Visual Tracking Based On Multi Discriminative Dictionary Learning

Visual tracking is a research hotspot in computer vision.It integrates the theory and technology of machine learning and pattern recognition,and has a broad application prospect in video monitoring,intelligent transportation and modern military fields.Visual tracking scenes often contain a variety of tracking difficulties,including video background clutter,target appearance changes,scale changes,illumination variation and partial occlusion,which will seriously interfere with the performance of target detection and recognition,making the tracking problem more complex.Therefore,it is a very challenging problem to track the target in the complex and changeable dynamic tracking scene.Based on the theory of sparse representation target tracking method,this dissertation mainly studies the target tracking framework based on sparse representation and dictionary learning,and proposes corresponding methods from the perspective of multi features,multi parts and model updating.The main research contents and contributions of this dissertation are as follows:1)Aiming at the problem of poor tracking performance of global model when the object's appearance changes and occlusion,a method of target tracking based on local weighted sparse representation is proposed.In this method,the target is divided into multiple local blocks,and the decision information of multiple local discriminant dictionary models is integrated to select the target.Firstly,sparse features are used to measure the similarity of samples and select candidate samples.Then the weighted reconstruction error is constructed by using the local weight of the sample and the reconstruction error information,and then the best sample is selected.In order to reduce the introduction of error information,sparse coefficient analysis is used to select the period with less interference intensity to update the model.Experimental results show that this method has good tracking performance in the task of target occlusion and target appearance change.2)Aiming at the problems of high dependence of single feature representation model on feature representation ability and poor scene adaptability,this dissertation proposes a target tracking method based on multi feature weight discrimination dictionary.This method constructs a dictionary model with weights on the basis of discriminative dictionary learning.In tracking,the dictionary weight information is used to measure and select candidate samples,and then the reconstruction error of candidate samples is compared to get the final tracking target position.In order to improve the adaptability of the model,the method of sparse coefficient statistical analysis is used to detect the target interference and update the model when strong interference begins.Experiments on open datasets show that the method has good performance in the case of blur of target motion,interference of background change and occlusion of target.To sum up,in view of the difficult problems in visual tracking,this dissertation starts with sparse representation and dictionary learning,and studies the appearance representation model,multi feature fusion and dictionary update strategy of the tracking target,so as to improve the robustness and accuracy of visual tracking.The effectiveness of this work is demonstrated by the comparative experiments and Analysis on multiple datasets.