Research On Object Tracking Algorithm In Complex Environment

As an important research field in computer vision,object tracking can provide not only the motion state and path of object,but also the vital data for high-level vision analysis,such as trend prediction,behavior analysis,scene understanding,etc,which is widely used in the fields of intelligent surveillance,human computer interaction,visual navigation,medical treatment,and so on.Although there have plenty of research achievements with object tracking algorithm in recent years,due to the influence of complex factors,which includes background interference,illumination variation,scale variation,rapid movement,rotation and deformation,motion blur,heavy occlusion,and others,the apparent model changes dynamically,it is still an important and challenging task to study effective object tracking algorithm to adapt to the complex environment.The paper gives serious analysis together with summary of existing object tracking technology,and then,according to the different model category framework,it presents the object appearance description as well as motion search strategy,and further provides focused research on the observation model buliding together with model update criterion design.Based on the sparse dictionary optimization,object motion continuity,classifier online learning and object detection theory,the paper proposes the novel object tracking algorithms and performs favorable tracking performance on various typical testing scenarios.The main research work and innovative contributions of this dissertation are as follows.To deal with the tracking drift problems caused by drastic object appearance change in complex scene,in the framework of bayesian theory,an object tracking algorithm combining spatial information and sparse dictionary optimization is proposed.Firstly,with the theory of likelihood estimation,it gets a more sparse cost function.Meanwhile,considering the spatial structure correlation,it intriduces the laplacian regularization,and obtains an optimized objective cost function.Then,the initialized dictionary is acquired with the use of clustering method,and the dictionary optimization is completed with the theory of lagrange dual and accelerate proximal gradient approach.Finally,according to the maximum pooling theory and spatial pyramid method,the coefficient of object template and candidate samples with the reduced dimension and more spatial information is obtained based on the optimized dictionary.The similarity criteria is utilized to get the optimum candidate sample that most similar to the object template,and then the current object state is obtained,which can achieve robust object tracking effectively.Aiming at the problem of fixed template information that can not adapt to the object appearance change,based on the sparse representation theory,the paper designs a novel object tracking algorithm on the basis of spatial structure and motion continuity.First of all,according to the local linear embedding theory,it reconstructs the local image block of each candidate with the aid of neighborhood information,and builds optimized objective cost function with the full use of space structure information.Then,it gets the optimization solution of objective function by solving the least squares problem,and obtains object template cofficient and candidate sample cofficient with more spatial structure information.The calculation of optimal candidate object can be completed with the similarity measure based on the theory of histogram interaction.Finally,motion continuity information is introduced into the template updating phase,togerther with the consideration of similarity between the current object tracking results and history template,the paper presents a new template update scheme about when to manage updating and how to realize the strategy,which enables tackle object appearance change effectively in a variety of complex scene.To make full use of the diversity and discrimination of the sample information,the paper proposes an object tracking algorithm on the basis of multi-feature fusion and classifier online learning.Firstly,it exploits the complementary characteristics to give the object appearance description,and trains the sub-classifier with different apparent feature on the basis of support vector machine principle.Then,according to the theory of maximum a posteriori estimation,it builds the loss function with the logarithmic likelihood as well as conditional entropy,and calculates the reliability of each sub-classifier.At the same time,the optimum object state estimation by means of the weighted fusion prediction results of each sub-classifier is obtained.Finally,the new update scheme of training sample set and classifier is designed,which updates the training sample set coarsely according to the nearest-farthest boundary principle as well as co-training theory,and gets more representative ones with the refined selection criterion.The novel method enables the sub-classifier has more generalization ability and can separate the object from the background effectively.To solve the tracking drift problem caused by the low discrimination of object appearance information in complex environment,based on the framework of correlation filter theory,the paper presents an object tracking algorithm on the basis of objectness detection.Firstly,it trains the initial classifier with the kernelized correlation filter,and obtains the preliminary object prediction state.Then,it generates the original proposal bounding box set with adaptive scale according to the objectness detection principle of proposal bounding box,and further gets optimized ones with the refined selection criterion.By introducing motion continuity,the prediction location and scale based on the proposal bounding box is calculated,and then the final optimum object state estimation is acquired comprehensively.Finally,taking into account the occlusion influence judge of object appearance at current frame,it designs the update strategy of coefficient matrix and the sample object model,which capacitates the classifer to get the stronger generalization ability.The proposed algorithm can separate object from complex background more accurately,and tackle the tracking drift problems caused by the object appearance change in complex scene more effectively.