基于时空惊奇计算的视频异常检测方法

Anomaly detection method in video based on spatio-temporal surprise computation

提出一种基于贝叶斯惊奇计算的视频异常检测方法.用块匹配运动估计方法提取运动特征(如运动幅度、方向),得到多尺度运动矢量直方图.使用空间维度与时间维度上惊奇计算相结合的度量方法,既可以检测"个体异常行为",也可以应用于"群体异常行为"检测.实验表明,该算法是鲁棒和实用的,且易于实现.

We propose an anomaly detection method in video based on Bayesian surprise computation. We use the block-matching motion estimation method to extract low-level motion features （such as magnitude and direction of motion） and then calculate multi-scale histogram of motion vector. We use both spatial surprise and temporal surprise to detect not only ＂individual abnormal behavior ＂ but also ＂ group abnormal behavior＂. Experimental results show that our algorithm is robust and applicable and it can be easily implemented.