The Research Of Key Problems In Intelligent Video Surveillance System

Intelligent video surveillance system is used widely in the fields of civilian and military at home and abroad.Without manual supervision,the system will take preprocess for surveillance video day and night,and analyze the recognized interesing targets,which will be automatically detected and tracked continually.After the system analyzes the identities and behaviors of these targets,their description will be given to recognize the whole scene behavior states further.It will send alerts and records information according to the preordained solutions,when the abnormal accidents happen.The practical applications in public security prove that the intelligent video surveillance system is taking an important role in detecting and providing the intuitive ways to criminal investigation.A smart application management platform and system of public security video surveillance based on intelligent video surveillance technology is proposed and implemented in the paper.It stores,manages and dispatchs the surveillance video resourses uniformly,providing the smart application services.This paper makes in-depth investigations and researches on three aspects: abnormal behavior detection,text detection and recognition and human activity recognition.And specifically,the main contributions of this paper are listed as follows.1.In the aspect of the anomalous behavior detection,this paper introduces an abnormal behavior detection algorithm based on Hidden Conditional Random Fields(HCRF).Firstly,in the stage of behavior feature representation,for the relationship of the patches' motion labels in time dimension,servial low-dimensional descriptors(PR,AIMB,IMB,COR,IIDMB)are proposed to describe behavior features.Secondly,in the process of video anomaly detection,the hidden variables structure of HCRF is used to overcome the limitations of the generative models that observations are assumed to be independent given the values of the latent variables.The model of HCRF could extract the behavior features in spatio-temporal neighborhood and fill the gap between low level and high level features well.2.In order to overcome the questions of high redundant information and high-dimensional data represention for surveillance video,this paper introduces group sparse coding method to detect abnormal events in the video.Firstly,histogram of optical flow is used to extract motion features of the video by different kinds of spatio-temporal distribution structure,showing the statistical information of the behavior in the video.Secondly,the abnormal event detection model based on group sparse coding is introduced.It learns the fully dictionary from normal activity set by using K-SVD algorithm.The coefficient of group sparse reconstruction is calculated based on regularized group orthogonal matching method.Last,the minimized group sparse reconstruction cost algorithm based on 2,0l norm is taken to compute the error of the group sparse's reconstruction for test video sample,and then determine whether an event is normal or not.3.For the application demand of text recognition in the complex natural scene,this paper studies a realtime framework and key technology of text detection and classification.For capturing the areas containing text information,a fast algorithm of text area detection is firstly introduced based on the features of text edges.The initial areas of character contour are extracted by morphology methods.The character rules are used to filter the real areas for capturing words precisely and quickly.The text classification model based on shape context could make robust description for the shape of character.The KNN classifier is trained from samples to ensure high performance of the text tecognition model.4.In the aspect of reflecting the major features of human behavior recognition,this paper introduces an algorithm based on nonlinear scale space that could denoise well without losing the details of target's edge.In the space dimension,the continuously nonlinear scale space is constructed by the nonlinear filter.In the time dimension,1-D gabor filter is used.The detection fuction of time-space interesting points based on these filters could extract these points precisely.The surround suppression theory is used to filter the backgound interesting points after being extracted which interfere the human behavior classification,and improves the efficiency of detection.Fisher coding is treated as middle semantic representation of the time-space interesting points' feature.All of these local representation are merged into a global description.Finally,the SVM classifier based on linear kernel is used to detect the human activities.For key problems in the intelligent video surveillance system,thress aspects are proposed in detail in this paper: abnormal behavior detection,text detection and recognition in the complex scene,and human activity recognition.The experimental results and applications show that our abnormal event detection algorithms based on HCRF and group sparse coding could reduce the redundant information efficiently and have good performance on anomalous behavior detection.The text recognition method based on shape context could improve the accuracy in the complex natural scene.The human behavior recognition algorithm based on nonlinear scale space could achieve a high recognition accuracy.