Crowd Dynamic Motion Understanding In Complex Scene Based On Deep Learning

With the continuous development of our social economy,the number and scale of social activities are increasing rapidly.In order to reduce the heavy manual audit work,automatic analysis and understanding of video semantic information has become a subject of great social significance.Video monitoring coverage gradually expands,which brings data basis for video monitoring understanding.Monitoring video not only includes visual features,but also contains complex and diverse crowd movements,which usually provides more important supplementary information for semantic understanding of crowd monitoring video.Therefore,the modeling method with more comprehensive scene generalization ability and higher semantic level for crowd dynamics has become the focus of crowd analysis algorithm.This paper researched the now existing anomaly detection algorithms and abnormal events detection datasets,and then mainly introduced anomaly detection method based on multiinstance learning.In this crowd anomaly detection algorithm,only video visual characteristics of the crowd are considered but crowd dynamic characteristics are not considered.Considering this disadvantage and inspired by attention model,this paper proposed an anomaly detection method based on attention module and multimodal fusion.Social force model models crowd behavior from two aspects: external environment and self-drive goal,which has higher level semantic information.First,through the trained 3D convolution network structure,the original video and social force video of the crowd were extracted into dynamic information features.Then put these features into neural network with attention module and optimized multi-instance ranking loss epoch by epoch.At last,the anomaly detection model predicts anomaly events' scores higher than normal events',in order to get the goal of detecting anomaly.We had experiments on public video dataset to prove the effectiveness of our proposed method.Secondly,we studied the dynamic modeling methods in this paper,and divided them into several types according to different crowd dynamic modeling methods and dynamic information extracting networks,and we also compares the advantages and disadvantages of different methods to determine our own research direction.This paper starts from two aspects: establishing a more comprehensive population dynamic descriptor and constructing a deep network structure which more capable of extracting dynamic information,and then proposed crowd dynamic attribute package.The package contains curl descriptor and divergence descriptor calculating from optical flow,and collectiveness,stability,conflictness descriptors which can describe inter-and intra-motion of groups.Crowd dynamic attribute package can depict crowd motion in complex scene more comprehensively because the package contains different levels of crowd dynamic.This article proposed two branch network for classification,using the video frame images and the corresponding dynamic package as input to recognize crowd behavior.The experiment performance on the public data set shows the effectiveness of our proposed algorithm.In conclusion,this paper investigated methods of popular dynamic descriptors and network structures,and proposed crowd behavior recognition algorithm based on the two stream network structure and crowd dynamic attribute package and crowd anomaly detection algorithm based on attention module and multimodal fusion.