Research Of Real-time Single Object Tracking Based On Kernelized Correlation Filters

With the constant improvement of the level of science and technology,upgrade of electronic components,in nowadays society the requirement of intelligent social is more and more higher,artificial intelligence technology combined with cutting-edge camera is also increasing,the progress of science and technology not only facilitates the human lives,but also proves the national comprehensive strength.One of the most important technical fields in the intelligent camera is computer vision,which is a very active frontier in the present situation,is closely related to people's life.Computer vision has been widely used in national security protection,traffic intelligent monitoring,autonomous vehicle,human-computer interaction and other fields,which is aroused the high attention of enterprise and university research institutes.Among them,the target tracking technology based on video is an important branch in the field of computer vision,and it also presents an increasingly prominent role and application scenarios.Currently,there are many superior algorithms for online target tracking,such as TLD,Struck,CT,etc.,but the track quickly and effective target tracking algorithm is not enough,Struck is about 20 frames per second,and the TLD is about 28 frames per second,only reach the requirements of the implementation for ordinary camera.Object tracking algorithm based on kernelized correlation filter can reach around 100 frames per second,part of the algorithm use circulant matrix technique to build classifier training samples,change the sample data matrix into a circulant matrix,such sampling greatly increased the number of samples,and improve the accuracy of the algorithm,at the same time,using a continuous label to mark samples,the label value of the distance near from the sample to the center of the target the label value tends to 1,and the label value the distance far from the sample center of the target tends to 0,effectively reflect the weight of each negative sample.Most tracking algorithm based on kernelized correlation filter through the character of the cyclic matrix,solving problem of classifier weights by changing time-domain convolution into dot-multiplication in Fourier domain,in order to avoid complex matrix inversion process and greatly improving the efficiency of the algorithm.Although the kernelized correlation filters tracking algorithm is superior to most mainstream algorithms in time,but it still can not solve the interference caused by the target tracking drift,fast motion,scale change,illumination change etc.The thesis in order to reduce the tracking performance degradation of kernelized correlation filters tracking algorithm in the conditions of target fast moving,scale change and light dark,proposes an online update algorithm based upon proportional,integration,derivative controller,and designed an effective anomaly estimation model to improve the drift problem.The algorithm updates the historical state of the tracked object by the form of oblivious proportion,at the same time,the differential state is added to predict the environmental change in advance,and using perceptual hash coding matching to judging whether the tracking is in a fault state,then control the update of classifier parameters:First of all,encoding the current frame of the tracking target and storing it,then repeat the same steps on the new incoming frame of the tracking target,compare the similarity between them;last but not least according to the similarity to determine whether to update the classifier parameters or to re-detect the target.In addition,the thesis uses the correlation of the upper and lower frames of the video to detect and fix the video jitter blur caused by the jitter of the imaging device.The experimental results show that algorithm mentioned in the paper is not only verify the strong robustness for scale change,fast moving,but also has strong robustness to other properties such as illumination change,occlusion and so on.Meanwhile tracking can still keeps high speed,the average processing speed can reach 100 frame per second.