Scholarlink

<p>The real world contains diverse objects, typically categorized as large or small based on appearance. However, in dynamic real-world tasks, object size is not solely determined by its intrinsic properties. For instance, from the perspective of unmanned aerial vehicles (UAVs), most objects appear as difficult samples at high altitudes, whereas at lower altitudes, most objects can be treated as easy samples. Additionally, in edge device scenarios, adjusting the input image size to maximize the detector performance based on the available computational resources is a satisfactory strategy, but we found that such strategy needs the model to be able to dynamically perceive drastic changes in the object scale. To address the inherent challenges in this field and improve the model’s dynamic perception capability, we proposed a multi-scale dynamic perception object detection method MSDP-YOLO. First, we designed a Lightweight Unified Dynamic Detection Head (LUDH) to enhance the model’s adaptability to scenarios with varying scales. Secondly, we developed two cross-level and cross-scale feature aggregation modules: the Efficient Multiple Scale Feature Fusion (EMSFF) module and the Two-layer Path Parallel Aggregation (TPPA) module, which enable bidirectional augmentation of both semantic and detailed information, and effectively alleviating issues with occluded dense small objects and limited critical information. Finally, a lightweight module, C2f with Dynamic Ghost Module (C2fDGM), was designed to integrate difference information across modules, productively extracting essential feature information from discrepancies and further strengthening the input of the detection layers. Extensive experiments demonstrate that, compared to mainstream excellent lightweight methods, our approach achieves superior performance on the VisDrone and PASCAL VOC datasets, which have drastically dissimilar on object scales. This indicates that our method is better suited to the demands of detection tasks in dynamic real-world environments.</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

MSDP-YOLO: Multi-scale Dynamic Perception Object Detection Method for UAVs

  • 期刊论文

  • 发表日期：

    2026年4月21日

查看全文
  • Tien-Wen Sung,
  • Ze-Hao Chen,
  • Chao-Yang Lee

摘要

The real world contains diverse objects, typically categorized as large or small based on appearance. However, in dynamic real-world tasks, object size is not solely determined by its intrinsic properties. For instance, from the perspective of unmanned aerial vehicles (UAVs), most objects appear as difficult samples at high altitudes, whereas at lower altitudes, most objects can be treated as easy samples. Additionally, in edge device scenarios, adjusting the input image size to maximize the detector performance based on the available computational resources is a satisfactory strategy, but we found that such strategy needs the model to be able to dynamically perceive drastic changes in the object scale. To address the inherent challenges in this field and improve the model’s dynamic perception capability, we proposed a multi-scale dynamic perception object detection method MSDP-YOLO. First, we designed a Lightweight Unified Dynamic Detection Head (LUDH) to enhance the model’s adaptability to scenarios with varying scales. Secondly, we developed two cross-level and cross-scale feature aggregation modules: the Efficient Multiple Scale Feature Fusion (EMSFF) module and the Two-layer Path Parallel Aggregation (TPPA) module, which enable bidirectional augmentation of both semantic and detailed information, and effectively alleviating issues with occluded dense small objects and limited critical information. Finally, a lightweight module, C2f with Dynamic Ghost Module (C2fDGM), was designed to integrate difference information across modules, productively extracting essential feature information from discrepancies and further strengthening the input of the detection layers. Extensive experiments demonstrate that, compared to mainstream excellent lightweight methods, our approach achieves superior performance on the VisDrone and PASCAL VOC datasets, which have drastically dissimilar on object scales. This indicates that our method is better suited to the demands of detection tasks in dynamic real-world environments.