摘要
公路中心标线的实时跟踪是公路巡检无人机视觉飞行中关键的一环。针对目前主流目标跟踪算法实时性差的问题,提出一种基于改进YOLO(you only look once)v3和Deep-SORT(deep simple online real-time tracking)的目标跟踪模型用于公路巡检无人机自主视觉飞行。通过引入并改进跨阶段局部网络,优化网络层级结构,使用泛化能力更好的激活函数,提升了公路道路标线的检测准确率和无人机平台的检测速度。对检测到的公路标线信息使用Deep-SORT算法进行公路中心标线跟踪。实验结果表明,与几类典型目标跟踪模型相比,在跟踪准确度基本不变的情况下,处理速度提升了数倍。
Real-time tracking of highway center markings is a key part of the visual flight of highway inspection drones.Aiming at the problem of leaky detection accuracy and speed of traditional object tracking method.This paper proposes an optimized algorithm based on YOLO(you only look once)and Deep-SORT(simple online real-time tracking)models to improve the accuracy and the speed of highway center marking tracking.This model proposes a new backbone network by introducing CSPnet(cross stage partial network)which has replaced the original network.To further heighten the performance of proposed model,this study also replaces the original activation function in the network to improve the ability of proposed model.In this study,Deep-SORT method is used to track the detected objects.Compared with several types of typical target tracking models,the processing speed is increased several times under the condition that the tracking accuracy is basically unchanged.
作者
韩建峰
赵志伟
宋丽丽
HAN Jianfeng;ZHAO Zhiwei;SONG Lili(School of Information Engineering,Inner Mongolia University of Technology,Hohhot,Inner Mongolia 010080,China;Inner Mongolia Key Laboratory of Perceptive Technology and Intelligent Systems,Inner Mongolia University of Technology,Hohhot,Inner Mongolia 010051,China;School of Aviation,Inner Mongolia University of Technology,Hohhot,Inner Mongolia 010080,China)
出处
《光电子．激光》
CAS
CSCD
北大核心
2021年第9期927-934,共8页
Journal of Optoelectronics·Laser
基金
内蒙古自治区关键技术攻关计划项目(基于无人机的公路路面健康状况监测,2019GG271)
内蒙古自治区高等学校科学研究重点项目(无人机航拍图像快速拼接的关键技术研究,NJZZ19068)资助项目。
关键词
目标跟踪
目标检测
无人机
深度学习
跨阶段局部网络
object tracking
object detection
unmanned aerial vehicle
deep learning
cross-stage partial network
