Scholarlink

<p>Unmanned aerial vehicle (UAV) imagery in emergency rescue scenarios faces severe challenges under complex conditions, including extreme scale variations, dense occlusions, and complex background interference (ruins, debris, vegetation, smoke plumes), which cause significant feature loss for small personnel and vehicle targets, thereby restricting real-time search and rescue decisions. This work presents efficient gated sparse multi-scale net (EGSMnet), a lightweight, parameter-efficient framework for personnel and vehicle detection in UAV emergency rescue imagery via adaptive feature enhancement. The spatially adaptive gated efficient attention (SAGE) mechanism generates spatially adaptive saliency masks to suppress disaster-area background noise while preserving the spatial integrity of small personnel and vehicle targets, unlike conventional attention mechanisms that apply uniform global pooling. The cross-scale multi-path adaptive aggregation (CMAA) network constructs ultra-fine-grained feature propagation paths with scale-adaptive asymmetric feature injection to avoid feature dilution during deep propagation. Enhanced ghost convolution (EGConv) achieves adaptive feature extraction with reduced parameters, overcoming the fixed feature-splitting ratio of standard GhostNet. These mutually reinforcing modules form a co-designed detection pipeline that significantly reduces both false negatives and false positives, providing reliable perception for disaster-area rescue operations. Experiments on VisDrone2019 (visible-light) and HIT-UAV (infrared) datasets show that EGSMnet achieves 41.6% mAP@50 and 24.8% mAP@50-95 on VisDrone2019 (improving over YOLOv8-n by 10.5 and 7.0 percentage points), and 95.1% mAP@50 and 61.5% mAP@50-95 on HIT-UAV, with only 2.5M parameters and 29.8 frames per second (FPS) on an NVIDIA Jetson Orin Nano, demonstrating a favorable accuracy–efficiency balance across visible-light and infrared modalities. Our code is available: <a href="https://github.com/AIQiQiaN/EGSMnet">https://github.com/AIQiQiaN/EGSMnet</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

Egsmnet: a lightweight framework for personnel and vehicle detection in UAV emergency rescue imagery

  • 期刊论文

  • 发表日期：

    2026年7月28日

查看全文
  • Peng Xiao,
  • Zhan Wen,
  • Meiqin Wu,
  • Wenzao Li

摘要

Unmanned aerial vehicle (UAV) imagery in emergency rescue scenarios faces severe challenges under complex conditions, including extreme scale variations, dense occlusions, and complex background interference (ruins, debris, vegetation, smoke plumes), which cause significant feature loss for small personnel and vehicle targets, thereby restricting real-time search and rescue decisions. This work presents efficient gated sparse multi-scale net (EGSMnet), a lightweight, parameter-efficient framework for personnel and vehicle detection in UAV emergency rescue imagery via adaptive feature enhancement. The spatially adaptive gated efficient attention (SAGE) mechanism generates spatially adaptive saliency masks to suppress disaster-area background noise while preserving the spatial integrity of small personnel and vehicle targets, unlike conventional attention mechanisms that apply uniform global pooling. The cross-scale multi-path adaptive aggregation (CMAA) network constructs ultra-fine-grained feature propagation paths with scale-adaptive asymmetric feature injection to avoid feature dilution during deep propagation. Enhanced ghost convolution (EGConv) achieves adaptive feature extraction with reduced parameters, overcoming the fixed feature-splitting ratio of standard GhostNet. These mutually reinforcing modules form a co-designed detection pipeline that significantly reduces both false negatives and false positives, providing reliable perception for disaster-area rescue operations. Experiments on VisDrone2019 (visible-light) and HIT-UAV (infrared) datasets show that EGSMnet achieves 41.6% mAP@50 and 24.8% mAP@50-95 on VisDrone2019 (improving over YOLOv8-n by 10.5 and 7.0 percentage points), and 95.1% mAP@50 and 61.5% mAP@50-95 on HIT-UAV, with only 2.5M parameters and 29.8 frames per second (FPS) on an NVIDIA Jetson Orin Nano, demonstrating a favorable accuracy–efficiency balance across visible-light and infrared modalities. Our code is available: https://github.com/AIQiQiaN/EGSMnet