Abnormal Crowd Detection Based On Mult-Scale Recurrent Neural Network

With the increase of population and activity diversity, the crowd phenomenon is becoming more and more common. It would cause significant loss once accidents happen in the crowd while the loss could be greatly reduced if the widely distributed real-time monitoring equipments can be used to catch abnormal situations. However, the great variations of crowd density and crowd dynamics, as well as the existence of many shelters in scenes, make the traditional video monitoring technology cannot be directly applied to the crowd scenes, which makes the abnormal crowd event detection are still challenging problem and hot topic of the crowd scene analysis. Therefore, the study of abnormal crowd event detection method has great research value.Based on spatial-temporal modeling of the crowd scene, this thesis proposes an abnormal crowd event detection and localization approach based on multi-scale recurrent neural network. Upon the proposed method, this thesis proposes a system construction scheme of anomaly crowd detection and localization based on multi-scale recurrent neural network. The main research content includes the following aspects:(1)This thesis uses the recurrent neural network to model the video sequence, to find out the spatial relation and temporal relation. Firstly, the crowd scenes are split into grids and presented using multi-scale histogram of optical flow (MHOF). Then, different grids are connected to obtain a global time series model of the crowd scene. Finally, a recurrent neural network is devised to detect and locate the abnormal event on the time series model of the crowd scene. In the recurrent neural network, the hidden layers are used to model the spatial relation, and the feedback loops are used to catch the temporal relation. And the output layers are used to determine whether grids are abnormal.(2) According to spatial locality, in order to improve the recurrent neural network, this thesis proposes a multi-scale recurrent neural network to better find abnormal events. The multi-scale hidden layers are used to model the spatial relation among different scale neighbors.(3)Compared with several typical methods, the effectiveness of the presented approach is verified, and this thesis analyzes the reason that this method is better than typical methods and that multi-scale recurrent neural network is better than recurrent neural network.(4)Upon the proposed method, a system construction scheme of anomaly crowd detection and localization based on multi-scale recurrent neural network is proposed. The system architecture, core function and the advantages of the system are introduced.