Research Of Observation Modeling Based On Online Single Target Visual Tracking

Visual object tracking is a key problem in computer vision,which has wide range of appli-cations in national defense,Visual navigation,Video surveillance and communication,human computer interaction,medical diagnosis,intelligent traffic controlling,etc.The goal of study visual tracking is to enable the computer to imitate motion perceptions of human vision func-tionality,that perceive the moving target in a video sequences,and provide an important feature sources for higher levels visual analysis.Designing a robust visual object tracking algorithm become very challenging problem due to the real environment conditions with intrinsic or ex-trinsic factors.Up to now,many enthusiastic scholars study visual object tracking at some ex-tent,and its theoretical framework is becoming more and more perfect.Although it has the ripe theoretical period,but there still exist many immature issues in its real environment applications.There for,Researching visual tracking not only has important theoretical significance,but also has wide application prospects and high practical value.This thesis mainly focuses on studying linear regression methods to develop an effective observation models for robust online visual tracking algorithm.The contributions of this thesis are as follows:First,this thesis aims at studying the existing visual tracking techniques,and presents a novel multi-task least-soft-threshold-squares regression based visual tracking algorithm(MLST).This algorithm aims to incrementally learn a feature subspace from the on online data,and represents the appearance variations of moving objects by linearly combining the subspace features with some additive i.i.d distributed Gausian-Laplassian noise.Its observation model designed as a multi-task joint-learning framework with a reliable maximum likelihood estima-tion function,that encourage the tracker has the great ability to detect the target quickly and accurately.In addition,the objective function optimization also uses the strong theoretical method named "Alternating Direction Method of Multipliers",and making the regression model achieves the good convergence to improve the representation accuracy and the robust-ness of tracking algorithm.Second,based on the MLST algorithm,this thesis also presents a robust visual tracking algorithm based on online integrative-template based model representation(OIMRT).The pro-posed algorithm is able to construct a representative feature template by using the online sub-space learning method with a negative feature extension mechanism.In its observation model,a suitable error estimation function is designed to effectively cope with the appearance changes caused by the abnormal noise such as occlusion or strong illumination.In addition,a novel observation likelihood measurement function is designed in OIMRT algorithm,that measures the likelihood with the standard of both target&bacgound authentication level.Third,in this thesis,a novel observation-modeling framework based on two-dimensional data prototypes is designed,and a matrix low-rank representation based robust visual object tracking algorithm(MLRT)is proposed.Its observation model treats each candidate samples as a two-dimensional matrix rather than one-dimensional vector,and the reconstruction error is assumed to be an additive small block noises with low rank constraints and outliers with sparse constraints.Meanwhile,a new observation likelihood function,which is based on the two-di-mensional data prototypes is proposed.The likelihood function not only handles the partial occlusion effectively,but also encourages the target well-align to the subspace features.The experimental results show that,the observation model is able to maintain the original structures of feature data,and reflects the excellent property to construct the exact candidate sample sig-nals and its errors.Compared to the traditional vector mapping method,it simplifies the data representation,so as to reduce the computational complexity and resource consumption of high dimensional feature mapping.Fourth,motivated by the success of MLST algorithm's efficient multi-task appearance learning advantages and MLRT algorithm's suitable model representation framework,a tensor nuclear norm regression based visual object tracking algorithm(TNRT)is proposed.In this algorithm,the multi-linear data analysis method is applied to the model representation,and maintains the original data prototypes by using tensor data structure.Thus,a structured data representation strategy gives full play to the superiority of the multi task joint learning frame-work,and provides more simple and compact modeling strategy to improve the efficiency and simulation ability of tracking algorithm.Finally,this thesis summarizes all the proposed tracking algorithms from the perspectives of the observation modeling,the minimum reconstruction error estimation,the maximum like-lihood measure,the template learning with benign updates,and so on.This thesis also uses an effective iteration algorithm to demonstrate all tracking algorithms.These algorithms are rea-sonable and have the good convergence property.this thesis also adopts numerous challenging public video sequences to synthetically evaluate all tracking algorithms proposed in this thesis,and compared with 6 classic and state-of-the-art tracking algorithms.