Research On Robust Object Tracking In Complex Scenes

Object tracking has traditionally been an important research topic in the field of machine vision.It plays a significant role in the fields of automatic object detection,intelligent transportation system,human computer interface,driverless cars,human action recognition and more.Although lots of researches on the object tracking algorithm have made significant progress in the past few decades,it is still challenging to design a truly reliable,accurate and efficient tracking algorithm to deal with complex tracking scenarios,such as scale variation,background clutter,geometric deformation and occlusion.In this paper,complex scenarios of visual tracking are discussed.On the basis of the remarkable achievements of the latest research in the field of object tracking,further research based on the most influential Kernelized Correlation Filter(KCF)tracker is conducted in our work.The main research achievements are as follows:(1)In order to overcome the drawbacks of KCF tracker,this paper improves the performance of KCF tracker from three aspects.Firstly,the original algorithm develops ridge regression model.This model is very sensitive to the outliers that occur in the tracking process.It causes over-fitting and results in tracking failure.Therefore,to improve the deficiency of this training model,a robust training model regularized by the L1 norm is proposed in this paper.This idea is inspired by tracking method based on the sparse representation.By using the computational advantage of the KCF tracker,an effective solution for our training model is proposed.This regularization model effectively alleviates the negative impact of outliers during the tracking process.Secondly,in order to solve the problem of scale evaluation that KCF tracker cannot handle,a separate scale filter is learned during the tracking process.This scale filter effectively solves the problem of scale evaluation.Thirdly,the traditional update model of KCF tracker can easily cause error accumulation and lose the target when it undergoes complex scenarios such as occlusion and fast motion.To overcome this problem,an adaptive update strategy is designed in this paper.Extensive experiments in various challenging situation demonstrate that the robustness of KCF tracker is improved by our methods.(2)Since the tracking algorithm based on correlation filter depends strongly on the spatial layout of the object,the algorithm is very sensitive to the deformation during the tracking process.Considering that the tracking algorithm based on color statistics has the advantage of dealing with deformation,this paper proposes that the improved KCF algorithm collaborates with the tracking algorithm model based on color histogram.In addition,an adaptive collaborative strategy is designed to achieve dual-model tracking.This tracking model is better able to deal with complex tracking scenarios.Extensive experiments in various challenging situations demonstrate that the robustness is further improved by collaborative model.(3)The previously improved KCF tracker uses a single feature to describe the target.However,a single feature cannot effectively describe the target in complex tracking scenarios due to various interference factor.Therefore,based on the original improved algorithm,this paper further combines multiple features to achieve object tracking.In addition,previously improved KCF tracker adopts a single template which is used to generate training samples to train the filter during the tracking process.Although it is easy to calculate for a single template,this training method often loses the useful information in the first few frames of the tracking results.To make the tracking algorithm more robust,this paper develops multiple templates during the training process and embeds them into the regularized training model of L1 norm.By using the properties of the resulting circulant matrix structure,an effective solution for our model is proposed in this paper.Both qualitative and quantitative evaluations on challenging sequences demonstrate that the proposed tracking algorithm performs favorably in complex scenes.