Research On Sementic Feature Extraction And Behavior Analysis For Crowd Motion

Intelligent surveillance technology can be implemented in the panoramic monitoring for public places, which helps to develop effective and efficient evacuation and flow control strategies to improve the public safety. In recent years, crowd behavior analysis has become one of the most challenging problems in the field of intelligent surveillance due to the difficulties of complex motion environment, object occlusion, etc. and it has drawn considerable attentions.The thesis is focused on semantic analysis for crowd motion and behavior analysis. The main contributions are outlined as follows:(1) A novel method is proposed for improving the traditional variational optical flow method using a data item combining the color invariants conservation and crowd motion gradient conservation assumptions. The method is effective to overcome the detection errors caused by illumination change and local exposure. Moreover, an adaptive anisotropic regularization term is used to solve the problem of lossing important details caused by the over smoothing in the existing optical flow method. The experimental results show that the optical flow can be estimated accurately under the circumstances of illumination change and the performances are much better than the similar methods.(2) An approach for feature dimensionality reduction based on LDA model is proposed to solve the curse of dimensionality in visual feature of crowd motion. Firstly, optical flow algorithm is used to estimate the motion information; Then,the motion feature vector can be defined with bag of visual words. Finally, Semantic feature of crowd behavior is represented using LDA model, which succeeds to reduce feature dimensionality of visual features.(3) A particle flow based bag of visual words model is proposed to solve the problem for lacking spatial-temporal context information of existing bag of visual words model. And, it’s an effective realization for the crowd behavior abstracted in advanced semantic level with features of bag of words. The key ingredient of model is that, feature variation for crowd motion information in spatial-temporal scale can be defined as propagation relationship of particles between the flow fields. The experiment results for this method show that recognition precision for crowd behavior is improved significantly, because the crowd behavior topics are represented with the visual words contain spatial-temporal structure.(4) A Category Guidance-Correlated Topic Model is proposed to solve the semantic ambiguity issue of latent behavior topics representation because of inappropriate topic number. And, it has the capability of modeling the correlated crowd motion pattern. Semantic features of crowd behavior can be extracted under this weakly supervised learning way. This kind of feature is possible to achieve the effective clustering of the same category of crowd behavior, leaving the behavior of different categories have a greater degree of differentiation.Experiments of crowd behavior recognition are carried on the bench data from PETS database with one-versus-one SVM classifier. Experimental results show that this method using CG-CTM model and Bo VW based on partial flow can achieve high accuracy to recognize crowd behaviors.