Research On Object Tracking Algorithm Based On Correlation Filtering

Object tracking is one of the basic problems of computer vision.Object tracking is important in many real-time visual applications such as intelligent monitoring systems,autonomous driving,security and surveillance,video communication and compression,traffic control,medical diagnosis and video editing.Important role.Many research teams have proposed a variety of classical algorithms,and continue to improve them.In recent years,object tracking algorithm based on correlation filtering has attracted the attention of scholars with its amazing speed and precision,and has obtained great results continuously.However,object tracking is a highly challenging task,for the moving target,the actual motion is very complicated,such as the change of illumination,the clutter of the background,or the change of target's appearance,such as rotation,occlusion,etc.In the process of the research,we need to pay attention on how to design a robust algorithm.To solve these problems,in this paper,we propose a correlation filter tracking algorithm based on feature integration and a correlation filter tracking algorithm based on feature integration depth features,so we can improve the performance of traditional correlation filter tracking algorithms.The following is the primary work and innovation.HOG Feature is calculated on the local cell of the image.It is invariance of the geometric and optical changes of the image,but the HOG feature is difficult to deal with the occlusion problem.And it is difficult to detect the change of object orientation changes or excessive range of postural movements,and it is very sensitive to noise.CN Feature is not sensitive of the size,orientation and viewing angle of the image,and is not affected by rotation and positional changes,so it is highly robust,but it can't describe local features of the target very well.LIOP feature describe the target appearance based on key point detection of gray image.It is robust to image rotation,compression,motion blur,scale change and so on.Therefore,we combine the 31 channel HOG feature,the 11 channel color feature and the 1 channel LIOP feature into a multi-channel fusion feature of 43 channel.For the scale variation,after obtaining the position with the highest accuracy,we establish a scale pool centered on the point,use another scale correlation filter to select the best scale.We design a reasonable model update method for the model pollution caused by occlusion.Deep learning is effective for many tasks in the visual field.Depending on its powerful representation ability,how to apply its unreplaced powerful representation to the object tracking is our key point to focus.But target tracking is only given the initial state of the target in the first frame,the training samples is scarcity,and it requires realtime,so it is unrealistic to apply the deep learning framework directly to the field.There are a lot of powerful deep learning network structure that has been trained now.Therefore,we use VGG-19 network structure to extract the target features,at the same time,in order to make better use of the features of each layer,we define each layer depth feature as a decision maker,and the relative error and the cosine theorem of the space vector are used to allot trust value for every decision maker,we design a response graph adaptive fusion method.Through the quantitative analysis of the proposed algorithm and other advanced algorithms on the OTB dataset,the experimental results show that the proposed algorithm is robust and accurate when the target in these situations,like fast motion,deformation,occlusion,and out of view.