Correlation Filter Tracking Research On Spatiotemporal Regularization And Feature Reliability Assessment

Object tracking is a research focus in the field of computer vision,which can be widely used in life fields,such as video surveillance,traffic,human-computer interaction and so on.In recent years,the performance of tracking algorithms has been improving,but it is still a challenge to design a robust tracking model in complex scenarios.In this thesis,aiming at the shortcomings of traditional tracking algorithms,the following works based on the correlation filtering framework are done:To solve the problem that spatial regularization weight is independent of the target content and model degradation during tracking process in traditional correlation filtering tracking algorithms,an object tracking algorithm based on temporal awareness and adaptive spatial regularization is proposed.Firstly,in the first frame,the initial spatial regularization weight with target information is obtained by using the image saliency detection algorithm.Then,an adaptive spatial regularization term is added to the objective function to alleviate the influence of the boundary effect on the correlation filter.Finally,the temporal regularization term is introduced to enable the correlation filter to learn the information between adjacent frames,thereby reducing the risk of overfitting when dealing with inaccurate samples.Experimental results demonstrate that the proposed algorithm has good robustness in complex environments on the OTB dataset.Aiming at the problem that traditional correlation filtering tracking algorithms ignore that various features have different representation ability to the target,an object tracking algorithm based on adaptive feature fusion is proposed.Firstly,a feature reliability evaluation method is proposed by combining two tracking quality evaluation indexes,peak to sidelobe ratio and average peak-to-correlation energy.Then,several filters are trained separately by using different features.Finally,the tracking response of multiple filters is weighted and fused according to the reliability coefficient to locate the object.This method effectively highlights the contribution of features with better representation ability in subsequent frames,so that the fusion tracking response map has less noise.Experimental results on the OTB dataset show that the proposed algorithm has achieved the desired results in terms of tracking success and accuracy rate.