Scholarlink

<p>Achieving efficient object detection on unmanned aerial vehicle (UAV) platforms is challenging due to the small scale of targets, complex backgrounds, and limited on-board resources. This paper proposes a novel re-parameterized object detection network named RPDet for UAV edge platforms. A key component of RPDet is the re-parameterizable feature extraction block, Rep-DCSA. During training, Rep-DCSA excels in feature extraction and aggregation. In inference, it simplifies the model structure through re-parameterization, significantly reducing model parameters and computational complexity. To advance feature extraction and interaction, we introduce the inverted bottleneck feature enhancement and aggregation (IFEA) module to enhance important channel features and the feature fusion module (FFM) to integrate multi-level information. In addition, we incorporate the global context (GC) module in the detection head, which leverages contextual information to effectively improve the detection performance of small objects in complex backgrounds. Extensive experiments on the public VisDrone and self-constructed HC-UAV datasets demonstrate that RPDet achieves a better trade-off between accuracy and speed than state-of-the-art methods. When deployed on edge platforms, RPDet’s accuracy on Orin Nano and RK3588-RT platforms improved by 3.4% and 5.8% compared to EdgeYOLO. Moreover, RPDet exhibited superior real-time performance and energy efficiency.</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

RPDet: a re-parameterized efficient object detection network for UAV edge platforms

  • 期刊论文

  • 发表日期：

    2025年6月5日

查看全文
  • Buhong Zhang,
  • Zhigang Wang,
  • Meibo Lv,
  • Xiaodong Liu,
  • Lei Zhang

摘要

Achieving efficient object detection on unmanned aerial vehicle (UAV) platforms is challenging due to the small scale of targets, complex backgrounds, and limited on-board resources. This paper proposes a novel re-parameterized object detection network named RPDet for UAV edge platforms. A key component of RPDet is the re-parameterizable feature extraction block, Rep-DCSA. During training, Rep-DCSA excels in feature extraction and aggregation. In inference, it simplifies the model structure through re-parameterization, significantly reducing model parameters and computational complexity. To advance feature extraction and interaction, we introduce the inverted bottleneck feature enhancement and aggregation (IFEA) module to enhance important channel features and the feature fusion module (FFM) to integrate multi-level information. In addition, we incorporate the global context (GC) module in the detection head, which leverages contextual information to effectively improve the detection performance of small objects in complex backgrounds. Extensive experiments on the public VisDrone and self-constructed HC-UAV datasets demonstrate that RPDet achieves a better trade-off between accuracy and speed than state-of-the-art methods. When deployed on edge platforms, RPDet’s accuracy on Orin Nano and RK3588-RT platforms improved by 3.4% and 5.8% compared to EdgeYOLO. Moreover, RPDet exhibited superior real-time performance and energy efficiency.