Scholarlink

<p>In-vehicle multi-object detection faces challenges due to the dynamic traffic environment, small object sizes, and complex backgrounds. To address these issues, we propose RSW-YOLOv8, a model designed for enhanced detection in such scenarios. The model replaces standard convolutions in the YOLOv8 backbone with RFCAConv for better spatial feature extraction, particularly for small objects. It also adds a four-level prediction head to improve shallow-level feature capture, incorporates SimAM for better attention on key regions, and uses Wise-IoU as the loss function to increase bounding box reliability. Testing on a custom traffic dataset shows RSW-YOLOv8 outperforms the baseline and other detection models, including Faster RCNN and YOLO series. The proposed model achieves a 0.4% increase in precision, a 6.6% improvement in recall, a 5.2% boost in mAP@0.5, and 35.3% faster inference, striking a balance between accuracy and real-time 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

Rsw-YOLOv8: an improved model for in-vehicle vision multi-object detection in complex traffic scenes

  • 期刊论文

  • 发表日期：

    2025年6月17日

查看全文
  • PeiQun Lin,
  • QianTong Li,
  • JingYi Mao

摘要

In-vehicle multi-object detection faces challenges due to the dynamic traffic environment, small object sizes, and complex backgrounds. To address these issues, we propose RSW-YOLOv8, a model designed for enhanced detection in such scenarios. The model replaces standard convolutions in the YOLOv8 backbone with RFCAConv for better spatial feature extraction, particularly for small objects. It also adds a four-level prediction head to improve shallow-level feature capture, incorporates SimAM for better attention on key regions, and uses Wise-IoU as the loss function to increase bounding box reliability. Testing on a custom traffic dataset shows RSW-YOLOv8 outperforms the baseline and other detection models, including Faster RCNN and YOLO series. The proposed model achieves a 0.4% increase in precision, a 6.6% improvement in recall, a 5.2% boost in mAP@0.5, and 35.3% faster inference, striking a balance between accuracy and real-time performance.