Crowd Abnormal Behavior Detection Based On Video Surveillance

In recent years,terrorist attacks and mass violence occur frequently.The intelligent video surveillance system arises at the historic moment in order to protect people's lives and property safety.Among lots of functions of the system,the abnormal behavior detection is one of the most important and it has attracted the attention of many researchers.Its purpose is to detect abnormal behavior in surveillance scenes and to alert in real-time by using the technology of computer vision and video content analysis.In this paper,two kinds of anomaly detection methods are proposed to detect local anomaly and global anomaly in surveillance video sequences.For local anomaly detection,a local abnormal behavior detection method based on motion features and appearance features is proposed,which is used to detect objects with speed anomaly and appearance anomaly in the surveillance scene.In order to detect anomaly caused by the change of motion features,the method extracts optical flow information of input image sequences and calculates motion energy value of each non-overlapping spatial-temporal cube by using frame-level and pixel-level motion vector.Then,the method estimates the boundary of motion energy value through limiting the data-increasing rate and such boundary is used to distinguish abnormal behaviors from normal ones.In order to detect anomaly caused by the change of appearance features,the method extracts spatial-temporal gradient as the feature descriptor and use SVDD to establish an optimal hypersphere with most of training samples inside it.The distance of testing sample and the center of hypersphere is calculated to distinguish abnormal behaviors from normal ones.For global anomaly detection,a global abnormal behavior detection method based on motion information entropy is proposed,which is used to detect escaping people in the square.The method utilizes the direction and magnitude of motion vectors extracted from optical flow field to calculate the motion information entropy which represents the uncertainty of motion information in surveillance scene.The normal samples are distributed around the center of a Gaussian distribution while the abnormal ones are distributed on the side.In this paper,two public datasets are used to evaluated the performance of proposed methods including ROC curve,AUC,EER and RD.Experimental results show that the proposed methods have better detection performance than most of other classic methods and the computational complexity is much lower than those ones' which are far from real-time detection.