Research On Visual Target Tracking Algorithm Based On Adaptive Space-time Perception

Target tracking is an important branch of machine vision,its purpose is to track the specified target in the video.From the perspective of tracking data set,it is mainly divided into ground surveillance video and UAV surveillance video.This theis considers several major UAV video datasets and one ground surveillance video dataset.Because of the various target tracking algorithms based on correlation filtering show good performance in video tracking,among which the adaptive space-time sensing algorithm is one of the more popular models recently.The two important indexes to evaluate the performance of target tracking are success plot and precision plot.Based on these two indexes,this thesis analyzes and improves the Automatic Spatio-Temporal model.In order to solve the problem that the update of correlation filter is not disturbed by redundant information,an optimal transport distance combined low-rank response method is proposed in this thesis.In this thesis,the empirical parameter distribution in temporal domain is learned by using the optimal translation distance,and the variance invariant property is used for temporal translation.The correlation filter is embedded before the temporal degradation successfully.In addition,the consistent low-rank constraint is used to approximate optimal response reasoning to achieve global response consistency.This approach utilizes the structural consistency of the correlation filter in adjacent sequence of the video,which is helpful to enhance the stability of the updating of the correlation filter.In several common standard data sets,the results demonstrate superior tracking performance of the new algorithm compared to the most advanced algorithms.When the correlation filter is updated,Automatic Spatio-Temporal method does not make full use of the static characteristics of the target and does not consider the noise interference caused by the discontinuity of the boundary gradient.In this thesis,a dual Gaussian mask weighting combined with dynamic and static guidance method is proposed.The global response of the target is weighted by binary Gaussian mask to preserve the global response of the center,in order to avoid the interference of discontinuous boundary gradient.In addition,this thesis uses the input characteristics of the static image in adjacent sequence of the video and the dynamic structure of the filter to joint guide the updating of the relevant filter,to reduce the influence of target deformation on the accuracy of model updating.Compared with the most advanced algorithms,the results in several data sets demonstrate the superiority of our algorithm.