Scholarlink

<p>Vehicle detection is a critical component in intelligent transportation systems (ITS), where its accuracy and real-time performance hold significant importance for the traffic management. However, complex factors in traffic environments, such as small objects, occlusion interference, and varying weather, pose significant challenges. This paper proposes the Reparameterized Multi-Scale Feature Fusion Network (RSF-Net), an improved model that leverages GPU-parallel computing capabilities, which are essential for high-performance computing (HPC) applications in real-time vehicle detection. The model incorporates several key innovations: a Reparameterized Multi-Scale Feature Fusion (RepHMS) module based on reparameterization principles to enhance feature extraction, a Boundary-Aware Feature Pyramid Network (BA-FPN) that reduces boundary information loss through bidirectional cross-scale connections and channel attention mechanisms, a Parallel Spatial Pyramid Pooling Fast (PSPPF) module to expand the receptive field, and a dedicated small object detection head. On the UA-DETRAC dataset, RSF-Net improves mean Average Precision (mAP@0.5) by 6.4% over YOLOv8s and reaches 186.8 FPS. On the BDD100K dataset, it improves mAP@0.5 by 3.9%. These results indicate the model’s effectiveness for vehicle detection in complex urban traffic scenarios.</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

RSF-Net: a robust multi-scale feature fusion network for vehicle detection in challenging traffic environments

  • 期刊论文

  • 发表日期：

    2025年10月18日

查看全文
  • Lisi Dai,
  • Hao Tang,
  • Yonghui Zhang,
  • Bo Xu,
  • Uzair Aslam Bhatti,
  • Jinxiong Gao

摘要

Vehicle detection is a critical component in intelligent transportation systems (ITS), where its accuracy and real-time performance hold significant importance for the traffic management. However, complex factors in traffic environments, such as small objects, occlusion interference, and varying weather, pose significant challenges. This paper proposes the Reparameterized Multi-Scale Feature Fusion Network (RSF-Net), an improved model that leverages GPU-parallel computing capabilities, which are essential for high-performance computing (HPC) applications in real-time vehicle detection. The model incorporates several key innovations: a Reparameterized Multi-Scale Feature Fusion (RepHMS) module based on reparameterization principles to enhance feature extraction, a Boundary-Aware Feature Pyramid Network (BA-FPN) that reduces boundary information loss through bidirectional cross-scale connections and channel attention mechanisms, a Parallel Spatial Pyramid Pooling Fast (PSPPF) module to expand the receptive field, and a dedicated small object detection head. On the UA-DETRAC dataset, RSF-Net improves mean Average Precision (mAP@0.5) by 6.4% over YOLOv8s and reaches 186.8 FPS. On the BDD100K dataset, it improves mAP@0.5 by 3.9%. These results indicate the model’s effectiveness for vehicle detection in complex urban traffic scenarios.