Video Monitoring And Downlink Anomaly Detection Based On Deep Learning

Pedestrian behavior anomaly detection is both an important topic in the field of computer vision and a crucial technical issue for promoting traditional video surveillance to intelligent monitoring.Although many scholars have made a great contribution to the research of pedestrian behavior anomaly and also acquired certain achievements,the difficulty of information extraction on pedestrian behavior in complex environments makes it slow to study pedestrians' behavior.To this end,this thesis studies an improved ViBe algorithm to solve the problem of pedestrian foreground's extraction which includes shadows,ghosts and spots.Furthermore,on the basis of intensively studying the structure characteristics and functional design of the basic convolution neural network,two improved convolutional neural networks suitable for pedestrian behavior detection in 2D and 3D environments are discussed.The main works and acquired achievements are summarized as follows:A.Aiming at the problem that the ViBe algorithm is sensitive to complex environments with the aspect of pedestrian foreground extraction,an improved ViBe algorithm is proposed.In the design of it,an adaptive threshold radius update strategy is designed by weighting background's gradient and mean square deviation,while a shadow detection rule is developed based on the pedestrian's foreground and the shadow's values of brightness,saturation and chromaticity in the HSV color space.On the other hand,a ghost elimination rule is constructed in terms of the background secondary update and foreground-neighborhood histogram.Comparative experiments show that the proposed algorithm is robust and can effectively obtain a relatively complete pedestrian prospect with computationally low cost.B.On the basis of exhaustively analyzing the principle,structure and functional design of the basic 2D convolutional neural network,an improved 2D neural network is established to deal with the problem of pedestrian behavior detection,by using the above improved ViBe algorithm,a Dropout mechanism and a data enhancement strategy.With the help of an open source data set,comparative experiments show that such an improved neural network is superior to the traditional convolution neural network with the aspect of recognition precision on pedestrian behavior detection while its training process is more stable.On the basis of intensively discussing the principle,structure and downsampling of the basic 3Dconvolutional neural network,an improved 3D neural network is developed to carry out the task of pedestrian behavior detection.In the design of it,a strategy of multi-feature fusion and the above improved ViBe algorithm are used to improve the quality of its input and produce valuable multiple-dimensional features,respectively;image's gradient and optical flow at pixels are also introduced to enhance recognition precision on pedestrian behavior.Experimental results show that the improved neural network has a significant improvement by comparison against the basic2 D and 3D convolutional neural networks while its training process is relatively stable.