Research On Object Tracking Algorithm Based On Deep Learning And Correlation Filtering

In recent years,the field of object tracking has developed rapidly.In particular,many excellent algorithms have emerged in deep learning and correlation filtering methods.However,there are still some problems in these algorithms.This paper studies the methods of deep learning and correlation filtering.The main work is as follows:In the first part of the thesis,we use the current popular convolutional neural network to extract features and propose an adaptive update object tracking algorithm based on multi-layer deep features.Use the deep feature as the target's appearance expression,combined with the shallow space features and high-level semantic features of the deep network,training two tracking filter templates respectively,and finally summing the tracking results of the two filter templates to obtain the final target position.In the algorithm,an adaptive update scheme for the target appearance model and the filtering template is also designed to better adapt the appearance change and occlusion of the target.This chapter adopts the traditional correlation filtering framework,which has certain drawbacks.In the following,the relevant filtering algorithms will be further improved.In order to solve the problem of the corrupted samples generated by the cyclic shift of the samples in the traditional correlation filtering algorithm,and to make better use of the information of the prior training samples,this paper proposes a tracking algorithm for the adaptive template templates with continuous sample weights.In the regression optimization calculation,the weight parameters of the filter template and the weight parameters of the training samples are added,the joint training framework is designed,and the filtering template and the sample weights are trained at the same time,so adaptive sample weights Can make better use of sample information.Get more accurate filter templates for better tracking.In addition,the multi-feature joint sparse representation method is used to fuse the advantages of target colors and gradient texture features.Experiments have shown that this algorithm is more accurate than conventional correlation filtering algorithms,especially for some fast motion and video where the target rotates,tracking can be better achieved.In addition to tracking by using training templates such as correlation filtering,when tracking is performed using the classification method,the traditional sample selection will generate multiple calculations,which is not conducive to the later classification.Aiming at these problems,this paper proposes a tracking framework based on convolutional neural network,and innovatively proposes directly extracting samples on the feature map,and training the size of the target box as a parameter,calculating the regression of the target box size,and adjusting the sample box,so that the bounding box can be more suitable for the target size.The overall tracking framework is performed on a convolutional neural network and trained using video data sets.In order to achieve the final classification task,the normalized pooling layer of the region of interest is added after the sample box is determined,and the sample frames are normalized to the same size.