Scholarlink

<p>For autonomous driving in complex urban street scenes, pedestrians and vehicles are the key detection objects. An improved algorithm based on YOLOv4 is proposed to address the issues of missed detection and low detection accuracy for small, occluded and overlapping objects. Firstly, to ensure that the model is available for embedded mobile devices, the ShuffleNetV2 network is introduced for the lightweight YOLOv4. Secondly, improving the attention mechanism of the convolutional block attention module (CBAM) and introducing it into the object detection algorithm ensures that the model pays attention to detection objects of various sizes. Finally, introducing SIoU loss and Soft NMS effectively improves the detection performance of overlapping objects. The results show that the number of frames reaches 43 per second, and the detection average precision (AP)@0.5 is 83.2%. Compared with YOLOv4, the difference in detection accuracy is not significant, but the model size is reduced to 1/6, only 43.4M, and the detection speed is improved dramatically. In addition, the comparison of the four versions of YOLOv5 also highlights the advantages of the improved model.</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

Research on an Improved YOLOv4-based Method for Pedestrian and Vehicle Detection in Urban Street Scenes

  • 期刊论文

  • 发表日期：

    2025年10月16日

查看全文
  • Qiming Wang,
  • Tianqi Yang,
  • Zhichao Lyu,
  • Xuan Wan

摘要

For autonomous driving in complex urban street scenes, pedestrians and vehicles are the key detection objects. An improved algorithm based on YOLOv4 is proposed to address the issues of missed detection and low detection accuracy for small, occluded and overlapping objects. Firstly, to ensure that the model is available for embedded mobile devices, the ShuffleNetV2 network is introduced for the lightweight YOLOv4. Secondly, improving the attention mechanism of the convolutional block attention module (CBAM) and introducing it into the object detection algorithm ensures that the model pays attention to detection objects of various sizes. Finally, introducing SIoU loss and Soft NMS effectively improves the detection performance of overlapping objects. The results show that the number of frames reaches 43 per second, and the detection average precision (AP)@0.5 is 83.2%. Compared with YOLOv4, the difference in detection accuracy is not significant, but the model size is reduced to 1/6, only 43.4M, and the detection speed is improved dramatically. In addition, the comparison of the four versions of YOLOv5 also highlights the advantages of the improved model.