Method Of Object Tracking Based On Likelihood Matrix And Attention Model

In recent years,visual object tracking,as an important research direction in the field of computer vision,has become one of the current research hotspots.Although object tracking technology has made great progress with the development of deep learning and the improvement of computer performance,it still faces many problems.The moving target has scale changes and deformations,and there are complex backgrounds,similar targets,and lighting changes around the moving target which lead the tracker to drift and fail.Therefore,solving the tracking problem and improving the accuracy of the tracker is the researchers' direction.On the basis of reading a great quantity of domestic literature,this thesis first introduces the related background,significance and current status of visual object tracking.Secondly,a brief introduction of the relevant knowledge about related filters,depth trackers and attention mechanisms is made.Combining the advantages of the correlation filter,the deep Siamese network and attention mechanism,this thesis proposes the object tracking model based on correlation filter with the target likelihood matrix,the object tracking model based on the module with channel and space attention in serial on the basis of Siamese network.To prove the reliability of the proposed method,a deep vision object tracking prototype system based on target likelihood matrix and attention module is designed.The main research work of this thesis is as follows:1)A visual object tracking model based on filter with a target likelihood matrix is proposed.The Siamese network based on deep learning visual object tracking is simple and can take both accuracy and speed into account.However,this model cannot handle the visual object tracking problems such as deformation and background clutter because the tracking template is not updated,and the filter can realize fast calculation.Learning real-time templates through filters can reduce system overhead,but it also introduces boundary effect problems.Therefore,a method is proposed to assist the filter to learn the filtering template through the target likelihood matrix in the deep Siamese network.The target likelihood matrix is composed of a weight matrix and a space matrix.It not only describes the statistical information of the target but also the position information of the target,which can suppress the boundary effect.Experimental results show that this method can obtain more precise position and target shape,thereby improving the performance of the visual object tracker.2)A target tracking model based on the module of channel and spatial attention in series is proposed.The task of learning filter template is based on feature map,while the deep features learned from shallow network are not sufficiently discriminable and the network treat the target and the background area equivalently.When the target changes drastically,it easily leads to tracking drift.At the same time,it increases the calculation work due to the introduction of irrelevant backgrounds,which affects the tracker's accuracy and real-time performance.Therefore,this thesis proposes that introducing a model of channel and spatial attention in series to the feature extraction layer of Siamese network.This model effectively fuses the detailed information of the bottom layer of the neural network and the high-level semantic information from the channel and space dimension,then highlights the salient areas of the target.The experimental results show that this method can improve the discriminability of the feature expression of the two branches of the Siamese network at the same time,thereby improving the accuracy of the visual tracker.3)Using the joint programming technology of matlab and C ++,with the help of graphical language Qt and computer vision library OpenCV,a prototype system for deep vision object tracking based on target likelihood matrix and attention model is developed.The system consists of three modules: model training,object tracking,and video playing.The system has a simple interface and good interactivity,which verifies the usability of the proposed visual tracking algorithm.