摘要
为了获取具有运动前景物体的初始背景,提出一种基于聚类识别的背景初始化算法。首先利用滑动可变窗口检测每个像素的所有不重叠平滑子序列,获取可能背景;然后选择每个平滑子序列的中值样本点构建分类序列集,根据未知类别的无监督聚类识别思想获取背景子集,实现背景初始化。选取不同交通状态的视频训练序列,将本文方法同中值法、一致性检测法进行了对比实验。结果表明,本文方法具有良好的适应性,可克服缓慢运动大型前景物体造成的影响,实现覆盖率大于50%的背景初始化。
To get the initialized background from the training sequence with foreground objects, a new background initialization algorithm was proposed based on clustering classifier. In this method, all stable sub-intervals in the training sequence were located for each pixel as possible background. Then a classify data set was constructed by the median values of each sub-interval. A background sub-set was obtained from the data set by unsupervised clustering. Accordingly the initialized background was obtained from the background sub-set. By experiments, the proposed method was compared with the median method and the sample consensus method under different traffic conditions. Results show that this method is robust, overcomes the influence of slow moving big objects, and tolerates over 50% foreground pixels and noises.
出处
《吉林大学学报（工学版）》
EI
CAS
CSCD
北大核心
2008年第1期148-151,共4页
Journal of Jilin University:Engineering and Technology Edition
基金
“973”国家重点基础研究发展规划项目(2006CB705500)
吉林省科技厅国际合作项目(20040705-2)
人事部归国优秀人员项目
关键词
计算机应用
视频检测
背景模型
背景初始化
交通流检测
无监督聚类
computer application
video detection
background model
initialization background
traffic flow detection
unsupervised clustering
