Object Tracking Algorithm Based On Correlation Filter

In recent years,object tracking is widely used in intelligent monitoring,smart transportation,medical diagnosis,intelligent robots,virtual reality,and so on.It becomes an important research topic in computer vision.Due to complex realistic environment and various object movement,it is difficult for object tracking algorithms to meet all kind of needs in robustness and real-time performance.Correlation filter tracking has become a research hotspot in recent years due to its excellent performance in robustness and real-time.However,the correlation filter trackers cannot accurately deal with occlusion,target drift,fast motion.Aiming at the problems that the correlation filters lose objects when fast motion happens,target drift and occlusion,this thesis proposes a tracking method that combines adaptive template update with target refiner.The main research content and innovative work are as follows:Firstly,after research on the kernelized correlation filter,this thesis fuses and optimizes the features of color,gradient and gray level co-occurrence matrix.The color features can reflect the image color label space,and HOG features emphasize the image gradient information,and gray level co-occurrence matrix can highlights the image texture information.This thesis fuses these three features to be the whole extracted feature so that the object description information can complement each other.The feature fusion enhances the target's discrimination,and then it is reduced dimension effectively by taking into account the tracking effect and time efficiency.Secondly,the template update based on the object velocity and feature change is researched and an adaptive learning strategy is proposed.According to the object velocity and change rule of the current frame feature and the filter template feature,two kind of models,that are the model of velocity correlation with template learning rate and the model of feature change correlation with template learning rate,are constructed and combined linearly.Two models are learned adaptively according to the speed and feature change to enhance tracker adaptability effectively.Thirdly,to solve the problem of low long-term tracking performance of the correlation filter,this thesis proposes object location refiner which combines short-term and long-term filter to track objects.To imitate the principle of long-term memory and short-term memory in the human brain memory models,long-term filter is built to store the target features and slowly updated during the tracking process.The short-term filter response value is used to determine the long-term filter cooperative tracking rule.When the object tracking result is deviated,the long-term filter corrects the position.When the target is occluded,the long-term filter intervenes in tracking and searching object.Till the refiner is successful,the short-term filter is updated to continue tracking.Compared with the state-of-art algorithms on MATLAB platform by using the Visual Tracker Benchmark,experimental results show that this proposed algorithm has strong robustness in occlusion,target drift and fast motion and can achieve accurate real-time target tracking.