Research Of Visual Tracking Via Dense Spatio-temporal Context Learning

In this paper,we did some research on the spatio-temporal context tracking algorithms which locates the tracking target by analyzing the surrounding regions of the object.The algorithm studied in this paper is fast tracking algorithm via dense Spatio-temporal Context(STC).STC models the statistical correlation between characteristics of the tracking target and its surrounding area based on the Bayesian framework,changing the tracking problem to calculated the maximum point in confidence map.The FFT is adopted for rapid learning and detection in the process of modeling which greatly reducing the complexity of the algorithm and achieves good performance in terms of robustness and real-time.But there are still some problems:(1)STC and some other tracking algorithm using the Bayesian framework for visual tracking which will make a blind average of all tracked frames and increase the influence of these interference factor on the tracking model.Moreover,STC doesn't evaluate the tracking effect of current frame.For example,STC will learn the occlusion feature when the target is occluded,which will accumulate the error of the tracking model and affect the following tracking.(2)STC treats the global context equally and updates the appearance model at a fixed rate.The weight of spatial context only depends on the distance from the target center which cannot take advantage of useful information,and increase the interference of the error information to the tracking model,eventually influence of the tracking effect.In order to solve the above problem,several improvements are made to STC.The research contents and achievements are as follows:(1)We proposed a multi-templates tracking structure to make up the deficiency of the single model structure of STC,which using linear structure and easily to accumulate error.By comparing the adaptability of multiple templates,we choose the appropriate template to carry on the model learning,and remove the frame with too much interference information,so that the tracking model will not learning wrong information and improve the tracking effect.(2)We improved the weight allocation method of the spatial context.STC cannot effectively distinguish the useful information from the interference information in spatial context,in order to increase the success rate,we proposed a parted-based tracking method.By distinguishing the information in the spatial context of trackingparts,the STC can utilize the effective information to track and improve the tracking accuracy of the algorithm.(3)Experiments and analysis of the improved algorithms are carried out respectively.The experiment shows that two methods we proposed can improve the tracking success rate and reduce the tracking center location error.The result shows that our methods can meet the requirements of real-time tracking.