The Real-Time Tracker Based On Correlation Filter

Object tracking is an important branch of computer vision,which has been widely used in many areas,such as,face recognition,intelligent traffic,robot and UAV.Object tracking is mainly based on the operation of feature extraction,detection,recognition and tracking of specific objects in the image.After that,the trajectory of the object is obtained,which provides accurate motion parameters information for application systems.From the development process of object tracking,the main algorithms can be divided into three categories,classical algorithm,deep-learning algorithm and the method based on tracking-by-detection or correlation filter.Although many algorithms have been proposed,there are still some problems to be solved.The main problem is how to balance the precision and speed of tracking.In order to solve the problem mentioned above,we propose a real-time tracking algorithm based on correlation filter.Through the experiment,we have obtained the following research results.First,the color histogpram and the histogram of oriented gradient(HOG)modified by DPM are used as the initial features of the object.HOG is the descriptor of the image texture information,which can ensure the local robustness of the image information in the case of illumination change.In addition,the histogram of oriented gradient improved by DPM can reduce the dimension of the histogram,and reduce the computational complexity of this algorithm.Color histogram is the feature that characterizing the overall color distribution of the image.When the object is deformed and rotated,the color can maintain the global robustness and maintain the invariant color.Therefore,combining the histogram of oriented gradient and color histogram as the basic characteristics of tracking object can ensure that the algorithm has higher accuracy under illumination,deformation and rotation.Secondly,we design a linear function to fuse color histogram and the improved histogram of oriented gradient according to the different weights,so that we can get the new fusion histogram.As the effective feature of the tracking process,the new histogram is applied to the follow-up filter tracking process.The feature is combined with histogram of oriented gradient and color histogram according to the different weights,making it more accurate to characterize the information of the image.Thus,the tracker will be more accurate.Thirdly,based on the fusion feature,we train a correlation filter model with the update of scale to predict the location of the object in the new frame,and update the relevant parameters in the filter according to the object location.The initial position of the object is predicted by the correlation filter based on the fusion characteristics.We can collect some candidate objects based on the scale transform of the initial position,and screen the most matched candidate objects as the final result in this frame.Meanwhile,we use the final result to update the relevant parameters of the correlation filter.Finally,we compare and discuss the real-time tracker based on the correlation filter in this paper.It can be concluded from the experiment that the average speed of the proposed algorithm is 81fps,and it can keep the real-time performance of the tracking.Moreover,under the condition of light change,occlusion,deformation,rotation and scale change,the center error of tracking can be reduced,and the accuracy of algorithm is improved.The success rate of algorithm reaches 0.761 on OTB-50 platform.In addition,the paper also introduces an improved TLD algorithm with Harris corner and color moment published by in 2016.The algorithm takes advantage of the characteristics of Harris corner and color moment to reduce the center error in tracking process and improve the accuracy of the algorithm in maintaining the speed of TLD algorithm.