Research And Application Of Object Tracking Based On Correlation Filters In Complex Scenes

Object tracking has long been a hot research branch in the field of Computer Vision(CV).In the case where only the target initial frame information is provided,the motion path of the target subsequent frame can be continuously tracked.As a research direction of multidisciplinary integration,it has been applied in many fields such as missile guidance,intelligent monitoring,drone tracking and medical image analysis.From the generation of model shifting algorithms such as early Mean-Shift,particle filters and optical flow method to the current mainstream correlation filters and deep learning,the continuous breakthrough of discriminant model tracking algorithms makes the existing object tracking technology greatly improve the performance of the algorithm.In particular,the rapidity of the correlation filters solves the real-time problem of the algorithm.However,there is still no general algorithm that can solve object tracking problems in complex scenes,such as illumination variation,deformation,occlusion,scale variation,and so on.Aiming at the above problems,this paper studies and improves the Kernelized Correlation Filters(KCF)and Efficient Convolution Operators(ECO)from the aspects of multi-feature weighted fusion,dynamic occlusion and fast scale estimation.The main work is as follows:(1)According to many research institutes,the use of multi-feature fusion methods tends to be better than single features,and it is also important to study how to carry out multi-feature fusion.Based on the in-depth study of Kernelized Correlation filters theory,this paper proposes a novel feature weighted fusion method for Histogram of Oriented Gradient(HOG)and color names(CN).The two features are tracked based on the KCF framework to obtain the corresponding detection response graph and normalized.The normalized response graph and the Euclidean distance of the expected Gaussian label are calculated as the weight coefficients of the feature,and the features are weighted and summed.Finally,using a weighted response map acquired to determine the predicted position of the target.(2)When the target occlusion,out of view,the fixed model update learning rate can easily learn the background information,resulting in tracking failure.Therefore,it is so vital to research the method of solving model update adaptation.This paper elaborates a model drift suppression method based on similarity measure.Firstly,a fixed-size target set is designed,and the target patch obtained in the tracking process is measured by similarity.The comparison results determine whether to update the target set and set the model update learning rate.(3)Although the correlation filters based tracking algorithm has achieved great improvement in accuracy and robustness,many early tracking algorithms have not solved the target scale change and the current mainstream scale estimation methods have computational redundancy and fixed scale factor.Defects,which affect the real-time problem of the algorithm,and changes in the target scale will seriously affect the stability of the model,which in turn affects the overall performance of the tracking.In view of the above problems,based on KCF,a fast and novel scale estimation strategy from coarse to fine is proposed.Using the product of the peak value and the corresponding response graph weighted value instead of the individual peak for comparison.Initially using three scale factors to roughly determine the direction of the target scale change,and then solving cyclically the optimal scale in that direction.In order to further verify the validity and portability of the method,the proposed fast scale estimation method replaces the ECO original scale evaluation method.In this paper,OTB-2013,OTB-100(Object Tracking Benchmark),Temple Color 128 and UAV123 data sets are compared experimentally,and the experimental results of various methods are improved quantitatively and qualitatively.The experimental results fully demonstrate the effectiveness of the three-point improvement proposed in this paper and can guarantee the real-time performance of the algorithm.Moreover,based on Matlab GUI technology,the corresponding target tracking system is designed,and the external camera is called in real time to further test the practical performance of the improved algorithm.