Target Tracking Algorithm Research Based On Space-time Constraints And Correlation Filter

With the continuous progress of computer technology,target tracking technology continues to improve and has a wide range of applications in video monitoring,driverless and other aspects,which is a hot spot in the field of computer vision.When the background or similar objects interfere with the target tracking algorithm,the accuracy will decrease.Therefore,it is very important to design an algorithm with good adaptability to complex background or similar objects.This paper proposes an adaptive space-time constraint based target tracking algorithm,which optimizes feature selection.The main research work of this paper is as follows.Firstly,The deformation of the target and the similar objects around the target may interfere with the solution of filter.To solve this problem and the problem of boundary effect in the filter training stage.Space-time constraints include space constraints and time constraints.The spatial constraint is to distinguish the foreground and background by using the characteristic histograms of the foreground and background of the sample area,so as to constrain the value of the relevant filter corresponding to the background.Time constraint is that the optimization of correlation filter should consider the filter of the previous image,that is,the norm of matrix obtained by subtracting the correlation filter of two consecutive images.Secondly,in the stage of fixing the target's location,In view of the problem that the sample image is polluted during the tracking process,this paper proposes a filter update decision mechanism under the condition of high confidence.After the target's localization,the maximum value of the response can be obtained from the convolution response of the filter and image features,and the average peak correlation energy(APCE)can be calculated.The joint criterion of these two values can be used to determine whether the target area is polluted and whether the correlation filter is updated in the tracking process.At last,the depth feature is introduced.The depth feature is extracted by the depth learning network of VGG,and the feature channel is selected by principal component analysis to reduce the redundancy of the feature.A new correlation filter is trained by the depth feature.The optimal scale of the target is determined by the filter trained by thetraditional feature,and the center position of the target is determined by the two filters trained by the traditional feature and the depth feature respectively.In order to objectively evaluate the target tracking algorithm in this paper,it is compared with different algorithms under the standard evaluation index.