Research On Visual Tracking Algorithms Based On Hierarchical Correlation Filters

Visual target tracking is one of the hotspots in computer vision research.It has been widely applied in sports events,precise navigation,intelligent monitoring,and so on.The task of visual target tracking is to design a tracking algorithm to determine the position and size of the target in the next frame according to the initial position and size of the target in the first frame.However,in the process of tracking,we will encounter the interference of target deformation,local occlusion,background interference and motion blur,resulting in the deviation of tracking results.Therefore,it is still a challenging task to design a practical and accurate tracking algorithm.Aiming at the problems of object’s appearance variation,background interference,spatial information,and time information,this thesis design some effective object tracking algorithms based on the framework of correlation filters for improving the accuracy and success rate.The main research contents of this thesis include a target tracking algorithm based on time-region sparse correlation filter,a hierarchical feature fusion tracking algorithm based on time-region sparse filtering,and a target tracking algorithm based on hierarchical correlation filtering.The specific contents are as follows:(1)Traditional correlation filter based tracking algorithms mainly learns the correlation filters from the features extracted from the sample area.However,some interference factors(e.g.occlusion and deformation)will affect the tracking performance,which results in the target drifting in the follow-up tracking.In order to solve above problem,a target tracking algorithm based on time-domain sparse correlation filter is proposed based on elastic network model.This algorithm uses binary mask to limit the range of filter value in the target region of the sample region,which makes the extracted features more representative.At the same time,this algorithm considers the adaptability of spatial weight,and fuses the spatial information and the temporal information to establish the relationship between the filters of two adjacent frames,which can effectively alleviate the filter degradation caused by the target occlusion,fast movement,and other factors.Experimental results show that our proposed tracking algorithm has better tracking performances than some existing popular algorithms.(2)Sample features,background information,spatial information,and temporal information play important roles in improving the performances of the correlation filter based tracking algorithms.In this thesis,a new hierarchical feature fusion tracking algorithm is proposed by developing the classical hierarchical convolution feature tracker using a temporal and region sparse filter with the sparsity and the peak side lobe ratio.Firstly,a temporal and region sparse filter is proposed,and it is trained by the extracted features with different layers of a deep convolutional network.Secondly,these correlation filters are applied to determine the target’s response maps for the next frame.Finally,the weights of different response maps are computed by the peak side lobe ratio.Then the fused response map is obtained to determine the location of the target.In our proposed tracking algorithm,the region sparsity is used to ignore the interference feature,the time regularization term is used to learn the relation of the last filter and the current filter,and a new response-map fusion method is applied to improve the performances of the tracking algorithm.Experimental results show that our proposed tracking algorithm has better tracking performances than some existing popular algorithms.(3)Existing correlation filter based tracking algorithms mainly use prior knowledge to design some regularization items,e.g.spatial regularization,temporal regularization,abnormal regularization,and put forward various regularization correlation filter appearance models.Therefore,the existing regularized appearance model becomes more and more complex,which increases the difficulty in solving the optimization model to a certain extent.In order to solve above problems,a universal correlation filter based tracking algorithm is proposed to represent the complex target tracking algorithm as a linear combination of some simple trackers.Taking the classical spatial-temporal regularized correlation filter based algorithm as an example,the filter is expressed as a linear combination of a discriminant filter,a spatial filter,and a temporal filter,where the combination coefficient is determined by the maximum value of the response graph calculated by each method.The proposed tracking algorithm can not only save the computational cost of appearance model,but also adaptively utilize the prior information.Experimental results show that the tracking algorithm proposed in this paper has better tracking effect than the corresponding complex tracking algorithm.