Research On Robust Target Tracking Method Based On Hard Threshold Sparse Representation

Object tracking technology,which is an important research topic in computer vision technology,has been widely used in video search,military,etc.The main purpose of object tracking is to extract and recognize the target from video sequence,so as to obtain its position,velocity and other information.In recent years,although the object tracking technology has been developed rapidly,due to the existence of all kinds of interference factors in practical application,such as changing illumination,rotation,occlusion,which leads the object appearance to change during the tracking progress,the object tracking technology is still facing with a great challenge on its accuracy and robustn ess.Sparse representation is a new method developed in the field of signal processing in recent years.It has been gradually applied to the field of visual tracking.In particular,the concept of compressive sensing in 2004 has further expanded the application of sparse representation in the field of visual tracking.It performs well in suppressing abnormal interference and is proved successful in occlusion target tracking(e.g.tracking the face occluded by glasses,hats and mask).In this paper,an improved target tracking method based on hard threshold pursuit sparse representation is proposed.Currently,most existing sparse representation methods have high computational complexity and require large storage capacity in robust target tracking.However,hard threshold pursuit converges in a few iteration cycles with high accuracy of reconstruction and robustness.In this paper,the iterative hard threshold tracking algorithm is applied to the sparsity-based collaborative model,which effectively improves the robustness of the collaborative model and makes the tracking faster.This paper presents a collaboration model based on hard threshold pursuit sparse representation.This model is a robust appearance model which considers the global and local representation of the target.It is mainly composed of a discriminant classifier based on sparse representation and a generating model based on sparse representation.In sparse representation discriminant classifier,firstly,according to the foreground and backgro und information of the initial frame,the feature selection method based on sparse representation is used to generate the positive and negative template bases with high discriminant characteristics.For each candidate target,the reconstruction error betwe en the target and the positive or negative template is calculated separately,so that the foreground image can be assigned a greater weight than the background image when calculating the confidence degree,so that the foreground target can be better separa ted from the background.In the sparse representation generation model,in order to deal with the occlusion,rotation and other problems,the local information of the image is considered,and based on this,the foreground target is modeled by histogram.Th e template updating algorithm in the tracking algorithm is analyzed.The discriminant classifier and the generated model are updated independently,and the latest observation results are combined with the initial template.It enables the tracker to deal ef fectively with the changes in the appearance of the target and the problem of slowing down the drift during the tracking process.