Research On Visual Object Tracking Method Based On Correlation Filter

Object tracking is a basic direction in the field of computer vision,which involves many theories such as signal processing,probability theory,neural network and multi-sensor data fusion,etc.,and it has a strong interdisciplinary and comprehensive nature.Thanks to the rapid development of digital and information technology,object tracking has been widely used in various social production and life,such as security video monitoring,sports analysis,abnormal event detection,human behavior recognition,auto autonomous driving,medical image processing,mobile robots,film post-processing and human-computer interaction.In recent years,correlation filters have been introduced into the framework of object tracking,and the good mathematical properties of cyclic matrix and Fourier domain have opened up a new direction in object tracking.Although the tracking method based on correlation filter has achieved remarkable results in both accuracy and speed,it still has many shortcomings,which are mainly reflected in the following aspects: 1)The reliability of tracking.For the object to be tracked,during the tracking process,the object may be occluded,the model is not accurate enough,the influence of interfering object and the interference of background information,etc.,which will inevitably lead to unreliable tracking results and the object cannot be tracked accurately.In order to adapt to changes in the appearance of the target during tracking,the tracking results at the current moment are used to update the tracker model.However,unreliable tracking results will cause the model to deviate from the true value,so repeatedly that the final model may be completely discredited and the tracking completely failed.On the other hand,most of the current conventional reliability detection methods only detect the object in the current image,and only stay in the static detection range,while the fast movement of the object and clutter interference are often a coherent influence,and have no detection ability for the dynamic characteristics between the two frames.2)Model update.In order to keep track of the object,the appearance model of the object must be able to adapt to such scenes as fast movement,background clutter and occlusion influence,so appropriate model parameter update is critical.Most trackers set a fixed update rate and update the current model with linear interpolation.Although this method is simple,the fixed-rate updating method does not distinguish the changes of different scenes in each frame,which makes the model prone to drift and even tracking failure when it encounters scenes with great changes.3)Fast scale estimation.Accurate and fast scale estimation of objects is a challenging research problem in visual object tracking.There are mainly two scale estimation methods for the tracker based on correlation filter,SAMF and DSST.SAMF advocates a multi-scale estimation method using a translational correlation filter.This method is simple but not efficient.In order to obtain accurate scale information,it needs to be estimated on a more detailed scale,while too many traversal scales lead to heavy computation,which is contrary to the real-time requirement in practical application.DSST advocates the use of independent correlation filters for scale estimation.In addition to the translation filter of position estimation,an independent scale filter is also used in this method.DSST has the advantage of tracking speed,but its tracking accuracy becomes its biggest bottleneck due to the use of fixed scale filter as position estimation.Through in-depth analysis and summary of the shortcomings and challenges of the existing methods based on correlation filter,this thesis carried out the following research work:(1)Research on the reliability of tracking process.In order to detect the dynamic characteristics of target confidence changes and avoid the introduction of unreliable information in the tracking process,this paper proposes a tracking reliability judgment method combining the average peak correlation energy threshold and its gradient threshold.Thanks to its gradient change can reflect the degree of dynamic object appearance changes,the proposed method not only can achieve in the current frame to judge the reliability of the tracking ability of "static",but also can detect different between two frames on the timeline goals change degree of "dynamic" reliability,so this method in fast moving and background clutter,challenging environment can well capture the object of overall motion tracking reliability,achieve good tracking effect.(2)Research on model updating.Aiming at the adaptability problem of fast-moving object tracking,a model updating method which can effectively capture fast motion is proposed,and the model updating rate is dynamically controlled by the results of previous reliability judgment studies.The tracking reliability factor is used to dynamically control the updating rate of the model and restrain the excessive accumulation of model error.Firstly,the reliability of the tracking results is measured according to the average peak-related energy threshold and its gradient threshold.Change update rate when tracking is unreliable.Different from the traditional tracking method,the threshold value of average peak correlation energy and its gradient threshold value are enhanced to increase the detection ability.At the same time,in order to get more object information as much as possible,the initial template information is used to assist the modification of model parameters.When tracing reliability is poor,appropriate heavy weights are assigned to the initial template to make it possible for the object to be recaptured.Because this method can update parameters according to the dynamic change control model of object tracking,it makes the model more adaptive and has a better and more significant effect on fast moving object tracking.In addition,the addition of the initial template information further improves the accuracy and robustness of the tracker.(3)Model updating method of simulation regularization in time domain.Aiming at the problem of tracking model "overfitting" and degradation,inspired by STRCF,a spatialtemporal regularization algorithm,this paper proposes an adaptive control of updating rate by using the variation degree of model parameters between the two frames.This method is simple and effective.On the one hand,it avoids the inconvenience of solving the closed solution and has good generalization ability.On the other hand,the proposed algorithm has fewer super parameters,which not only makes the debugging parameters simple and effective,but also avoids the "overfitting" phenomenon among parameters while reducing the control link,further improving the robustness of the algorithm.(4)Research on fast scale estimation.In order to solve the complexity problem of multiscale traversal estimation,a method based on scale change direction and tracking reliability is proposed for adaptive estimation of target scale.Based on the scale estimation method of SAMF,the original seven fixed scales are reduced to three,and an adaptive scale is added.Three fixed scales are used to determine the direction of scale change,and the adaptive scale of the next frame is controlled by the rate of change of the average peak relative energy of the current frame and the previous frame.Finally,the optimal scale is determined according to the maximum response results.On the one hand,the proposed method avoids the biggest bottleneck problem caused by DSST algorithm using fixed scale filter as position estimation,and on the other hand,it reduces the problem of SAMF algorithm that too many scale factors lead to heavy computation and low tracking speed.The experimental results show that the proposed algorithm reduces the computation and can not only estimate the object scale quickly and improve the tracking speed,but also has a high tracking accuracy.