Research And Realization Of Intelligent Video Surveillance Of Groups Of Abnormal Event Detection Method

With the social and economic development, especially in China, the acceleration of urbanization, increasing urban population density, the city's public spaces are often crowed with large amount of people. And group accidents often occur in these crowed regions. If these accidents can not be detected in real time, they often lead to very serious consequences. Abnormal crowd behavior detection deserves more and more attention, because it is an urgent need by reality. Abnormal crowd behavior detection needs to find the hidden and specific information from the video to characterize the abnormal event. And then a model is been obtained using these information for training. When the abnormal event happens next time, this model can detect it and immediately alerts people for discovering and handling the anomaly, avoiding further expansion. Researching in this area relieve people from the tedious work, which requires staring at the monitor screen all day. It has broad application prospects. Therefore, abnormal crowd behavior detection has become a focus of current research.The system of abnormal crowd behavior detection is often applied to monitoring the abnormal behavior in densely populated areas in the medium cities. And omissions and misstatements may bring great distress and losses. The existing algorithms of abnormal crowd behavior detection have shortcomings in detection rate and time efficiency and can not be applied in practical applications. Therefore, it is necessary to conduct more in-depth research about crowd behavior, explore new ideas, propose new algorithms to improve accuracy and time efficiency, reduce the false negative rate and false positive rate. Then the abnormal crowd behavior in congested scene could be detected more quickly and efficiently. We could get more valuable information.In this thesis, we start our research from the abnormal crowd behaviors which are easy to cause serious consequences, such as gang fights, panic, riots, etc.. Through researching the various anomaly detection algorithms at home and aborad, expecially in the detection rate and time effciency, we firstly propose the size-adapted spatio- temporal(SAST) cuboid and high-frequency and spatio-temporal(HFST) cuboid for the deficiency of existing algorithms. Based on these two characteristics, we conduct in-depth study of the abnormal crowd behavior detection scheme, and propose a new anomaly detection algorithm and solutions. We have achieved the following results:1)The algorithm of SAST can select the best size and location of spatio-temporal cuboid automatically, based on the information contained in the video sequence, and extract it. Its size, location and number are determined by the video sequence. And based on this, we redefine the rules of similarity judgments between spatio-temporal cuboids which have different scales. We also propose a complete anomaly detection scheme combining the SAST and LDA model. The accuracy of anomaly detection has been improved.2)We propose the concept of high-frequency spatio-temporal cuboid, and propose the algorithm for its calculation. The traditional HMM is modified in this thesis. We propose a complete anomaly detection scheme which is used to key areas for local monitoring combining the HFST with HMMs which has a competitive mechanism. Our method improve the detection rate of abnormal events in local area greatly; We also propose a complete anomaly detection scheme which is used to global zone combining the HFST and LDA model. Without reducing the detection rate, our method reduces the time complexity greatly compared with traditional method. Our anomaly detection scheme can be used to practical application. And experimental result shows that our algorithm has very good detection rate and time efficiency.