基于典型相关树加权置信传播的运动目标检测  

Moving target detection based on canonical correlation tree weighted belief propagation

作　　者：董安国[1] 李聪[1] 

机构地区：[1]长安大学理学院,陕西西安710064

出　　处：《江苏大学学报（自然科学版）》2015年第6期686-690,共5页Journal of Jiangsu University：Natural Science Edition

基　　金：国家自然科学基金资助项目(11171043;11201038);中央高校基本科研业务费专项资金资助项目(CHD2012TD015)

摘　　要：针对基于视频的运动目标检测中视频背景复杂多变的特点,提出一种基于典型相关树加权置信传播的算法进行运动目标提取.首先将视频图像分成大小相同的子块,构造出有环图模型;再运用tree-reweighted算法将环路分解成生成树的形式,进行有环图模型的目标优化检测;然后利用典型相关分析求解相邻子图像块之间的典型相关系数值,选取典型相关系数值最大的两子图像块进行连接,组成新的环路;最后利用树加权置信传播算法迭代更新信息,实现视频运动目标的检测.试验结果表明,该算法的运行时间为9.8 s,与原图像的相似度可达到95%以上,因此它可以比较准确检测分离出视频序列运动目标,且稳定性好,运算时间较短,适合于对运动目标的实时检测.To solve the complexity and variability of video background for moving object detection in video, a new algorithm was proposed based on the canonical correlation tree weighted belief propagation. The image was separated into some blocks with the same size to establish loop model. The tree-reweighted algorithm was used to decompose the loop into spanning tree, and the detection of moving targets on the loop model was achieved. The canonical correlation analysis was used to solve the canonical correlation coefficients between adjacent sub-blocks, and then the two sub-blocks with the maximum value were linked to form new loop. The iteration of the tree weighted belief propagation was adopted to update the information, and the moving targets were detected. The experimental results show that the run time of the proposed algorithm is 9. 8 s, and the similarity with the original image is over 95%. The proposed algorithm can detect and separate the moving target accurately with reliable speed, and it is suitable for real time detection of moving targets.

关 键 词：运动目标检测 置信传播 典型相关 有环图模型 生成树 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]