摘要
设计实现了一种基于相关滤波目标跟踪的特征模型融合算法。该算法利用卷积神经网络提取深度特征,通过相关滤波器得到滤波响应图模型;同时利用传统的特征提取方式,得到颜色直方图特征概率模型,用来弥补相关滤波算法的边界效应问题和模板类方法不能有效处理目标变形的问题;之后将两模型进行融合,最终确定目标位置。实验证明,本文算法在VOT测试数据集上可达到的较为理想跟踪效果,与CF2算法相比提高了目标跟踪的性能。
Visual target tracking technology is one of the most active research fields in machine vision,target tracking algorithm based on correlation filter has become a hotspot in recent years because of its high real-time performance.But the correlation filter tracking algorithm is easy to lose the target when the target moves too fast and changes.To solve this problem,this paper designs and implements a feature model fusion algorithm based on correlation filter target tracking.The algorithm uses convolutional neural network to extract depth features and correlation filters to obtain the filter response graph model.At the same time,the probability model of color histogram feature is obtained by using traditional feature extraction method,which can compensate for the boundary effect of correlation filter algorithm and the problem that template methods can not effectively deal with target deformation.Then,the two models are fused to determine the final target.Label position.Experiments show that the proposed algorithm achieves ideal tracking effect on VOT test data sets and improves the performance of target tracking.
作者
吴雨秋
李朝晖
WU Yu-qiu;LI Zhao-hui(Information Engineering School,Communication University of China,Beijing 100024,China)
出处
《中国传媒大学学报（自然科学版）》
2020年第2期50-58,共9页
Journal of Communication University of China：Science and Technology
关键词
目标跟踪
相关滤波
颜色直方图
深度特征
Target tracking
correlation filtering
color histogram
depth feature
