Scholarlink

<p>Real-time vehicle detection forms the foundation of intelligent traffic management. Traditional detection methods suffer from limitations in timeliness and adaptability to complex environments. This study proposes a real-time vehicle detection model based on an improved lightweight detection model to achieve a balance between accuracy and real-time performance in edge computing scenarios for real-time vehicle detection models and to further enhance the robustness of detection methods. Specifically, replacing the Stem module reduces the computational load by 23.6%, and Lightweight Head compression achieves a 48.7% reduction in parameters, ensuring high hardware compatibility. The research method optimizes lightweight detection models by integrating global–local information networks, replacing the Focus module with the Stem module. The New Backbone then extracts multi-scale features, which undergo multi-level feature fusion and attention-weighted aggregation via the simple attention module-path aggregation network. The Lightweight Head module ultimately outputs the detection results. The results showed that compared with existing lightweight models (YOLO-NAS and RT-DETR), the research method performed better on key indicators. Notably, under illumination variations (0–10,000 lx), mAP degradation was minimized to 7.4%, while occlusion tests showed 89.9% recall at 100% occlusion ratio. Among them, the mAP value of the research method only dropped by 7.4% in the light intensity range of 0-10000lx, while the two comparison models dropped more significantly. Its FPS value on edge devices (such as Jetson Nano) could reach 29.8 frames/s, exceeding the real-time threshold (24 FPS) for video processing, with only 28.9ms inference delay at 640 × 640 resolution. Meanwhile, the recall rate remained at 89.9% with an occlusion ratio of 100%, significantly improving the robustness of the real-time vehicle detection model in complex environments. In summary, the proposed model is highly precise, computationally efficient, stable, and robust. It effectively addresses the limitations of traditional methods regarding timeliness and environmental adaptability.</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 vehicle detection methods based on an improved YOLO-Lite approach in edge computing scenarios

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年2月13日

查看全文
  • Miao He,
  • Changhao Li,
  • Jinfeng Yang,
  • Xiaomin Ning

摘要

Real-time vehicle detection forms the foundation of intelligent traffic management. Traditional detection methods suffer from limitations in timeliness and adaptability to complex environments. This study proposes a real-time vehicle detection model based on an improved lightweight detection model to achieve a balance between accuracy and real-time performance in edge computing scenarios for real-time vehicle detection models and to further enhance the robustness of detection methods. Specifically, replacing the Stem module reduces the computational load by 23.6%, and Lightweight Head compression achieves a 48.7% reduction in parameters, ensuring high hardware compatibility. The research method optimizes lightweight detection models by integrating global–local information networks, replacing the Focus module with the Stem module. The New Backbone then extracts multi-scale features, which undergo multi-level feature fusion and attention-weighted aggregation via the simple attention module-path aggregation network. The Lightweight Head module ultimately outputs the detection results. The results showed that compared with existing lightweight models (YOLO-NAS and RT-DETR), the research method performed better on key indicators. Notably, under illumination variations (0–10,000 lx), mAP degradation was minimized to 7.4%, while occlusion tests showed 89.9% recall at 100% occlusion ratio. Among them, the mAP value of the research method only dropped by 7.4% in the light intensity range of 0-10000lx, while the two comparison models dropped more significantly. Its FPS value on edge devices (such as Jetson Nano) could reach 29.8 frames/s, exceeding the real-time threshold (24 FPS) for video processing, with only 28.9ms inference delay at 640 × 640 resolution. Meanwhile, the recall rate remained at 89.9% with an occlusion ratio of 100%, significantly improving the robustness of the real-time vehicle detection model in complex environments. In summary, the proposed model is highly precise, computationally efficient, stable, and robust. It effectively addresses the limitations of traditional methods regarding timeliness and environmental adaptability.