• 中国计算机学会会刊
  • 中国科技核心期刊
  • 中文核心期刊

计算机工程与科学

• 人工智能与数据挖掘 • 上一篇    下一篇

基于融合特征的多尺度快速相关滤波跟踪算法

火元莲,曹鹏飞,董俊松,石明   

  1. (西北师范大学物理与电子工程学院,甘肃 兰州 730070)
  • 收稿日期:2018-04-18 修回日期:2018-06-20 出版日期:2019-03-25 发布日期:2019-03-25
  • 基金资助:

    国家自然科学基金(61561044);甘肃省高等学校科研项目(2016A004)

A multi-scale fast correlation filter tracking
 algorithm based on fusion features

HUO Yuanlian,CAO Pengfei,DONG Junsong,SHI Ming   

  1. (College of Physics and Electronic Engineering,Northwest Normal University,Lanzhou 730070,China)
  • Received:2018-04-18 Revised:2018-06-20 Online:2019-03-25 Published:2019-03-25

摘要:

针对复杂场景下目标遮挡和尺度变化所导致的跟踪效果不佳问题,提出一种基于融合特征的多尺度快速相关滤波跟踪算法。首先,对目标的3种特征降维融合构成特征矩阵;其次,采用主成分分析思想实时地提取显著特征,重构特征矩阵,在有效降维的同时训练位置相关滤波器;最后,利用融合特征矩阵训练尺度相关滤波器,从而准确预测目标位置和尺度。实验部分将改进算法与目前流行的相关滤波跟踪算法进行比较,结果表明,改进算法在目标遮挡和尺度变化场景下跟踪精度较高,平均跟踪速度达到52.5 frame/s。

关键词: 目标跟踪, 相关滤波, 特征融合, 主成分分析

Abstract:

We propose a multiscale fast correlation filter tracking algorithm based on fusion features to solve the problem of poor tracking effect caused by target occlusion and scale change in complex scenes. Firstly, the dimensions of the three features of the target are reduced and fused to form a feature matrix. Secondly, the principal component analysis is used to extract the salient features in real time, reconstruct the feature matrix, and position correlation filters are trained while reducing the dimension effectively. Finally, the fusion feature matrix is adopted to train scale correlation filters, thus the position and scale of the target is accurately predicted. We compare the improved algorithm with popular correlation filter tracking algorithms by experiment, and the results show that the improved algorithm has a higher tracking accuarcy and an average tracking speed of 52.5 frame/s in scenarios of target occlusion and scale change.
 

Key words: object tracking, correlation filter, feature fusion, principal component analysis