Research On Object Tracking Algorithm Based On Correlation Filtering

With the development of the time and the improvement of people's living standards,computer vision has occupied an increasingly important position.As an important part of computer vision,object tracking algorithm is applied to various scenarios such as military operations,civil security,road traffic and so on.Researchers have proposed a variety of research algorithms to deal with the challenges brought by various complex tracking scenarios.It is still a problem with the current algorithm that the contour and scale change and the object cannot be effectively tracked when it is occluded.Based on the flow and characteristics of the object tracking algorithm,the principle and model of the correlation filtering algorithm are analyzed and studied in this paper.It points out that the most important part of the correlation filtering algorithm is the extracted object features.Taking the two classic algorithms of MOSSE and CSK as an example,the secret of high-speed filtering algorithm is to convert the correlation operations in the time domain into multiplication operation in the frequency domain.It is analyzed that the correlation filtering algorithm needs to sample a large number of samples during tracking,which leads to the contradiction between high computational cost and tracking speed.This contradiction can be solved mathematically by simulating dense sampling by the method of circulant matrix,which improves the tracking accuracy and speed simultaneously,and the nonlinearity of low-dimensional samples is mapped to the linear relationship of high-dimensional by using the kernel method.The frequency domain solution of the correlation filtering tracker is given,which simplifies the complexity of the tracking algorithm and improves the tracking quality.Based on the object scale deformation problem faced by the tracking algorithm in the actual scene,the influence of the object features on the algorithms such as SAMF,CN and KCF is analyzed.By analyzing and verifying the LBP texture features,Hog gradient features and CN color feature extraction methods and effects of the object itself,a scale adaptive algorithm based on DSST algorithm is proposed.The improved SAMF feature is applied to the DSST position filter to enhance its discriminating ability.Aiming at the boundary effect problem generated by the circulant matrix,the mask matrix is used to minimize the error function,and the optimal filter parameters are solved by ADMM iteratively.Finally,the indicators on the OTB test benchmark and the test of multiple challenging sequences are compared with correlation filtering algorithms on the dataset to verify the effectiveness of the tracking algorithm and analyze the adaptability of the improved algorithm itself.Aiming at the lack of long-term tracking algorithm in the correlation filtering algorithms,the TLD tracking algorithm is introduced in detail,and a long-term tracking algorithm based on DSST and TLD tracking algorithm is proposed based on the components of TLD algorithm,and the implementation principle of the improved algorithm is presented.The algorithm is tested and compared on the corresponding sequences of OTB,and then according to the test results,advantages and disadvantages of the improved algorithm are analyzed,and effectiveness and limitations of the algorithm in the tracking scenario are obtained.After analyzing the algorithms mentioned in the paper,the limitations of the improved algorithm and existing current tracking algorithm are obtained,and it is pointed out that the improvement and optimization direction of the correlation filtering tracking algorithm in the future will tend to solve the high precision of the correlation filtering tracking and the low-speed of deep learning tracking.