  • 发表日期：

    2025年10月27日

查看全文
  • Hongjie Zhou,
  • Hongya Tuo,
  • Zhongliang Jing,
  • Lindong Wang,
  • Heshan Dong

摘要

Unmanned aerial vehicles (UAVs) drive demand for advanced 3D perception. However, existing approaches face critical limitations: dense Bird’s-Eye View (BEV) methods incur high computational costs and poor long-range performance, while sparse query methods exhibit deficiencies in dynamic detection.​ To address these, we propose Depth-aware Position Embedding TRansformation (DPETR), a novel 3D detection paradigm. The DPETR integrates three core modules: a hybrid depth estimation network that combines probability buckets with direct regression for precise absolute depth; a 3D adaptive query generation module enhanced through auxiliary 2D proposals; and Geometry Self-Attention module, which leverages depth maps as priors to guide the fusion of RGB and depth features.​ Experiments on real-world (nuScenes) and simulated UAV (Carla Drone) datasets validate its superiority. The DPETR boosts mAP by 3.2% on nuScenes validation and 8.2% on CDrone test versus baselines. It shows significant improvements in long-range detection accuracy and cross-scenario generalization.