Research On High Order Robust Models For Video Foreground/Background Separation

The foreground and background separation of video have always been one of the hot research topics.This task is widely used not only in a series of practical tasks such as monitoring,tracking,and early warning,but also serves as the basis for different kinds of high-level video processing tasks such as semantic understanding and driverless driving.How to effectively distinguish the two based on the different data distributions and structural characteristics of the video foreground and background is the core of the task.In this thesis,based on analysis of the high-order and high-dimension structural characteristics of video data,as well as aiming at the problem that existing methods are very sensitive to noise and dynamic background detail changes in video background separation task,some vector represented models(quantization clustering etc.)are firstly extended into the tensor space to better conform to the characteristics of video data.In addition to the quantization part,low-rank representation and sparse representation are employed separately,which result in two hierarchical background models.Then,joint optimization models are built up by further integrating some proposed foreground priors.Verified by a series of numerical experiments,the efficiencies of the models are convincing.The proposed new algorithms are as follows:1)Targeting at the problem that all the present algorithms have ignored the elementwise relationship in the natural high order structure of the data,the proposed quantization clustering algorithm,that is based on high order representation and data aggregation degree,has produced more effective codebooks that are robust enough to the noise.Based on the structural distribution characteristics of high-order data and highdimensional data,the high-order product quantization and distribution sensitive product quantization are proposed as to achieve fast and effective distance approximation and cluster analysis in tensor space.For high-order data,the regionalized element correlation relationship is used to guide the space segmentation problem in the quantization task,which effectively improves the representative ability of the clustering codebook.For high-dimensional data,the imbalance of local data distribution is fully considered.A series of indicators such as data distribution degree and matching index are used to measure and quantify the imbalance of different local data,and then adjust the focus of the coding task to achieve a more effective clustering codebook.2)Targeting at the problem that most low-rank based models are sensitive to the detail changes in the background and are usually time-consuming,the proposed robust low-rank model,which is based on quantization clustering and multi-resolution analysis,has achieved improving performances when solving the foreground/background separation task.Starting from the data characteristics,multi-layer model is constructed to describe different components of the video: the derived rank-1 model formed by the combination of the main value tensor and the changing tendency matrix is used to describe the stable averaging background;based on the fact that quantization clustering codebook is robust to data noise,a detailed background layer model is proposed;based on different distributions of the noise from the videos under different resolutions,a multi-layer fusion foreground detection model is proposed;numerical experiments on the I2 R and CDnet datasets verify the advantages of the proposed joint model over other algorithms.3)Aiming at the poor update accuracy and poor noise robustness of existing sparse representation models,a robust sparse representation model is proposed based on quantitative clustering,error modeling and image gradient analysis.The model has effective background separation performance and the corresponding algorithm is more efficient.The model captures the main background information by sparse representation and dictionary learning,which is based on the observation that sparse representation model can record much more background status than the low rank one;further introduces the spatial correlation of background elements into the detailed background layer which is modeled based on quantization;fits the distribution of the video noise by mixture of Gaussian,that to reduce the influence of noise on the video processing task;a more accurate comprehensive foreground model is proposed,by considering both the smoothness of the foreground area and the continuity of the foreground objects;finally,the experiments on the I2 R and CDnet partial datasets verify the validity of the proposed joint model.In summary,this paper employs a variety of machine learning models and algorithms to solve different sub-problems in video foreground and background separation task.The proposed detail background layer improves the robustness of the foreground and background separation models to the dynamical variations in the video background.A large number of experiments show that the proposed models work more effectively on the high-order and high-dimensional video data when compared with various existing models.