Research On Object Tracking Algorithm Based On Convolutional Neural Network And Attention Mechanism

With great development have taken place in computer vision technology,visual object tracking has become an important research field,and is used in video surveillance,human-computer interaction and other analysis tasks.Given the initial state information of the target,the algorithm estimates the scale,position and other state information of the target in the subsequent frames,so as to achieve continuous tracking.Although relevant scholars have proposed various excellent algorithms,there are many tracking interference factors in real tracking situations,which make it more difficult to track the target accurately and effectively.This thesis takes the problem of target tracking as the starting point of algorithm research,and conducts deep research and analysis.The main research contents are as follows:(1)In order to solve the problem that traditional methods cannot extract robust features and have poor generalization ability,an object tracking algorithm based on the similarity measurement of convolutional neural is proposed.First,the number of layers and different layer feature of the convolutional network are analyzed.For the influence of target feature extraction,a convolutional neural network suitable for image feature extraction in tracking tasks is constructed,and then a similarity measurement target tracking framework for tracking and positioning is constructed,on which the initial frame generates multiple templates to improve matching effects.In addition,a regression module is introduced on the basis of matching calculation to further improve the tracking effect of the algorithm.After experimental comparison and verification on the OTB2015 dataset,the robustness of the algorithm has been significantly enhanced,and the tracking effect has also been improved in a variety of tracking scenarios.(2)In order to solve the problem of mismatching and inefficient use of historical frames in similarity measurement target tracking algorithm,a discriminative object tracking method is proposed.The algorithm constructs a target classification decision tracking framework based on the spatial and channel attention mechanism of the convolutional neural network based on the characteristics of the extracted features of the convolutional network.The framework includes a spatial attention module and a channel attention module,which can suppress background interference,extracting discriminative features,improving the ability of the algorithm model to extract features.When tracking,the discriminative module integrates the long-term tracking and shortterm tracking modules of the target organically,effectively using the historical data in the tracking process.In addition,a regression module is used in the framework to further improve the tracking effect of the algorithm.After comparison and verification on the OTB2015 experimental dataset,the algorithm's target tracking accuracy in various tracking scenarios can be significantly improved.