摘要
视频跟踪是计算机视觉的重要研究方向,具有广泛的应用。但视频跟踪过程中存在运动模糊和形变等问题,导致了跟踪的性能降低。提出多特征卷积融合的相关滤波视频跟踪算法,通过卷积神经网络提取视频图像中多个特征,并根据提取特征后产生的多个响应值进行融合,从而实现目标定位。首先,通过卷积神经网络提取视频序列同一图像块由浅层到深层的多个特征;其次,将提取的图像特征通过目标响应置信函数产生目标响应值并进行融合,根据最小惩罚值来计算目标区域;最后,采用线性插值的方法进行目标模板和滤波参数更新,实现目标跟踪。多特征卷积融合的相关滤波算法在OTB-2013、OTB-2015和VOT-2016公开的数据集进行实验,与5个视频跟踪算法进行对比分析。实验表明:该算法表现出较好的性能,特别是在目标跟踪的运动模糊和形变方面的性能优于对比的相关滤波跟踪算法。
Video tracking is a vital part of computer vision and has a wide range of applications.However,there are motion blur and deformation problems in the video tracking process,which lead to the degradation of tracking performance.This paper proposes a correlation filtering video tracking algorithm of multi-feature convolution fusion,which extracts multiple features from video images through the convolutional neural network and fuses them according to multiple response values generated after extracting features to achieve target positioning.Firstly,multiple features of a video sequence from the shallow layer to deep layer are extracted by the convolutional neural network.Secondly,the extracted image features are fused to generate the target response value through the target response confidence function.The target region is calculated according to the minimum penalty value.Finally,the target template and filter parameters are updated by linear interpolation.The correlation filtering algorithm of multi-feature convolution fusion was tested in the data sets of OTB-2013,OTB-2015 and VOT-2016.Compared with five video tracking algorithms,the experimental results show that the proposed algorithm has better performance,especially in the motion blur and deformation problem of target tracking than the traditional correlation filter tracking algorithm.
出处
《漳州职业技术学院学报》
2020年第2期67-74,共8页
Journal of Zhangzhou Institute of Technology
关键词
计算机视觉
目标跟踪
卷积神经网络
相关滤波
融合响应
computer vision
object tracking
convolutional neural network
correlated filtering
fusion response
