Pedestrian Behavior Recognition Methods In Surveillance Video Big Data For Public Security

As the main body of social activities in urban public places,the public safety of pedestrian has become a critical issue in the management of social security and the construction of intelligent security since the frequent occurrence of public emergencies in recent years such as terrorist violent attacks,demonstrations and stampedes.At present,video surveillance is well developed and is a core part of the security monitoring system in China.High-resolution cameras have been widely distributed in most areas of cities,such as squares,stations,shopping malls,capturing a large number of monitoring data.How to analyze the state change of crowd behavior and to discover clues of the mass emergencies from these monitoring video data have increasingly become sharp scientific problem in the fields of public security.The problem of crowd behavior recognition has been paid more and more attentions by researchers,technology companies and other research institutions at home and abroad.However,existing studies are on a basis of the image features and classification methods without utilizing comprehensive knowledge.As a result,the waste of resources is arising due to the lack of the whole knowledge organization and the redundancy of the sub algorithm during the application of their algorithms in the analysis of the monitoring video big data.As crowed behavior is characterized by the state change of the pedestrian attributes in videos,this dissertation studies the skeleton information,the counting information and the behavior information of pedestrian,and verifies the validity of the model and the algorithm in terms of instance validations and system implementations.The main works of this dissertation are as follows.(1)A framework for mining pedestrian behavior information based on knowledge is proposed to deal with pedestrian behavior recognition in surveillance video big data.This framework is on a basis of system science theory and the basic knowledge element model,and implements the management and organization of prior knowledge involved in the processing of surveillance video.Considering image element of pedestrian as a type of attributes of pedestrian knowledge element,the mining process of some attributes of image element is guided by the mapping relationship between attributes,realizing the identification of the type of crowd behavior based on the prior knowledge of the attributes of image element.With the guidance of cognitive knowledge,we flexibly implement the integration of algorithms to meet various demands of pedestrian analysis under multi-camera in different surveillance scenes.(2)The skeleton attribute mining algorithm for single person image element and the individual behavior recognition method based on skeleton information are studied to cope with the problem of skeleton-based individual behavior recognition.As a basic algorithm of individual behavior recognition in videos,skeletonizing algorithm needs to have a fast processing speed under the premise of ensuring the topological structure.In this dissertation,the key nodes of shape polygons are obtained according to iterative process of discrete curve evolution.Then we truncate shape polygons by parallel line clusters and mark the midpoints of truncated segments as skeleton points.Finally,a graph structure of pedestrian skeleton is determined by utilizing the correlation among skeleton points.The experimental results on the recorded video set verify the effectiveness and rapidity of the proposed skeletonization algorithm.Furthermore,we summarize some skeleton attributes of different individual behavior types as a prior knowledge,and construct fuzzy logic rules of prior knowledge about individual behavior to identify the individual behavior in video.(3)People counting problem and knowledge-based crowd behavior are studied to meet the requirement of crowd behavior control under the surveillance cameras.Due to the various perspective-related distortions in video cameras,regression approaches of people counting may give different counting results for the same people group in different frames.In the dissertation,we propose a digraph model to represent the relationship of different groups of moving people among frames.Based on the counting results for each group resulting from regression-based counting algorithms,a quadratic programming method,characterized with network flow constraints is proposed to improve the performance of the proposed method.In order to further improve the accuracy and processing speed of the algorithm,the histogram of block feature and the network contraction method are proposed.The experimental results based on the standard dataset show that the network constraints always smooth out oscillations in the curves of original person counting.Furthermore,the prior knowledge we summarized according to the human experience in the analysis of abnormal crowd behavior is regularized based on the fuzzy rules.A crowd behavior recognition method based on the prior knowledge is proposed,which can flexibly append the scene knowledge and the crowd prior knowledge.It helps identify the abnormal crowd behavior accurately.(4)A distributed processing algorithm on MapReduce framework is proposed to deal with information mining problem in massive video data.We take the pedestrian counting algorithm under multiple cameras as an example.The distributed algorithm of pedestrian detection is used to collect the training set for the regression function.Then we storage the instantiation files of the knowledge element of group relations in the distributed system and implement the people counting algorithm.Finally,some relevant systems are shown under the guideline of the proposed theoretical model and algorithms.This dissertation belongs to interdisciplinary study involving social public safety,emergency management,big data mining and artificial intelligence,and it has certain theoretical and practical significance for realizing crowd behavior analysis under monitoring video big data.