欢迎访问《图学学报》 分享到：
图学学报

图学学报

• 视觉与图像 • 上一篇    下一篇

抑制背景下多重滤波目标跟踪

  

  1. 1. 三峡大学电气与新能源学院，湖北 宜昌 443002；
    2. 三峡大学湖北省水电工程智能视觉监测重点实验室，湖北 宜昌 443002
  • 出版日期:2017-06-30 发布日期:2017-07-06
  • 基金资助:
    国家自然科学基金项目(61273183)；湖北省教育厅科学技术研究计划重点项目(D20151204)；湖北省创新群体项目(2015CFA025)

Target Tracking with Background Suppression and Multiple Filtering

  1. 1. College of Electrical Engineering & New Energy, Three Gorges University, Yichang Hubei 443002, China;
    2. Hubei Provincial Key Laboratory of Intelligent Visual Inspection of Hydropower Engineering, Three Gorges University,
    Yichang Hubei 443002, China
  • Online:2017-06-30 Published:2017-07-06

PDF (PC)

271

摘要： 为解决运动目标稳定跟踪与实时性要求，提出一种背景抑制下多重滤波(BSMF)
跟踪算法，对跟踪目标的中心点计算、动态模板学习和变尺度问题进行研究。首先，以目标
为中心的跟踪区域的RGB 图像映射为11 维颜色，提取目标主要颜色的描述向量，生成抑制
背景的权值矩阵；然后与跟踪区域灰度图的对应点相乘，用相关滤波方法计算抑制背景后的
跟踪区域目标中心位置；最后将定位的目标区域进行金字塔滤波计算，得到尺度变化系数。
该算法采用动态的匹配模板学习策略，降低各种干扰因素混入匹配模板内，在树莓派3 代板
嵌入式系统中，平均可达17 帧/秒的跟踪速度，且比同级别复杂度跟踪算法具有更稳定的跟踪
性能。

关键词: 抑制背景, 多重滤波, 相关滤波, 目标跟踪, 动态学习

Abstract: In order to design a moving target tracking algorithm with respect to real-time and stable
tracking process, an approach named background suppression multiple filtering (BSMF) was
proposed, which researched the problem of computing target center position, dynamic matching
temple update, and scale-variant. Firstly, the algorithm maps the RGB figure of target area to 11
dimension color vectors to extract target main color describe vectors, get a background suppression
weighted matrix, and then it multiplies with the gray tracking area pix matrix, the target center
position in tracking area was calculated by correlation filter method after background suppression.
Finally, located target area was calculated by pyramid-filter to get scale change coefficient.
Meanwhile, BSMF algorithm was used to decrease various interference factors mixing up in matching
template by dynamic matching template strategy. And the algorithm could achieve 17 FPS when run
at Raspberry Model 3 and tracking performance is better than other algorithms which have the same
computational complexity.

Key words: background suppression, multiple filtering, object tracking, correlation filter, dynamic learning