Crowd Behavior Analysis In Complex Scenes

In recent years,the public security accidents caused by crowd phenomenon have received sufficient attention.To prevent the accidents,it is crucial to analyze the human behavior in crowd scenes.Consequently,crowd behavior analysis becomes increasingly important in intelligent video surveillance,and has been widely used real-world applications,such as scene segmentation and crowd tracking.The objective of crowd behavior analysis is to capture the motion patterns of crowds.In complex scenes,pedestrians tend to from groups with the surroundings and show collective behaviors.Each group corresponds to a specific motion pattern.Groups are more stable than the pedestrians,and provide semantic information about the crowds.Therefore,the investigation of group behavior has became the main studying content of crowd behavior analysis,and is involved in many research topics,such as anomaly detection,crowd behavior prediction.However,the complex property of crowds brings three major difficulties in group behavior analysis.First,the pedestrians in the same group may come from a manifold structure,and show difference behaviors.Second,due to the serious occlusion,it is impractical to extract the pedestrians directly.As an alternative approach,most of the existing works take feature points or particles as study objects.However,the points and particles are too microscopic to represent the individuals.Thirdly,since the crowd density may be extremely high,most image features,such as color and texture,cannot be used for analyze the group behavior.To tackle the above problems,this thesis introduces a series of methods,which can be summarized as follows.1.Anchor-based crowd behavior analysis.There are always some representative individuals in crowd scenes,which reflect the motion patterns of the corresponding groups.Inspired by this observation,we propose to locate the anchor feature points according to the local interaction.After obtaining the anchors,the topological relationship between the points is exploited with the manifold ranking algorithm.Then,based on the topological relationship with the anchors,the points are divided into groups.Furthermore,a coherent merging strategy in designed to combine the groups with similar motion patterns.In this way,the group behaviors are finally recognized.2.Patch-based topic model for crowd behavior analysis.Since the feature points and particles can just reflect the local crowd behaivor,we propose to segment the crowd images into patches,and consider the patches as study objects.Compared with the local points,the patches are more stable to represent the individuals.To further deduce the latent semantic information,the markov field is combined into the latent dirichlet allocation model.Paches with the same topic are combined.In addition,the intra-class distance is employed to evaluate the quality of the groups and select the most appropriate groups.3.Multiview parameter free crowd behavior analysis.To mitigate the affects of tracking noise,a structural context descriptor is put forward,which is able to profile the structural property of feature points.Besides,a self-weighted multiview clusering algorithm is presented to integrate the similarity graphs from both the motion direction and context views.The proposed clustering method is able to group the feature points without any post-processing.Finally,the tightness-based merging method is used to combine the similar groups to get the final result.The proposed framework does not involve any parameter or threshold,so it can handle the crowds with various densities.4.Patch-based multiview clustering behavior analysis.To fully exploit the crowd features,we propose to construct the similarity graph of feature points from four views,including interaction relationship,spatial location,motion direction and motion transformation.A diversity-regularized multiview clustering method is introduced to learn the optimal graph from the multiple initial graphs.The proposed clustering method emphasizes the complementary information from different views with a diversity regularization term.The final groups are obtained by merging the clustering result according to the spatial location and motion direction.