The Crowd Abnormal Event Detection Based On Surveillance Video

In recent times,there have been an increasing number of terrorist attacks,crowd stampedes and other public safety events.Detection of abnormal crowd events which is the foundation of monitoring,analysis and early warnings for crowd movement,has become one of the most urgent problems in intelligent surveillance.However,since the crowd movement is complicated,it is almost impossible to list all possible anomaly event in the crowd.Therefore,the detection of abnormal events is not a typical classification problem.To solve this problem,we adopt the following technical route.Firstly,we train a model on a set of surveillance video which only contain normal crowd events.Then,we judge whether the crowd movement is abnormal or not by calculating the degree of deviation between pre-model and crowd movement feature of testing video.Specifically,we divide the detection of abnormal events into the two stages: one is video feature extraction namely event representation,the other is model training and anomaly detection.For feature extraction,two different methods are proposed in this paper.The one is based on repulsive force model.To avoid some problem resulting from tracking all objects,especially in highly dense crowds,a holistic grid particle advection is adopted to model crowd motion.Then,the force matrix obtained by a repulsive force model,which can reflect crowd movement precisely,is regard as feature.The other is based on CNN.Although the first method has achieved good performance in a variety of scenes,it is not sensitive to the scene with tiny change of the repulsive force.Hence,we adopt the method which integrate an improved CNN model,a sliding window and PCA to extract temporal and spatial features of crowd video.For model training and abnormal event detection,this paper propose a group of visual dictionary based on sparse coding to decrease the computational complexity and difficulty in maintaining a single dictionary.For abnormal event detection,we use each dictionary in turn to reconstruct a testing word.Once the testing word can be sparsely represented by a dictionary,we can consider this word is normal and put it into the word pool which is corresponding to the dictionary.If all dictionaries can't sparsely represent a testing word,this word is regarded as anomaly.To solve the problem of dictionary degradation and concept drift with new video increasing or the dynamic changes of video scene,we propose a fully unsupervised global and local online updating algorithm,based on sparse reconstruction and a group of word pools.We experimentally verify our method compared with other methods on the UMN dataset,the UCSD dataset and the Web dataset separately.The results indicate that our method not only increase the accuracy and efficiency of crowd abnormal detection,but also the method based on CNN improve the disadvantages of the method based on repulsive force model in the scene with tiny change of the repulsive force.