Research On Object Tracking Algorithm Based On Correlation Filter

Object visual tracking is an important research field in pattern recognition and computer vision.It has been widely used in many practical applications,such as smart cities,security monitoring,medical treatment and so on.It is a research direction with wide application and many technical points.With decades of research and development,there are many excellent target tracking algorithms,that improved tracking precision and speed.However,as the target tracking environment is complex in practical applications,there are still many difficulties,such as dealing with occlusion,illumination variations,shaking and motion blur.Given the challenge of these problems in visual tracking,there is still a long way to go for how to achieve robust tracking.This dissertation focuses on visual tracking,analyzes the advantage and disadvantage of various object tracking algorithm and reviews the history and development of Correlation filter(CF).In addition based on the CF,we mainly focus on the problem such as occlusion,rotation and many changes on the object appearance,as well as the effects of fast motion,background clutter and illumination variations during the movement.To improve the accuracy and precision of object tracking,the main contributions of this dissertation are summarized as follows:(1)The standard Kernel correlation filter algorithm(KCF)is computationally efficient.However,it loses accuracy in complex backgrounds.Meanwhile,KCF cannot fully use particle information of the image.Mean Shift algorithm(MS)could adapt to changes in the appearance of the targets and complex backgrounds by using statistical histogram information to build target model.In this work,block MS is introduced into KCF.During the tracking process,MS could adjust the target locations,which are obtained by KCF,and locate the correct positions.The experiment result illustrates that the improved algorithm achieves good performance in handling images with backgrounds clutter,illumination variations,occlusion and rotation problems.(2)In order to improve the accuracy of CF algorithm in target tracking under the occlusion and background clutter,adjacent frame subspace representation and color histogram are combined with CF algorithm.First,CF algorithm continuously updates the target appearance model in order to overcome the challenges of background clutter and motion blur.Second the subspace representation between adjacent frame is sensitive to huge adapt variation,which is robust to backgrounds clutter,illumination variations.Third,color histogram utilizes the color information,with the ability of occlusion-aware and rotation-aware.Combining the three aspects,it can be more robust and accurate to adapt diversify.Experiment results in tracking benchmark shows that the improved algorithm achieves outstanding performance when dealing with occlusion,rotation,deformation,background clutter and etc.(3)The CF algorithm is sensitive to similar objects around the target in the tracking process,which cause the tracking failures.Therefore,adaptively finding patch that are similar to the target and adding the similar patch them as the negative samples in the training process can avoid the influence of similar patches on the classifier during detection.The experimental results show that this method is robust to target occlusion and rotation,and improves the accuracy of target tracking.