基于行为减法的视频异常检测研究

本文提出了一种基于行为减法的视频异常检测研究的方法。与传统的视频异常检测方法相比，该方法不需要先对目标进行标签、识别、归类和跟踪，因此，需要的计算量和内存消耗较少，实时性良好。基于物理世界中的事件都是时空相关的，该方法很好地利用了事件的时空特性。在动态特征检测的预处理基础上，直接在像素点上进行操作。在训练阶段，对每一个像素点，先建立一个时空共生模型，通过建模，计算正常事件概率；然后在检测阶段，采用相同的模型，将计算获得的概率值经过阈值比较的方法，确定该点是否为异常。通过实验证实，该方法在视频异常检测中具有高效性，并且可以应用在很多场合。

Video anomaly detection using behaviour- subtraction-based method

In this paper,we proposed a behaviour-subtraction-based method to detect anomalous events in the traffic videos.Different with the traditional approaches,objects identifying,classifying or tracking is not needed at first in our method,so it calls for less computation and memory consumption.Because the physical world events are spatiotemporal related,our approach makes good use of this characteristic.while motion label extraction is done,our method works at the pixel level to build spatiotemporal co-occurence models of normal data and then attempts to detect deviations from the normal models in observed data.Confirmed by experiment,our approach has a high efficiency,and can be used in many occasions.