Scholarlink

<p>The perception system is the core component of self-driving cars, and a reliable perception system is a key condition for the secure and reliable operation of autonomous vehicles in various complex situations. To improve object detection performance in complex scenes, this study proposes a lightweight and efficient detection framework by integrating three key components. First, Receptive Field Attention Convolution (RFAConv) is introduced to optimize spatial feature representation within the receptive field, enhancing the model’s ability to focus on target regions and improving detection accuracy for small and cluttered objects. Second, the Top-K Sparse Kernel Attention (TKSA) mechanism selectively attends to the most relevant feature areas in the backbone, improving long-range perception while maintaining computational efficiency. Finally, Focaler-IoU is adopted as the bounding box regression loss to dynamically adjust the contribution of samples based on their IoU, improving stability and accelerating training convergence. Together, these enhancements significantly boost detection accuracy and robustness, particularly in challenging environments. Assessments performed on the KITTI datasets show that our detection algorithms achieve mAP@0.5 and mAP@0.5:0.95 of 83% and 55.4%, respectively. In comparison with the baseline model YOLOv11n, our approach delivers notable performance gains, with improvements of 5.2% and 3.9%.</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

RTF-YOLO: an efficient object detection framework for autonomous vehicles

  • 期刊论文

  • 发表日期：

    2025年9月5日

查看全文
  • Yang Sun,
  • Mengyao Zhen,
  • Yinuo Ma,
  • Haiyang Wang,
  • Guanyu Chen,
  • Xiaoran Gao

摘要

The perception system is the core component of self-driving cars, and a reliable perception system is a key condition for the secure and reliable operation of autonomous vehicles in various complex situations. To improve object detection performance in complex scenes, this study proposes a lightweight and efficient detection framework by integrating three key components. First, Receptive Field Attention Convolution (RFAConv) is introduced to optimize spatial feature representation within the receptive field, enhancing the model’s ability to focus on target regions and improving detection accuracy for small and cluttered objects. Second, the Top-K Sparse Kernel Attention (TKSA) mechanism selectively attends to the most relevant feature areas in the backbone, improving long-range perception while maintaining computational efficiency. Finally, Focaler-IoU is adopted as the bounding box regression loss to dynamically adjust the contribution of samples based on their IoU, improving stability and accelerating training convergence. Together, these enhancements significantly boost detection accuracy and robustness, particularly in challenging environments. Assessments performed on the KITTI datasets show that our detection algorithms achieve mAP@0.5 and mAP@0.5:0.95 of 83% and 55.4%, respectively. In comparison with the baseline model YOLOv11n, our approach delivers notable performance gains, with improvements of 5.2% and 3.9%.