Research On Tracking Algorithm Based On Kernel Correlation Filtering Framework

There are many factors affecting target tracking,which results in that the tracking effect of the algorithm can not meet expectations.For more accurate and robust target tracking,in this paper,based on the framework of kernel correlation filter tracking,combined with the excellent tracking algorithms and knowledge in depth learning field in recent years,the following work is mainly done to solve the problems and shortcomings in tracking.Aiming at the one-sidedness of traditional features,the traditional tracking model for the problem of detecting the model drift problem and the lack of remedial measures,a kernel correlation filter tracking algorithm for residual depth feature and drift detection is proposed.The hierarchical feature is extracted by convolutional neural network,and the residual structure is added to the convolutional neural network to connect different network layers to achieve the fusion of shallow and deep features.There is no need for artificial design feature fusion,and the network structure can automatically implement the feature fusion function.The depth feature distinguishes the target and the background,and the resolution is more accurate than the traditional feature.The tracking result is more accurate.In order to judge whether the model drifts during the tracking process,the response strength down counter is designed.Each frame judges whether the model drift occurs according to the value of the counter.And take the corresponding model update plan as a remedy.The response strength drop counter detection model drift strategy can handle the task of tracking targets in different scenarios and achieve robust tracking.The kernel correlation filter tracking algorithm results for residual depth feature and drift detection are verified on the OTB dataset and have achieved certain expected results.For the expression of target features in the target tracking algorithm,the traditional feature expression has no deep feature,but the single-resolution depth feature is lacking for multi-scale expression of the target.In this paper,multi-resolution fusion features and kernel-correlation filter tracking of adaptive scale transform are proposed.The features of different resolutions are combined to construct multi-resolution fusion features to describe the target more comprehensively.At the same time,an adaptive scale transform method is proposed to integrate the position information of multiple candidate samples,reposition the target prediction position,adapt to the scale change of the target,and achieve accurate tracking target.The multi-resolution fusion feature and the adaptive scale-based kernel correlation filter tracking algorithm results are verified on the OTB dataset.Compared with the existing algorithms,the algorithm has a great improvement in accuracy and success rate.