Research On Abnormal Event Detection In Crowded Scene

Abnormal event detection in video is an important research in the intelligentmonitoring system, and it is becoming a focused application field in recent years. Thecrowded scene is the public place where population has a big liquidity and density, suchas subway, square and so on. In these places, mass disturbances happen easily. If it can’tbe dealt well, it will produce serious influence. Therefore, it is required to researchdeeply on abnormal event detection in crowded scene. In the uncrowded scene, trackingthe target is usually used to infer the abnormal event. However, in the crowded scene, itis hard to separate every target or event, so more efficient methods are using localfeatures to represent abnormal event, which contains space-time information, motionfeatures and optical flow histogram. According to region of detection, the abnormalevent detection contains two parts: local abnormal event detection and global abnormalevent detection. The topic researches the local abnormal event detection and the globalabnormal event detection in the crowded scene separately.The topic research a method based on space-time MRF framework to detect thelocal anomaly. To construct the space-time MRF, choose a fixed length of frames invideo, and divide each frame into small regions that represent the nodes in space-timeMRF. The features of node can be represented by optical flow, and the L-K opticalflow method is used in the topic. However, the original optical flow can’t represent themotion features of node well, which means the features need to be modeled. As theGaussian Mixed Model, MPPCA (Mixture of Probabilistic Principal ComponentAnalysis) combine several PPCA (Probabilistic Principal Component Analysis) torepresent the complex data better. So this topic uses MPPCA to model the optical flowfeatures of nodes. The energy minimization problem of MRF is NP-Hard, but it can becomputed by approximation algorithm, such as Belief Propagation. The label of eachnode in space-time MRF can be computed by Belief Propagation, which stands for thestate (normal or abnormal) of node. And the labels are regarded as the basis to detectlocal abnormal event.With regard to the global abnormal event, the topic research a method based onspares representation and improve the algorithm. The optical flow histogram of eachframe can be extracted from video, and if represent the features vector based on agiven features vector using sparse representation, then the sparse efficient of differentfeatures vector is different. Choosing the features vector of normal frames in video tolearn dictionary, a features vector basis can be required, which is a dictionary. If werepresent the features vector of test frames using the learned dictionary, the sparseefficient value of abnormal frames fluctuates widely. For quantizing the extent ofanomaly, a sparse representation cost function is constructed based on the sparse efficient value, which is regarded as the judgment whether the frame is abnormal. Inthe process of sparse representation, a fast approximation sparse representation methodcan be achieved by reducing the items in dictionary, which can improve the speed ofthe process. Then it can be added to the video anomaly detection framework, andimprove the processing speed of each frame. As a result, this method can ensure thejust-in-time of the process of abnormal detection.