摘要
视频目标跟踪是当前计算机视觉领域中的一个关键问题。基于场景分类提出了一种视频目标特征的实时在线选择方法。离线时首先采用改进型的"空间金字塔匹配"算法完成先验视频场景的分类;然后采用不同的描述算法实现对目标和背景的特征描述,通过对数形式的方差比计算出目标和背景的特征距离;最后融合均值和熵值对特征距离进行统计分析,建立场景相关的描述子显著性排序。在此基础上,在线时利用初始帧完成测试视频的场景判定,结合当前场景下的描述子排序,实时选择最优特征,应用于粒子滤波跟踪系统,在公开视频库上进行跟踪测试,验证排序的正确性和选择机制的必要性。
Video target tracking is a key issue in the field of computer vision.Based on the scene classification an online method for video target feature selection is proposed.During off-line,the modified spatial pyramid matching algorithm（NEW-SPM）is used to complete the sample-videos classification at first.And then the rough localization of target and background region is realized with TLD tracking algorithm,and the characteristics description the target and background is completed with different descriptors.After this,the feature distance between target and background is calculated through the form of log variance ratio.And finally,the fusion mean and entropy statistical analysis is carried out on the characteristics of distance for the saliency sorting of different descriptors.On this basis,the scenarios related characteristics saliency lists are constructed.Combining with the testing scenes judgment,the lists are applied to the on-line particle filter tracking system to select the best descriptor in the current category,and it is tested on the public library of tracking,the validity of the lists and the necessity of selection method are proven.
出处
《光学技术》
CAS
CSCD
北大核心
2014年第6期551-559,共9页
Optical Technique
基金
光电控制技术重点实验室和航空科学基金联合资助项目(20125186005)
关键词
目标跟踪
TLD跟踪算法
显著性排序
实时特征选择
粒子滤波
跟踪精度
target tracking
TLD tracking algorithm
saliency list
on-line feature selection
particle filter
tracking precision
