Scholarlink

<p>Unmanned aerial vehicles (UAVs) have become indispensable in various applications, including post-disaster rescue and urban surveillance, due to their flexibility and ability to operate in diverse terrains. However, UAV-based object detection faces significant challenges, such as dispersed large object features, poor small object detection performance, and cross-modal differences between visible light and infrared images. To address these issues, we propose STAR-Det, a state space-enhanced and reparameterized bidirectional fusion network. STAR-Det incorporates a state space enhancement module (SS2D Block) for capturing dispersed object features and a hypergraph feature bridge module (HFBridge) combined with a reparameterized bidirectional feature pyramid network (RepBFPAN) for efficient small object detection. Experimental results on the VisDrone and HIT-UAV datasets demonstrate that STAR-Det achieves 44.9% and 85.6% mAP, respectively, outperforming existing models by 6.7% and 2%, while maintaining a good balance between parameter count and computational complexity. These findings highlight STAR-Det’s effectiveness in complex UAV scenarios and its potential for practical applications. Notably, the consistent performance improvements across both visible light (VisDrone) and infrared thermal imaging (HIT-UAV) datasets validate STAR-Det’s robust cross-domain adaptability, making it particularly suitable for round-the-clock UAV surveillance applications such as urban monitoring and disaster response operations. Our code is available at: <a href="https://github.com/gdzengzeng/Star-Det">https://github.com/gdzengzeng/Star-Det</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

Enhancing multi-domain UAV object detection via state space modeling and hypergraph feature fusion

  • 期刊论文

  • 发表日期：

    2026年1月19日

查看全文
  • Weizeng Qin,
  • Zhaolong Zeng,
  • Xiaofeng Hu

摘要

Unmanned aerial vehicles (UAVs) have become indispensable in various applications, including post-disaster rescue and urban surveillance, due to their flexibility and ability to operate in diverse terrains. However, UAV-based object detection faces significant challenges, such as dispersed large object features, poor small object detection performance, and cross-modal differences between visible light and infrared images. To address these issues, we propose STAR-Det, a state space-enhanced and reparameterized bidirectional fusion network. STAR-Det incorporates a state space enhancement module (SS2D Block) for capturing dispersed object features and a hypergraph feature bridge module (HFBridge) combined with a reparameterized bidirectional feature pyramid network (RepBFPAN) for efficient small object detection. Experimental results on the VisDrone and HIT-UAV datasets demonstrate that STAR-Det achieves 44.9% and 85.6% mAP, respectively, outperforming existing models by 6.7% and 2%, while maintaining a good balance between parameter count and computational complexity. These findings highlight STAR-Det’s effectiveness in complex UAV scenarios and its potential for practical applications. Notably, the consistent performance improvements across both visible light (VisDrone) and infrared thermal imaging (HIT-UAV) datasets validate STAR-Det’s robust cross-domain adaptability, making it particularly suitable for round-the-clock UAV surveillance applications such as urban monitoring and disaster response operations. Our code is available at: https://github.com/gdzengzeng/Star-Det.