The Research On Target Tracking Based On Spatio-temporal Context

The target tracking based on the vision is an important topic in the field of computer vision research.Target tracking is the process of detecting,extracting,judging,matching and recognizing the target from the image sequences.In this process,we can obtain the moving parameters of the target,such as position,velocity and acceleration to provide the support for the further research and analysis.Visual tracking,as an important branch of computer vision,has a wide range of practical applications,such as visual navigation,human-computer interaction,video surveillance and so on.Designing a robust visual tracking algorithm has a great significance in theory and practice.However,it is a very challenging issue to designing a general robust tracking approach,due to the complexity of environment and the variety of the object's appearance.In order to achieve the target tracking in complex environment such as illumination change,background disturbance,target occlusion,scale change and so on.This paper mainly focus on the use of target temporal and spatial context information and the scale change in complex environment.The main research process of this paper is:First of all,considering the shortcomings of the Fast Tracking via Spatio-Temporal Context Learning(STC)that can not accurately estimate the target scale,this paper presents a new algorithm for multi-scale tracking based on spatio-temporal context(SASTC).The algorithm is based on the framework of correlation filtering tracking,and the 1-dimensional object scale is calculated on the basis of the principle of2-dimensional target tracking.This algorithm divides the tracking process into two separate parts: the translation estimation and the scale estimation,and the scale estimation part can be used and applied by other algorithms under any tracking framework.Secondly,due to the two separate parts,we can choose different feature description and feature calculation methods to train and test the tracking process in each individual part.To improve the overall performance of tracking,we use the commonly used featureimage gray value in the position estimation part,and PCA-HOG feature in the scale estimation part.The different features can construct a more perfect and effective targetappearance model,and enhance the robustness of the algorithm.Finally,this paper optimizes the algorithm details on the whole.On the one hand,the image normalization is used in the process of algorithm implementation,which can effectively reduce the illumination interference caused by uneven light on the target image without changing the contrast of the image.On the other hand,to filter out the high frequency portion of the image edge and reduce the frequency effect of the image boundary,a low pass filter-Hamming window is added to the image before the Fourier transform.