赵梓杉,桑海峰.基于改进的 YOLOv5 的交通锥标检测系统[J].电子测量与仪器学报,2023,37(2):56-64 |
基于改进的 YOLOv5 的交通锥标检测系统 |
Traffic cone detection system based on improved YOLOv5 |
|
DOI: |
中文关键词: 无人驾驶方程式赛车 交通锥标 深度学习 目标检测 目标跟踪 相似三角形测距 |
英文关键词:driverless formula racing traffic cone deep learning target detection target tracking similar triangle ranging |
基金项目:辽宁省教育厅科研项目(LJGD2020006)资助 |
|
|
摘要点击次数: 863 |
全文下载次数: 812 |
中文摘要: |
针对中国大学生无人驾驶方程式赛车目标检测系统当前算法检测速度较慢,在不同场景下容易出现检测精度低,漏检、
误检现象严重等问题,文章设计了一种适用于锥标颜色的整体检测系统,在识别模块中,首先,为了提高原 YOLOv5 基础模型检
测速度和识别精度,采用 CIoU 作为边界框回归损失函数,针对训练时收敛速度慢和算法识别精确度低的问题,将原加权非极大
抑制方式更改为 DIoU_NMS,测试精度为 0. 963,相较于原算法提高了 2. 1%,结果表明改进后的算法更适合比赛场景下锥标颜
色识别。 其次,在跟踪模块中,对深度表观特征锥标颜色重识别模型进行训练,将单目标跟踪算法改为可以对多种类别目标进
行跟踪,相比于单一的目标检测算法,有效降低漏检现象,最后,添加测距模块,利用检测框高度信息进行车载摄像头到锥桶的
测距,90 m 以内的平均误差小于 9%。 整个系统的帧率达到 20 FPS,实现锥标颜色识别和距离的有效测量,为比赛提供更多的
数据支持。 |
英文摘要: |
In view of the slow detection speed of the current algorithm of the target detection system for Chinese college students’
driverless formula racing cars, the low detection accuracy and serious missing and false detection in different scenarios are easy to occur.
In the recognition module, first of all, in order to improve the detection speed and recognition accuracy of the original YOLOv5 basic
model, the paper uses CIoU as the boundary box regression loss function. To solve the problems of slow convergence speed and low
recognition accuracy of the algorithm during training, the original weighted nonmax suppression method is changed to DIoU_NMS in this
paper, the test accuracy is 0. 963, which is 2. 1% higher than the original algorithm. The results show that the improved algorithm is
more suitable for cone color recognition in the competition scene. Secondly, in the tracking module, the depth apparent feature cone
color recognition model is trained, and the single target tracking algorithm is changed to be able to track multiple types of targets.
Compared with a single target detection algorithm, the phenomenon of missing detection is effectively reduced. Finally, the ranging
module is added to use the height information of the detection frame to distance the vehicle camera to the cone barrel. The average error
within 90 meters is less than 9%. The frame rate of the whole system reaches 20 frames/ second, realizing cone color recognition and
effective distance measurement, and providing more data support for the game. |
查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|