Scholarlink

<p>The rapid pace of urbanization and the increase in vehicle ownership have significantly raised the challenges associated with vehicle detection in complex traffic environments, which are further complicated by occlusions, multi-scale target distributions, and adverse weather conditions. Existing YOLO-series algorithms still present the potential for optimization in achieving a balance between detection accuracy and real-time performance, especially in contexts involving dense traffic and small-target detection. To overcome these challenges, an enhanced MD-YOLO (Multi-scale and Dynamic YOLO) model is proposed in this study. The model enhances cross-layer feature representation via a Multi-Modal Adaptive Feature Aggregation Network (MAFAN), incorporates a Dynamic Multi-Scale Adaptive Spatial Attention Gate (DMSAG) to mitigate background noise and highlight critical regions, and introduces an Inner-Wise-Powerful IoU (IWP-IoU) loss function to improve bounding box regression. Experiments show that on the KITTI vehicle detection task, MD-YOLO achieves an accuracy of 91.1%, a recall of 81.4%, and an mAP50 of 89.3%, outperforming other popular algorithms of the same type across all these metrics. Cross-domain evaluation on the self-constructed CARS dataset further confirms the model’s robust adaptability to real-world traffic scenarios. Notably, MD-YOLO is specifically designed to address the high-performance computing (HPC) challenges associated with large-scale, city-level real-time video stream analysis in intelligent transportation systems. The model’s multimodal and dynamic architecture enhances detection accuracy while offering considerable parallelization capability, permitting efficient deployment on both HPC cluster-based cloud control platforms and high-performance edge nodes. Beyond presenting an advanced detection algorithm, this work provides a highly efficient supercomputing-oriented solution for traffic perception tasks. Furthermore, its modular design establishes a technical foundation for subsequent optimization on heterogeneous computing architectures.</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

MD-YOLO: research on the innovation and performance optimization of vehicle detection technology

  • 期刊论文

  • 发表日期：

    2026年1月20日

查看全文
  • Keyou Guo,
  • Jiangnan Wang,
  • Haibing Jiang,
  • Pei Zhang,
  • Huangcheng Qin

摘要

The rapid pace of urbanization and the increase in vehicle ownership have significantly raised the challenges associated with vehicle detection in complex traffic environments, which are further complicated by occlusions, multi-scale target distributions, and adverse weather conditions. Existing YOLO-series algorithms still present the potential for optimization in achieving a balance between detection accuracy and real-time performance, especially in contexts involving dense traffic and small-target detection. To overcome these challenges, an enhanced MD-YOLO (Multi-scale and Dynamic YOLO) model is proposed in this study. The model enhances cross-layer feature representation via a Multi-Modal Adaptive Feature Aggregation Network (MAFAN), incorporates a Dynamic Multi-Scale Adaptive Spatial Attention Gate (DMSAG) to mitigate background noise and highlight critical regions, and introduces an Inner-Wise-Powerful IoU (IWP-IoU) loss function to improve bounding box regression. Experiments show that on the KITTI vehicle detection task, MD-YOLO achieves an accuracy of 91.1%, a recall of 81.4%, and an mAP50 of 89.3%, outperforming other popular algorithms of the same type across all these metrics. Cross-domain evaluation on the self-constructed CARS dataset further confirms the model’s robust adaptability to real-world traffic scenarios. Notably, MD-YOLO is specifically designed to address the high-performance computing (HPC) challenges associated with large-scale, city-level real-time video stream analysis in intelligent transportation systems. The model’s multimodal and dynamic architecture enhances detection accuracy while offering considerable parallelization capability, permitting efficient deployment on both HPC cluster-based cloud control platforms and high-performance edge nodes. Beyond presenting an advanced detection algorithm, this work provides a highly efficient supercomputing-oriented solution for traffic perception tasks. Furthermore, its modular design establishes a technical foundation for subsequent optimization on heterogeneous computing architectures.