稀疏表示下大场景视频图像运动目标跟踪仿真
Sparse representation of large scene video image motion object tracking simulation
刘洪鹏 1孙永佼2
作者信息
- 1. 宁夏理工学院计算机科学与工程学院,宁夏 石嘴山 753000
- 2. 东北大学计算机科学与工程学院,辽宁 沈阳 110819
- 折叠
摘要
大场景视频中存在复杂的背景信息,这些背景信息与目标的外观相似,导致难以准确跟踪目标.为了解决这一问题,提出基于稀疏表示的大场景视频图像运动目标跟踪方法.通过提取视频图像haar-like特征并对特征投影压缩,在字典中融入背景信息,构造超完备字典.改进传统粒子滤波,提出基于残差的Unscented粒子滤波算法,在传统稀疏表示中添加判别函数生成判别稀疏表示,利用L1 范数最小化求解候选目标稀疏系数,在字典不断更新下通过基于残差的Unscented粒子滤波算法,实现大场景视频图像运动目标跟踪.实验结果表明,所提方法目标跟踪精度和重合度均在 0.9 以上,且跟踪平均时间短.
Abstract
In large scene videos,there is a large amount of complex background information,which is similar to the appearance of the target,so it is difficult to accurately track the target.To solve this problem,this article put for-ward a method for tracking motion targets in large scene video images based on sparse representation.Firstly,Haar-like features were extracted from video images and compressed.Secondly,background information was incorporated in-to the dictionary to construct an over-complete dictionary.Then,the traditional particle filter was improved.Mean-while,an Unscented particle filter algorithm was proposed based on residual error.Moreover,a discriminant function was added to the traditional sparse representation,so that a discriminant sparse representation was generated.Based on L1 norm minimization,the sparse coefficients of the candidate target were solved.Under continuous dictionary up-dates,the motion target tracking was realized through the residual-based Unscented particle filter algorithm.Experi-mental results show that the target tracking accuracy and coincidence rate are more than 0.9,and the average tracking time was short.
关键词
稀疏表示/大场景视频图像/运动目标跟踪/haar-like特征/粒子滤波Key words
Sparse representation/Large scene video/Moving object tracking/Particle filtering引用本文复制引用
出版年
2024
NETL
NSTL
万方数据