中文摘要：

提出了一种基于光流块统计特征的视频异常行为检测算法.该算法首先对训练集视频序列的光流场进行分块及预处理,而后提取光流块的统计特征,所提取的块统计特征同时包括了光流块的幅度信息和相位信息,通过训练集得到的光流块统计特征训练出对应的正常行为的高斯混合模型（GMM）.测试集通过同样的方式提取光流块统计特征,通过计算所提取统计特征以多大的概率属于GMM判定所检测光流块的异常程度.实验结果表明,该算法能够在一定程度上解决运动物体一致性和部分遮挡问题,并提高了异常行为检测的准确率.

英文摘要：

An anomaly detection algorithm based on the statistic feature of optical flow block was proposed.First,the whole optical flow field of training video sequences were obtained.Then,each optical flow field was divided into blocks and each block was preprocessed in order to extract the statistic feature considering both magnitude and phase information of the block.The Gaussian mixture model（GMM）was employed to establish the probability model of normal behaviors by feeding the statistic feature into it.The abnormal degree of the optical flow block was judged by the output posterior probability of the GMM probabilistic model.The experimental results show that the method proposed considers both the consistency information of moving objects and the partial occlusion issue,at the same time,improves the accuracy of anomaly detection.