Visual Tracking Based On Spatial-Temporal Relevance Learning

As an important branch of computer vision,visual tracking has been widely used in intelligent monitoring,human-computer interaction,intelligent driving and so on.In the complex environment,the change of the target’s shape or the interference of environmental factors on the target’s appearance in the process of moving can easily lead to the decline of the tracker’s performance,which makes it a difficult to track the dynamically changing target robustly.Therefore,how to use the relevance information between samples to provide richer information for the tracking model,and how to construct a reasonable tracking model according to the changes of targets are the key to improve the performance of the tracker.To this end,based on the deep learning and correlation filter model of the mainstream visual tracking methods,combining the temporal and spatial information of samples,this article focuses on the adaptive suppression of boundary effect based on correlation filter tracking method,and the optimization of sample information based on deep network tracking model.The main research contents and methods of this article are summarized as follows:(1)Visual tracking based on Bayesian correlation filter learning is proposed.For the boundary effect suppression of correlation filter tracking model,to make the tracker deal with the morphological changes of the target in the process of motion,the research on the adaptive suppression of boundary effect is the key to improve the tracking performance of correlation filter.To solve the above problems,this article proposes a tracking method using structured Gaussian scale mixture model to decompose correlation filter.Specifically,the correlation filter is decomposed into the product of a sparse multiplier and a Gaussian scale vector by Gaussian scale mixture model.The sparse multiplier adaptively estimates the target and background information.By combining with Gaussian scale factor,the filter is constrained by sparse multiplier while learning Gaussian coefficient,so as to realize the adaptive suppression of boundary effect of correlation filter.Secondly,considering the correlation between adjacent pixels of the target,a structured Gaussian scale mixture model is introduced to model the spatial correlation of the target and background.Finally,the results of multiple test sets show that the proposed model is conducive to the improvement of the tracking performance of the correlation filter.(2)Correlation filter tracking method based on temporal-spatial information modeling is proposed.The boundary effect and model attenuation of correlation filter are the key factors that restrict the tracking performance.To solve the above problems,this article proposes a correlation filter tracking method based on temporal-spatial information modeling.Through the simultaneous modeling of temporal and spatial information,the adaptive suppression of boundary effect and the mitigation of model attenuation are realized.Specifically,the Gaussian scale mixture model is used to sparse model the tracking target,and the adaptive suppression of the boundary effect of the correlation filter is realized.Then,the Gaussian scale mixture model with non-zero mean is used to decompose the correlation filter,and the learned filter coefficients are transferred to the update learning of subsequent models through the mean value,so as to realize the adaptive modeling of temporal information.The acquisition of temporal information does not involve additional parameters and does not affect the overall computational complexity of the model.Furthermore,considering the correlation between the internal features of the target,a non-zero mean structured Gaussian scale mixture model is introduced to model the spatial correlation information of the target,so as to realize the correlation filter tracking method based on the decomposition of temporal-spatial Gaussian scale mixture model.Finally,through the comparative experiments in different tracking frameworks,it is proved that the proposed model is helpful to improve the positioning accuracy of the tracker.(3)Visual tracking method based on self-attention mechanism is proposed.The tracking algorithm based on the Siamese network performs tracking by calculating the similarity between the target template and the search sample.When the apparent of the target changes greatly during the movement,the target template can not provide effective matching features,which affects the tracking performance of the model.Reasonable use of historical data that is highly related to the search sample is beneficial to the tracking model to deal with the apparent change of the target.To solve the above problems,this article proposes a visual tracking method based on self-attention network learning.Firstly,the multiple historical samples are used as target templates.Compared with the method of using a single target template,the multiple templates can provide more abundant target diversity information.Secondly,the self attention module is used to calculate the correlation between the search sample and the template data,and obtain the correlation information of the samples in the temporal dimension.Then,according to the spatial structure information of the search samples,the query-read module is used to filter out the feature information from the template samples that is conducive to the tracking.The combination of temporal and spatial information module realizes the screening of template samples,and improves the target identification ability and tracking performance of the model.Furthermore,using the anchor free method to process the target anchor is helpful for the model to deal with the change of the target appearance.Finally,the effectiveness of the proposed model is verified by the results of different testsets.