人群中的异常行为是一大潜在威胁,自动检测监控中的异常行为成为近年的研究热点之一。然而,由于异常的未知性与复杂性,已有的检测方法仍然存在检测率低、定位精度差的问题。为此,提出了对视频监控中的人群异常行为自动检测与定位的方法。结合灰度值与光流场的分布提取运动区域;对运动区域分割得到有效的运动块,从中提取表示外观和动态的两种特征,即局部H梯度方向直方图G和局部H光流方向直方图F特征;使用k-means方法对运动块进行聚类,对每类样本使用一类分类器进行建模。最后,加入运动连续性约束,以抑制干扰噪声。在两个复杂的异常行为数据集上的实验结果表明,本文方法明显优于已有的检测方法,且可以满足正确率高、抗干扰能力强等实际工程需求。

The anomaly in the crowd is a great potential threat, and the automatic detection of abnormal behavior for surveillance has become a hot topic in recent years. However, because the anomaly is unknown and complex, the previous detection methods still suffer from a low detection rate and poor location accuracy. To this end, a method is proposed for anomaly detection and location in the crowded surveillance videos. First, the motion regions are extracted according to the distributions of the gray-scale value and the optical flow field. Second, the effective motion blocks are obtained by segmenting the motion regions. Two features, namely the local H histogram of gradient G and the local H histograms of flow F, are extracted from the motion blocks, representing the appearance and dynamics. Third, the motion blocks are clustered with the k-means method, and each cluster is modeled using a one-class classifiers. Finally, the motion continuity constraint is added to suppress the noisy noises. Experimental results on two complex abnormal behavior datasets show that the proposed method is obviously better than previous detection methods. It could meet the practical engineering needs such as high accuracy and strong anti-interference ability.