Research On Correlation Filter Tracking Algorithm Based On Improved Transforme

Visual target tracking is a basic but still challenging task in the field of computer vision.It is dedicated to giving target information from the first part of a video sequence to specify the target frame of the target in each screen after the continuous video sequence.In recent years,with the development of computer technology and the rise of human intelligence,target tracking technology has made great progress.It has important significance and broad application prospects in the fields of human intelligent communication and automatic driving.Correlation filter tracking algorithm has attracted extensive attention in the field of target tracking because of its huge advantage in following speed.With the rise of deep learning,compared with the traditional use of handcraft to extract features,the features extracted by deep convolutional neural network are invariant to translation,rotation and shrinkage,and have stronger robustness.Therefore,based on deep features and association Modeling of filter tracking is one of the leading trends in the field of visual object tracking.Whether it is a correlation filtering algorithm based on traditional artificial feature extraction or a correlation filtering algorithm based on neural network feature extraction,each frame of image is processed independently to complete the tracking task,and there is no way to integrate these frames with each other.The inter-frame information establishes the connection between frames and ignores the rich temporal and spatial information between frames,which is especially important for tracking tasks.Accuracy is critical.Recently,Transformer has made great achievements in fields such as natural language processing,and the core attention mechanism of Transformer is famous for its ability to naturally integrate sequences of global relational information.Therefore,we try to introduce Transformer into the depth correlation filter tracking algorithm to deal with the above limitations.But at the same time,the application of the attention mechanism in the classic Transformer in the field of view is performed pixel by pixel on the image features,and the computational complexity is relatively large.At the same time,the overall information of the target is also ignored to a certain extent.This paper studies the above problems and proposes a correlation filter tracking algorithm based on improved Transformer.The main work is as follows:First,on the basis of TrDiMP,this paper designs a multi-frame feature fusion module based on Transformer,so that the model can establish the connection between frames and integrate the spatial and temporal information between the first frame,the current frame and the historical frame,to maximize the use of the global information of the entire video sequence to improve the tracking accuracy.Second,on the basis of the first problem,introduce the idea of Window based attention computing mechanism in the multi-frame Transformer feature fusion module,introduce multi-scale circular shift window attention,and raise attention from pixel level to window level,for visual object tracking.Multi-scale attention across windows has the advantage of aggregating attention at different scales and generating the best fine-grained matches for target objects.In addition,the cyclic shift strategy improves the accuracy by expanding the window samples by using the position information,and at the same time saves a lot of computing power by eliminating redundant calculations and improves the speed of tracking.Third,on the GOT-10K,TrackingNet,and OTB2015 datasets,the proposed correlation filter tracking algorithm based on multi-frame Transformer feature fusion and multi-scale cyclic shift window attention was experimentally verified.Experimental results demonstrate the effectiveness of our proposed method.