Scholarlink

<p>In the Formula Student Autonomous China (FSAC) competition, accurate detection of traffic cones is the key to ensuring that autonomous racing cars can accomplish their autonomous driving mission. In the field of autonomous racing cars, a more excellent detection range enables the racing cars to obtain more forward-looking environmental data, thereby facilitating the planning of an optimal trajectory to improve the single-lap speed. In addition, the limited computing power of the racing cars requires that the network parameters be kept as low as possible. In order to solve the above problems, this research proposes the TCone-YOLO network based on YOLOv5, which introduces the SPD-Conv module in the backbone network and proposes the MF-BiFPN(Micro Focus BiFPN), based on BiFPN. The SPD-Conv and MF-BiFPN modules enhance the detection capability for long-distance traffic cones, thereby effectively expanding the network’s detection range. Among them, MF-BiFPN can reduce the number of network parameters. Furthermore, this research proposes a method to select the number of anchors based on the silhouette coefficient to obtain a more optimal number of anchors according to the dataset, considering the sizes of traffic cones in this competition are fixed. Moreover, an open-source traffic cone dataset for the FSAC competition, Formula Student Autonomous China Cone in Context (FSACCCO), is created. Experiments on FSACCCO show TCone-YOLO improves mAP50 by 3.83% and mAP50:95 by 1.74% over YOLOv5s, while reducing parameters by 7.97% and extending detection range by 20.59%.</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

TCone-YOLO: enhanced long-range and low-parameter traffic cone detection for autonomous racing cars

  • 期刊论文

  • 发表日期：

    2025年9月5日

查看全文
  • Wei Huang,
  • Zhu Liao,
  • Zhichen Wei,
  • Yihua Lu,
  • Hai Li

摘要

In the Formula Student Autonomous China (FSAC) competition, accurate detection of traffic cones is the key to ensuring that autonomous racing cars can accomplish their autonomous driving mission. In the field of autonomous racing cars, a more excellent detection range enables the racing cars to obtain more forward-looking environmental data, thereby facilitating the planning of an optimal trajectory to improve the single-lap speed. In addition, the limited computing power of the racing cars requires that the network parameters be kept as low as possible. In order to solve the above problems, this research proposes the TCone-YOLO network based on YOLOv5, which introduces the SPD-Conv module in the backbone network and proposes the MF-BiFPN(Micro Focus BiFPN), based on BiFPN. The SPD-Conv and MF-BiFPN modules enhance the detection capability for long-distance traffic cones, thereby effectively expanding the network’s detection range. Among them, MF-BiFPN can reduce the number of network parameters. Furthermore, this research proposes a method to select the number of anchors based on the silhouette coefficient to obtain a more optimal number of anchors according to the dataset, considering the sizes of traffic cones in this competition are fixed. Moreover, an open-source traffic cone dataset for the FSAC competition, Formula Student Autonomous China Cone in Context (FSACCCO), is created. Experiments on FSACCCO show TCone-YOLO improves mAP50 by 3.83% and mAP50:95 by 1.74% over YOLOv5s, while reducing parameters by 7.97% and extending detection range by 20.59%.