Scholarlink

To address the challenge of vehicle detection in drone perspectives for dim target detection, this paper proposes a YOLOv8s-based detection framework. By reconstructing the annotation system of the VisDrone dataset—retaining vehicle-related categories and removing interfering targets such as pedestrians—the model’s capability to learn vehicle-specific features is optimized. Experiments employ a pretrained YOLOv8s model trained on the filtered VisDrone dataset for 276 epochs, achieving precise vehicle detection with an mAP50 of 80.3% and an inference speed of 8.4 ms per frame. The study demonstrates that target category filtering effectively enhances the model’s focus on specific tasks, providing a lightweight solution for drone-based vehicle detection in dim target scenarios.

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

Dim Target Detection Using Deep Learning Network

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Ying Wang,
  • Haoting Liu,
  • Hao Li,
  • Kai Ding,
  • Haiguang Li,
  • Xiaofei Lu,
  • Qing Li

摘要

To address the challenge of vehicle detection in drone perspectives for dim target detection, this paper proposes a YOLOv8s-based detection framework. By reconstructing the annotation system of the VisDrone dataset—retaining vehicle-related categories and removing interfering targets such as pedestrians—the model’s capability to learn vehicle-specific features is optimized. Experiments employ a pretrained YOLOv8s model trained on the filtered VisDrone dataset for 276 epochs, achieving precise vehicle detection with an mAP50 of 80.3% and an inference speed of 8.4 ms per frame. The study demonstrates that target category filtering effectively enhances the model’s focus on specific tasks, providing a lightweight solution for drone-based vehicle detection in dim target scenarios.