| Group sudden aggregation events have the characteristics of suddenness and destructiveness,from the response of events,it is very important to quickly find and identify the precursor characteristics of group sudden aggregation events.At present,the identification,monitoring and early warning of group sudden aggregation events mainly focus on the detection of group abnormal behavior,there is a lack of analysis of the characteristics of group sudden aggregation event scene correlation,the behavior pattern of the population in the event is not clear and so on.In view of these shortcomings,this paper puts forward a combination of scene characteristics and group abnormal behavior detection of the group sudden aggregation event identification and early warning method.First of all,139 cases of sudden cluster events of groups were collected in the past ten years,the key feature elements such as the location,time of occurrence,number of people involved and induced factors were extracted,and the key feature elements of the case were analyzed by Apriori algorithm.The analysis found that there is a strong correlation between the occurrence location of group sudden aggregation events and the time of occurrence and the induced factors,and based on this,the knowledge factor identification tree of group sudden aggregation events is established.Then,the corresponding volunteers were designed and organized to carry out the simulation of the abnormal behavior of the sudden aggregation events of the four groups,and used the high-definition video capture equipment to collect data,and the characteristic extraction method based on YOLOv3 and Deep-Sort algorithm to extract the movement characteristics of the population and analyze them.The results showed that the population in the abnormal behavior of fixed-point gathering showed a relatively stable disorder state,the population in the abnormal behavior of the parade showed a relatively stable and orderly state,the population in the abnormal behavior of violent conflict showed a disordered state,the movement speed curve appeared several obvious peaks,the population in the abnormal behavior of the killing showed the orderly state with a large overall speed,and the population in normal behavior showed a disordered state with a small overall speed.Subsequently,according to the characteristics of group abnormal behavior movement,based on fixed-point gathering,procession,violent conflict,killing as a negative sample,with normal behavior as a positive sample,a two-classification model based on LSTM neural network was established,and a multi-classification model based on LSTM neural network was established,and a multi-classification model based on LSTM neural network was used to evaluate the model.The results showed that the AUC curve of fixed-point gathering,procession,violent conflict and beheading behavior in the two classification models was0.97,0.90,0.84,0.95,and the AUC of fixed-point gathering,procession,violent conflict,beheading and normal subclass ROC in the multi-classification model was0.98,0.97,0.82,0.96,0.98.Finally,the core function of the group sudden aggregation event identification and early warning system is realized by using python’s streamlit library.The results of this paper will help video surveillance big data analysis and predict and warn of sudden group gathering events. |
