Research On Abnormal States Detection And Prediction In Crowd Videos

Recent years,public places,such as sports events,festival celebrations,performance venues,and large-scale business districts,often attract a large number of enthusiastic participants.Stampedes are most likely to occur in such places.Crowd disasters appear frequently,thus,researchers should pay more attention in crowd behavior analysis technologies based on video surveillance.The applications of crowd abnormal states detection and prediction can find crowd abnormal events promptly.Supported by these technology,the loss of crowd disasters can be reduced.Crowd abnormal events are usually detected through detection models,which are often learned from normal crowd behavior samples.The models tend to judge the events of low probability to be normal behaviors as abnormal events.Crowd abnormal events prediction technology can find the abnormal states before their occurrences.The prediction methods can provide more time to avoid human disasters.These methods usually deduce the future states of crowds,then,abnormal events can be predicted in the future states.The primary researches in this study included:(1)According to the analysis of high semantic features of crowded scenes,a crowd abnormal states detection method based on slow feature analysis is proposed.Usually,the responses of retinal receptors or the gray pixel values of a CCD camera vary quickly within a short period of time;on the other hand,high level responses in a human brain tend to vary slowly for a longer time.Slow feature analysis models this action of human brain,and transforms the fast varying low semantic visual features into high semantic features by a transform function.High semantic features represent discriminative information of the scene,which make the abnormal behavior can be detected easier.This work divides human scenes into many equal spatio-temporal regions.Considering each local region as a node,a graph model is established.By the representation of weights between nodes,the information of feature,time,and space domains are included in our model.Experimental results suggest that the extraction of high semantic feature with spatio-temporal information can improve the detection performance.(2)By analyzing the disadvantages of outline learning methods,A new method called double sparse representation is proposed to detect crowd abnormal behaviors.The method contains two sparse representation processes.One of the processes contains a dictionary filled by normal samples,the other contains a dictionary filled by abnormal samples.A test sample's result is given by integrating the results of two sparse representation processes.As discriminative information is included in the model,the performance of the model improves.Besides,the method contains an online learning process to add testing samples into the two dictionaries,which is called dynamic dictionary updating process.The process makes the dictionary more complete,which produces more accurate detection results.The results of experiments conducted on various datasets show that the proposed method achieves higher accuracy than state-of-the-art methods in local and global abnormal events detection.(3)After analyzing solution of error accumulation problem generated by online learning methods,An Crowd abnormal state detection learning method based on controlled random walk and adaptive sample selection is proposed.A manifold learning method based on controlled random walk graph model is incorporated to generate reliable predictions on unlabeled data for online learning of abnormal events detection.The method makes the labelling action avoid dense areas,which may contain wrong labelled samples.Meanwhile,it let the labelling actions reach the labelled samples.Then,the wrong labelled samples have the chance to be corrected.An adaptive sample selection process,which aims to minimize the expect classification error rate,is proposed to find the optimal testing samples and add them into the training set automatically.The self-training algorithm embedded by the controlled random walk graph model and the adaptive strategy overcome the disadvantages caused by wrong labelled samples in online learning.The experiment results show that the proposed self-learning algorithm can improve the detection results of most abnormal events detection models.(4)By analysing the evolution and prediction of crowd states,a purpose driven lattice Boltzmann model is proposed.A dense crowd is similar to a fluid.Both of them have collision and streaming processes.In addition,individuals in dense crowd have the trend to follow the crowd's main motion.The trend is generalized as purpose drive in this work.Then,the purpose drive is added into the lattice Boltzmann model.Through purpose driven lattice Boltzmann model,the future states of dense crowd can be predicted.By behavior entropy model,the abnormality of predicted states can be detected.Experimental results show that the purpose driven lattice Boltzmann model has a strong ability to predict abnormal crowd behaviors.In this work,the slow feature analysis is used to extract high semantic features for abnormal motion detection.Double sparse representation is used to include abnormal samples in the learning of detection model.With a dynamic dictionary updating process,the training samples are used to optimize the detection model.The problem of small sized training sample is reduced by this online method.The use of controlled random walk and adaptive sample selection overcome the disadvantages of wrong labelled samples in online learning of detection model.After analyzing the differences and similarities between dense crowds and fluids,the purpose drive of human is added into the lattice Botlzmann model.The purpose driven lattice Boltzmann model can deduce the crowd future states,then,the abnormal events can be predicted.