Scholarlink

<p>Real-time small-object detection and tracking from UAVs is inherently challenging due to tiny object sizes, rapid viewpoint changes, and stringent accuracy–speed constraints, yet remains essential for mission-critical defense, security, and disaster response applications. We propose YOLOv11-EFAC, using YOLOv11n as baseline, a UAV-oriented detection framework employing a multi-level optimization strategy: an EfficientNet-B0 backbone for lightweight, high-quality feature extraction; hybrid FPN+PANet for enhanced multi-scale fusion; Squeeze-and-Excitation attention for adaptive channel weighting; and a custom loss function tailored for small-object localization. For tracking, we comparatively evaluate multiple state-of-the-art algorithms, including DeepSORT, EKF, ByteTrack, SORT, CenterTrack, FairMOT, and the transformer-based TrackFormer under realistic UAV operational conditions. Additionally, we introduce a hybrid DeepSORT+EKF approach to better handle non-linear motion, achieving 89.9% MOTA, an 11.5% improvement over standalone DeepSORT, with reduced identity switches. A 42,500-image hybrid dataset (58.1% small objects) combining COCO, VisDrone, and UAVDT improves robustness and generalization. Experimental results demonstrate 83.7% mAP@0.5 at 89 FPS on embedded hardware&#xa0;with a 21.4% small-object detection improvement, surpassing YOLOv8,&#xa0;YOLOv12, YOLOv13, and multiple YOLOv11 variants. These results position YOLOv11-EFAC as a robust, real-time solution for mission-critical UAV applications under operational constraints.</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

Real-time multi-object detection and tracking in UAV systems: improved YOLOv11-EFAC and optimized tracking algorithms

  • 期刊论文

  • 发表日期：

    2025年9月2日

查看全文
  • Rabia Kıratlı,
  • Alperen Eroğlu

摘要

Real-time small-object detection and tracking from UAVs is inherently challenging due to tiny object sizes, rapid viewpoint changes, and stringent accuracy–speed constraints, yet remains essential for mission-critical defense, security, and disaster response applications. We propose YOLOv11-EFAC, using YOLOv11n as baseline, a UAV-oriented detection framework employing a multi-level optimization strategy: an EfficientNet-B0 backbone for lightweight, high-quality feature extraction; hybrid FPN+PANet for enhanced multi-scale fusion; Squeeze-and-Excitation attention for adaptive channel weighting; and a custom loss function tailored for small-object localization. For tracking, we comparatively evaluate multiple state-of-the-art algorithms, including DeepSORT, EKF, ByteTrack, SORT, CenterTrack, FairMOT, and the transformer-based TrackFormer under realistic UAV operational conditions. Additionally, we introduce a hybrid DeepSORT+EKF approach to better handle non-linear motion, achieving 89.9% MOTA, an 11.5% improvement over standalone DeepSORT, with reduced identity switches. A 42,500-image hybrid dataset (58.1% small objects) combining COCO, VisDrone, and UAVDT improves robustness and generalization. Experimental results demonstrate 83.7% mAP@0.5 at 89 FPS on embedded hardware with a 21.4% small-object detection improvement, surpassing YOLOv8, YOLOv12, YOLOv13, and multiple YOLOv11 variants. These results position YOLOv11-EFAC as a robust, real-time solution for mission-critical UAV applications under operational constraints.