摘要
目前主流的判别式目标跟踪模型大多使用灰度、颜色等手工特征,在目标快速移动或受到视频序列背景等因素干扰情况下,目标跟踪器可能在跟踪目标时学习到错误特征而导致跟踪失败。因此,提出一种结合深度特征的相关滤波跟踪算法。首先将待跟踪目标图像输入至卷积神经网络中,提取出较高层的卷积特征,然后将提取的卷积特征输入相关滤波器中得到响应,最后根据响应峰值得到追踪结果。以VOT2016中包含人体运动的视频序列为实验数据集,并分别与CN、SAMF及KPDCF模型进行对比。实验结果表明,结合深度特征的相关滤波算法具有较好的追踪性能,在不大幅降低追踪速度的情况下,提升了追踪精度和稳定性。
At present,most of the dominant discriminative target tracking models use manual features such as grayscale and color,so that when the target moves quickly or is interfered by factors such as the background of the video sequence,the tracking may fail for the target tarcker may learn the learn wrong features in tracking.Therefore,a correlation filter tracking algorithm combining depth fea⁃tures is proposed.Firstly,the image of the target to be tracked is input into the convolutional neural network to extract the convolution features of the higher layer,and then the extracted convolution features are sent to the correlation filter.Get a response,and finally get the tracking result based on the peak in the response.The video sequence containing human motion in VOT2016 was used as the experi⁃mental data set and compared with CN,SAMF and KPDCF models respectively.The experimental results show that the correlation fil⁃tering algorithm combined with the depth feature has better tracking performance,which can improve the tracking accuracy and stabili⁃ty without greatly reducing the tracking speed.
作者
蒋宇
袁健
JIANG Yu;YUAN Jian(School of Optoelectronic Information and Computer Engineering,University of Shanghai for Science and Technology,Shanghai 200093,China)
出处
《软件导刊》
2020年第1期89-94,共6页
Software Guide
基金
国家自然科学基金项目(61775139)
关键词
手工特征
相关滤波器
深度特征
目标追踪
卷积神经网络
人体运动序列
manual feature
correlation filter
depth feature
target tracking
convolutional neural network
human motion sequence
