摘要
在目标跟踪算法中,相关滤波算法近几年来发展迅速,成为了该领域的研究热点。相关滤波跟踪算法具有速度快、效果好等优点,但受限于传统手工特征对目标表达能力不足,仍然难以应对诸如形变、遮挡、模糊等情形。最近,卷积神经网络在诸多领域取得了极大的成功,研究人员将相关滤波与卷积特征相结合,克服了传统手工特征缺少目标语义信息的缺点。为了有效处理目标外观变化,文中提出一种融合多层卷积特征的相关滤波运动目标跟踪算法。该算法将目标跟踪分为预测位置和估计尺度两个步骤:提取多层卷积特征并在每个卷积层上估计目标位置,通过固定权重将所有卷积层的结果融合以确定目标的最终位置;确定位置后通过提取目标多个尺度的方向梯度直方图特征来估计目标的最佳尺度。在公开数据集中选取20段视频来验证所提算法,并将该算法与4种运动目标跟踪算法进行比较。实验数据表明,与次优的基于传统手工特征的DSST算法相比,所提算法的距离精度提高了48.9%,重叠精度提高了51.9%;与同样使用卷积特征的HCFT算法相比,其距离精度提高了19.1%,重叠精度提高了25.2%。文中提出的算法较好地克服了传统手工特征表达能力弱的缺点,其性能优于使用手工特征的传统相关滤波跟踪算法,相比同样使用卷积特征的相关滤波算法也有所提高。在目标发生遮挡、模糊等复杂情况下,该算法仍然能够准确跟踪目标。
In the visual object tracking,correlation tracking algorithm is a hot topic and has developed rapidly in recent years.Correlation filter tracking algorithms have the advantages of fast speed and good effect.However,the traditional hand-crafted features are insufficient for target discrimination,and fail to handle challenging situations such as deformation,occlusion and blurring.Recently,convolutional neural networks have achieved great success in many fields.Researchers have combined correlation tracking and convolutional features to surmount the shortcomings of hand-crafted features that lack target semantic information.In order to cope well with the above problems,this paper proposed a vi-sual object tracking algorithm based on correlation filter with hierarchical convolutional features.The proposed algorithm divides the object tracking into two steps,including position prediction and scale estimation.Multi-layer convolutional features are trained and the target position on each convolutional layer is predicted with a coarse-to-fine searching approach.The Histogram of Oriented Gradient features is used to estimate the optimal scale of target.Comprehensive experiments on 20 challenging sequences were performed to verify the proposed algorithm,and the proposed algorithm was compared with other four trackers.The results show that the proposed approach significantly improves the performance by 48.9%and 51.9%in distance precision and overlap precision respectively compared to the DSST tracker based on hand-crafted features.Moreover,the proposed method outperforms the HCFT tracker using convolutional features by 19.1%and 25.2%in distance precision and overlap precision,respectively.The proposed algorithm overcomes the shortcomings of poor representation skills of traditional manual features,and its performance is better than the correlation filtering tracking algorithms using manual features.Compared with the same correlation filtering algorithms using convolutional features,the tracking performance has also been imp
作者
李健鹏
尚振宏
刘辉
LI Jian-peng;SHANG Zhen-hong;LIU Hui(Faculty of Information Engineering and Automation,Kunming University of Science and Technology,Kunming 650504,China)
出处
《计算机科学》
CSCD
北大核心
2019年第7期252-257,共6页
Computer Science
基金
国家自然科学基金项目(61462052,11873027)资助
关键词
目标跟踪
相关滤波
卷积特征
尺度估计
Visual object tracking
Convolutional feature
Scale estimation
Correlation filter
