Visual Tracking Algorithms Based On Temporal-Spatial Regularized Correlation Filters

With the rapid development of modern computer vision technology,object tracking,as one of important parts,has attracted more attention from researchers.Although object tracking has been widely applied in military,industrial,civil and other fields,it still has great spaces to be improved in its performance due to the influence of fast motion,rotation change,illumination change and other factors.This thesis mainly studies the object tracking algorithms under different appearance models in framework of correlation filters,and proposes three research work: visual object tracking using PCA correlation filters,visual tracking based on temporal regularized background-aware correlation filters,and visual tracking based on multi-temporal spatial regularized correlation filters.Specific research contents are as follows.1.Many existing trackers use some feature extraction methods and the exhaustive scale methods to solve accurate translation and robust scale estimation,respectively.Generally speaking,the more representative the selected object features are,the more accurate the center point of the target is.At the same time,the more suitable the candidate scale is,the closer the target scale estimation is to reality.To overcome above two problems,this thesis in the framework of correlation filters,proposes an efficient tracker that applies Principal-Component-Analysis(PCA)features and robust scale estimation.The proposed tracking algorithm can predict the location of the target more accurately,and keep the good performance for the scale variation via an accurate scale estimation method.Experimental results show that our proposed tracker has a better accuracy for predicting the location of the target and a higher percent in the average overlap precision than some other methods on the benchmark sequences with illumination variation,occlusion,background clutter.2.For the boundary effect and the similarity of the adjacent correlation filters,based on the background perceptron and regularization theory,this thesis proposes an efficient and improved tracker,named temporal regularized background-aware correlation filters(TRBACF)tracker.Firstly,the step of circular shift operation for collecting training samples in our tracker can obtain more background information and target information to solve the unnecessary boundary effect in some extent.Secondly,the new added temporal regularization term in our appearance model for training the filters can consider the similarity of the correlation filters for the adjacent frames aptly,which can obtain more accurate target location and scale and raise the tracking results mostly.Lastly,the alternating direction method of multipliers(ADMM)is applied to solve the appearance model efficiently.Numerous experiments illustrate that the proposed TRBACF tracker performs favorably against several the state-of-the-art trackers.3.For the boundary effects problem and tracking drift when the target is seriously occluded,a new tracker MTSRCF based on multi-temporal spatial regularization correlation filters is proposed in this thesis.Compared with the traditional trackers,a new coefficient matrix in our MTSRCF tracker can enhance the positive influence of the target block,and reduces the influence of the similar patches in the background in the tracking process,so to gradually reduce the background influence from the target edge.At the same time,by introducing multiple temporal regularization terms,the tracking results of our MTSRCF tracker are effectively improved when the target is occluded or the tracking drift occurs.In addition,the appearance model in our proposed tracker can be solved quickly by using ADMM for satisfying the real-time requirements of target tracking.Experiments illustrate that the proposed tracker has better results than some other the state-of-the-art trackers.