Research On Abnormal Behavior Detection Based On Improved ViBe Algorithm

As people pay more and more attention to public safety,the market demand for intelligent video surveillance is also growing.Research on abnormal behavior detection can promote the progress of intelligent video surveillance,which is of great significance for maintaining social stability and improving people’s happiness.The research on abnormal behavior detection includes many aspects,and the object of this thesis is people.The research of human abnormal behavior detection is mainly divided into two parts: moving target detection and human abnormal behavior recognition.The research content of moving target detection is mainly to extract the foreground image of moving target.The better the extraction effect,the more conducive to the subsequent recognition research.The research content of human abnormal behavior recognition is mainly to classify behavior based on the extracted behavior features,and to judge whether the behavior is abnormal according to the classification of abnormal behavior.The selection of algorithm will affect the correct recognition rate of behavior,so it is very important to select the appropriate algorithm.Therefore,this thesis mainly focuses on how to eliminate the influence of interference in the process of moving target detection,extract the ideal human foreground target and select the appropriate algorithm to identify the human abnormal behavior.The research work of this thesis is as follows:In the part of moving target detection,according to the shortcomings of ViBe algorithm,it is improved.Firstly,shadows are detected according to the chromacity method,and then the detected shadows are removed by adding the difference from the average background intensity on the current pixel.The background extracted by the mean value method is used to establish the initial background model to eliminate the ghost phenomenon caused by the presence of moving objects in the first frame image during the target detection process.In the initial stage of moving target detection,by reducing the update factor,the update speed of the background model is accelerated,and the problem of unclean background extracted by the mean value method when there are too many moving objects is quickly eliminated.The fixed threshold is replaced by the adaptive threshold,and the color distortion value is introduced,and the classification of pixels is determined by the color distortion value and the adaptive threshold.A small area connected domain removal operation is performed on the foreground target to reduce the interference of noise and dynamic background.Finally,the closed operation is carried out to improve the foreground target.The experiment is verified on the CDnet 2014 dataset.The experimental results show that the improved ViBe algorithm can effectively remove shadows and ghosts,improve the accuracy of pixel classification,reduce the interference of dynamic background and noise,and the detection effect is ideal.In the part of human abnormal behavior recognition,firstly,based on PCA,features are extracted for the foreground target extracted by the improved ViBe algorithm.The performance of the three classification algorithm models is evaluated through Shuffle Split cross validation,and the BP neural network model with the best classification effect is selected.The extracted features are sent into the trained BP neural network classification model to classify human behavior.According to the classification of abnormal behavior,it can judge whether the recognized behavior is normal behavior or abnormal behavior.The Experiment is verified on video datasets containing abnormal and normal behaviors constructed on the basis of Weizmann dataset,KTH dataset and Multicam fall dataset.The above experiment shows that the classification performance of the algorithm in this thesis is relatively good,and it can judge whether the recognized behavior is normal behavior or abnormal behavior according to the classification of abnormal behavior,so as to achieve the purpose of detecting abnormal behavior.