Deep Kernel Nonlinear Network In Video Behavior Analysis

With the development of computer storage capabilities and the popularity of mobile smart terminals,video data has gradually become the main data form that people consuming and producing.Faced with a large amount of video data,it is particularly important to build an intelligent video analysis system.Video behavior analysis is an indispensable and important function of the intelligent video analysis system.It is mainly to detect,extract,analyze and identify human behavior in video.In recent years,deep learning plays a paramount role in video behavior recognition.However,the un-explainability of “black box” makes researchers try to expand the traditional machine learning methods with clear principles into deep structures for video behavior analysis,and better results are achieved.As a classical and effective machine learning method,the kernel method has also received the attention of researchers.This thesis studies the basic theory of deep kernel nonlinear network in video behavior analysis.The main research work is as follows:1.Verification of behavior recognition algorithm based on improved dense trajectory is performed.HOG,HOF,and MBH feature descriptors are used to express dense trajectories.Descriptors are coded using BoVW model,and behavior classification results are verified on the Weizimann standard motion database.2.Proposition of a kernel-based nonlinear representor coding method based on the kernel-based nonlinear representor.The method breaks the cohesion of the kernel function and maps the features explicitly into the high-dimensional feature space.And it is applied to video dense trajectory feature coding,which effectively improves the classification accuracy of the original word bag model.3.Video behavior representaiotn from the perspective of video block generation,usin the idea of deep structure in deep learning.It is proposed that the kernel-based nonlinear representor coding and the Fisher vector-based coding are stacked to form a deep kernel nonlinear network model.Experiments show that the model increases the ability to express video behavior and improves the recognition rate of video behavior.This thesis clarifies the current research status and technical difficulties of video behavior analysis.Kernel method is introduced in video behavior analysis in detail.At the same time,based on kernel-based nonlinear representor coding and Fisher vector,a deep kernel nonlinear network model for the expression of video behavior is proposed.Experimental results on Weizimann show that the proposed model outperforms the traditional coding model.