摘要
随着深度学习与人工智能技术的不断发展,视频目标跟踪已经成为计算机视觉的重要研究内容,在公安布控、人机交互、交通管制、军事等各个领域起到越来越重要的作用;尽管现在国内外学者提出了多种目标跟踪算法,也搭建了较为完善的目标跟踪系统,但是算法的鲁棒性依然是一个比较大的挑战;文章对运动目标跟踪系统结构进行了简要介绍,并从特征提取及融合、外观模型、目标搜索等方面详细阐述了目前主流运动目标跟踪算法;然后对目标跟踪算法在深度学习大环境下的新发展进行了分析,从基于深度学习的目标跟踪及目标检测算法角度分析了深度学习在提高目标检测算法鲁棒性方面的有效性,最后概述了深度学习在视频目标检测算法中的具体应用并对其未来发展进行了展望。
With the development of deep learning and artificial intelligence technology,video object tracking has become an important research content of computer vision.Video object tracking plays a more and more important role in public security,humancomputer interaction,traffic control,military and other fields.Although a variety of object tracking algorithms have been proposed by scholars,and a relatively perfect object tracking system has also been built,the robustness of the algorithm is still a big challenge.In this paper,the structure of moving object tracking system is briefly introduced.At the same time,the main moving object tracking algorithms are described in detail from feature extraction and fusion,appearance model,object search and so on.Then,a new development of object tracking algorithm in deep learning environment is analyzed.From the perspective of object tracking and object detection algorithm based on deep learning,the effectiveness of deep learning in improving the robustness of object detection algorithm is analyzed.Finally,the specific application of video object detection algorithm is summarized and its future development is prospected.
作者
王海涛
王荣耀
王文皞
王海龙
刘强
Wang Haitao;Wang Rongyao;Wang Wenhao;Wang Hailong;Liu Qiang(School of Automation Engineering,Nanjing University of Aeronautics and Aeronautics,Nanjing 211106,China;Shandong Great Wall Computer System Co.,Ltd.,Yantai 264003,China;Jiangsu Mingyuan Rail Transit Co.,Ltd.,Nanjing 210044,China)
出处
《计算机测量与控制》
2020年第4期1-6,21,共7页
Computer Measurement &Control
关键词
目标跟踪
特征提取
外观模型
深度学习
神经网络
object tracking
feature extraction
appearance model
deep learning
neural network