Crowd Anomaly Detection Based On Hidden Markov Model

Video technology that has been widely used bring a large amount of video data,so it is impossible to detect abnormalities in surveillance video rely on human operators only.In order to reduce the burden of human resources and economic and improve the accuracy of anomaly detection,people are constantly looking for the method which could detect abnormalities in video automatically.Crowded scenes are very susceptible to accidents,so it is especially important to strengthen the anomaly detection in crowded scenes.A method is proposed to detect and classify abnormalities in crowded scenes by Hidden Markov Model(HMM).In this paper,abnormalities in crowded scenes are defined as abnormal targets that appear in dense crowd,such as motor vehicle,bike and skater.In addition,group behavior in crowd are also abnormalities,such as population flee suddenly.The purpose of this paper is to detect the abnormal targets and abnormal group behavior,and classify the abnormal targets.Hidden Markov Model(HMM)can model a group of discrete variables based on the temporal-spatial context of variables,HMM is selected in this article is also for this reason.In the anomalous detection stage,rigidity of moving targets is obtained by using optical flow.And then rigidity of abnormal target is exploited to obtain the pre-detection results,based on this,HMM is used to build a temporal context anomaly detection model and the optimal hidden sequence which is the final result of detection is obtained by using Viterbi algorithm.Positions of abnormal objects are obtained at the same time.In the stage of anomalous classification,a SVM classifier combined with the Radon feature of abnormal targets.We use the classifier to obtain the pre-classification result,and another HMM is used to build a model of classification.The final result is achieved through decoding.In order to verify the effectiveness of our algorithm,we experiment on UCSD PED2 and UMN.And the experiment is conducted in two stages,the first is to detect the abnormal target and abnormal group behavior,the second is to classify the abnormal target which is detected accurately.The experimental results show that our algorithm can detect and locate anomalous accurately and classify the abnormal targets effectively.