Visual Object Tracking Based On Correlation Filters

In the field of computer vision,visual object tracking is an important branch of research.By using the collected video images,the video frames are processed and further analyzed to realize the location and tracking of the interested objects in the video.This technology has important applications in many fields.In recent years,many good visual object tracking algorithms have been proposed.However,there are still many challenges in object tracking,such as illumination variation,scale variation,shape variation and occlusion problem.These factors will generally affect the tracking effect.Among many visual target tracking algorithms,the algorithm based on correlation filters has been widely studied in recent years,because of its fast tracking speed.In this paper,the principle of visual target tracking algorithm based on correlation filters and the thought and flow of related algorithm are studied,and the improvement method is proposed and verified experimentally.The main work and innovation of this paper are as follows:(1)First of all,the basic methods of visual object tracking are studied and analyzed,and the challenges and difficulties are introduced.At the same time,the basic process of most visual target tracking algorithms is introduced.Then,the basic principles of correlation filters and the idea of applying it to visual object tracking are described.The formula of object tracking algorithm based on kernelized correlation filters is deduced in detail,and the steps of object tracking based on kernelized correlation filters are expounded.(2)The traditional correlation filters tracking algorithm usually only extracts single feature,in order to make the representation of the tracking target more reliable,this paper adopts the combination of HOG and LBP features for the feature representation of the object in the correlation filters tracking,which makes the representation of the target more accurate and anti-jamming,so that the tracking of the object is more accurate and robust.A multiscale correlation filters tracking framework is used to achieve more accurate object tracking by using the proposed multi-channel fusion characteristics as the input of the algorithm.The algorithm is tested and validated on the dataset and compared with other algorithms,the results show the effectiveness of the algorithm.(3)In order to solve the problem of the object tracking algorithm based on kernelized correlation filters,such as the variation of object scale and occlusion processing,the strategy is improved.Object tracking algorithm based on kernelized correlation filters,introduces the idea of cyclic sampling and kernel function,making sample collection more convenient and fast,and processing data more quickly.However,it still has a single feature,it cannot adapt to the object scale variation,cannot better deal with occlusion and other problems.In this paper,we use the fusion of color feature and gradient feature as the description of the object,so that the object is more abundant.For multi-scale changes of objects,the multi-scale detection idea of DSST algorithm is introduced into the algorithm,so that it can adapt to multi-scale changes in object motion,and make tracking more accurate.Aiming at the traditional mechanical model updating strategy,we decide whether to update the model by judging the Peak-Sidelobe Ratio,which can better handle the interference conditions such as occlusion.The algorithm improves the tracking algorithm from different parts of the tracking and it is verified by experiments that the improved algorithm has a good improvement effect.