系统工程与电子技术 ›› 2019, Vol. 41 ›› Issue (5): 951-957.doi: 10.3969/j.issn.1001-506X.2019.05.03

• 电子技术 • 上一篇    下一篇

多特征融合的尺度自适应相关滤波跟踪算法

张红颖, 胡文博   

  1. 中国民航大学电子信息与自动化学院, 天津 300300
  • 出版日期:2019-04-30 发布日期:2019-04-26

Scale-adaptive correlation filter tracking based on multiple features

ZHANG Hongying, HU Wenbo   

  1. College of Electronic Information and Automation, Civil Aviation University of China, Tianjin 300300, China
  • Online:2019-04-30 Published:2019-04-26

摘要:

针对传统的核相关滤波跟踪算法特征表达不足且不能解决尺度变化导致的跟踪失败问题,提出了一种多特征融合的尺度自适应相关滤波跟踪算法。首先,在位置预测阶段,利用Hue和梯度方向直方图两种特征和各自的位置滤波器作用得到的输出响应值,自适应分配权重实现目标位置的预测。其次,在尺度预测阶段,通过在目标位置采集多尺度图像,组成样本独立训练尺度滤波器,并根据样本的尺度滤波响应值,对目标的尺度进行估计,使跟踪算法能够适应目标的尺度变化。最后利用两帧图像间的差异性,自适应调整学习率,进行位置滤波器的更新。实验结果表明,改进算法相较于传统的核相关滤波跟踪算法,在跟踪成功率和跟踪精度上分别提高了25%和13%。当目标尺度变化较大时,改进算法仍能稳定跟踪。

关键词: 目标跟踪, 相关滤波, 特征融合, 尺度自适应

Abstract:

Aiming at the tracking failure caused by scale variation and single feature in the traditional kernel correlation filter tracking algorithm, the scale-adaptive correlation filter tracking based on multiple features is proposed. Firstly, in the position prediction stage, the characteristics of Hue feature and histogram of oriented gradient are analyzed, and the target location is predicted by distributing the weight of the output response values of the two features and each translation filter. Then, in the scale prediction stage, a scale filter is trained independently by the multi-scale image sampling in the target position, and the scale of the target is estimated according to the scale filter response value of the sample, which enables the tracking algorithm to adapt to the scale change of the target. Finally, the difference between two frames is used to adjust the learning rate adaptively to update the model of the translation filter. The experimental results demonstrate that this algorithm is higher than the traditional kernel correlation filter tracking algorithm by 25% in the tracking success rate and by 13% in the tracking precision. Furthermore, the improved tracking algorithm is still able to track the moving targets stably and accurately in the case of scale variation of the targets.

Key words: targets tracking, correlation filter, feature fusion, scale-adaptive