Scholarlink

<p>The complex traffic environment poses significant challenges to the visual analysis of autonomous vehicles. Existing general object detection algorithms often encounter drawbacks in traffic scenes, such as high model complexity, slow inference speed, and subpar performance on small objects. To address these issues, we propose a novel framework called YOLO-TS based on YOLOv5s specifically tailored for traffic scenes. Our approach includes several key components. Firstly, we introduce a Global Feature Extraction Block (GFEB) with varying dilation rates to capture more comprehensive spatial features of objects by enlarging the convolutional receptive field. This enables the model to have a better understanding of object contexts in the scene. Secondly, we propose the Mix Spatial Pyramid Pooling Fast (MSPPF) module, which effectively integrates features from different scales in the deep convolutional neural network. This allows the model to leverage multi-scale information for robust object detection. Moreover, we design an Enhanced Feature Pyramid Aggregation Network (EFPAN) to fuse multi-scale feature maps and reduce the semantic gap between different layers, thereby enhancing the detection performance of small targets. To optimize the bounding box loss function and enable fast convergence during training, we employ an efficient intersection over union (EIOU) loss. This loss function facilitates efficient training and improves the model’s accuracy. Overall, YOLO-TS combines efficient contextual information, achieving a better balance between accuracy and complexity in traffic scenes. Extensive experiments have been conducted to evaluate our method. The results demonstrate that YOLO-TS significantly outperforms the YOLOv5s baseline. When compared to existing mainstream object detection architectures, YOLO-TS achieves competitive performance.</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

Multi-scale object detection with feature enhancement for traffic scenes

  • 期刊论文

  • 发表日期：

    2026年1月13日

查看全文
  • Qinliang Chuai,
  • Tieying Zhu,
  • Xuezhen Wang,
  • Zhentao Wang,
  • Xuan Fu

摘要

The complex traffic environment poses significant challenges to the visual analysis of autonomous vehicles. Existing general object detection algorithms often encounter drawbacks in traffic scenes, such as high model complexity, slow inference speed, and subpar performance on small objects. To address these issues, we propose a novel framework called YOLO-TS based on YOLOv5s specifically tailored for traffic scenes. Our approach includes several key components. Firstly, we introduce a Global Feature Extraction Block (GFEB) with varying dilation rates to capture more comprehensive spatial features of objects by enlarging the convolutional receptive field. This enables the model to have a better understanding of object contexts in the scene. Secondly, we propose the Mix Spatial Pyramid Pooling Fast (MSPPF) module, which effectively integrates features from different scales in the deep convolutional neural network. This allows the model to leverage multi-scale information for robust object detection. Moreover, we design an Enhanced Feature Pyramid Aggregation Network (EFPAN) to fuse multi-scale feature maps and reduce the semantic gap between different layers, thereby enhancing the detection performance of small targets. To optimize the bounding box loss function and enable fast convergence during training, we employ an efficient intersection over union (EIOU) loss. This loss function facilitates efficient training and improves the model’s accuracy. Overall, YOLO-TS combines efficient contextual information, achieving a better balance between accuracy and complexity in traffic scenes. Extensive experiments have been conducted to evaluate our method. The results demonstrate that YOLO-TS significantly outperforms the YOLOv5s baseline. When compared to existing mainstream object detection architectures, YOLO-TS achieves competitive performance.