Research On Target Tracking Based On Deep Learning

As a branch of computer vision,video target tracking technology has gradually developed into a research hotspot in the field of artificial intelligence.The traditional particle filter tracking algorithm has the advantages of mature framework,good real-time performance and simple feature expression.Because of its non-linear and non-parametric characteristics,it can be well applied to visual tracking research.However,in the process of tracking,it is often disturbed by various environmental changes or changes of the target itself,including its own deformation,complex background,changes in illumination intensity and occlusion.The interference factors mainly affect the construction of moving target feature model in the process of tracking,that is,matching according to the selected features.On the other hand,with the continuous development of deep learning theory and the great advantage of feature representation,the convolutional neural network(CNN)based target tracking algorithm has attracted more and more attention.Convolutional neural network has the advantages of good generalization and strong expressive ability.It can make up for the shortcomings of traditional features to a certain extent and achieve better tracking effect.However,when adjusting the parameters of each layer,the traditional convolution neural structure can easily converge to the local optimum rather than the global optimum.The average method is used in the pooling layer,which is time-consuming and inaccurate for the target location of image blurring.In this paper,the existing problems are studied,the traditional convolution neural structure is improved,and a deep learning target tracking algorithm based on siamese network is proposed.The work of this paper is mainly from the following aspects:1.Improvement of the convolution nerve structure: In view of the problems existing in the traditional convolution nerve structure,this paper improves the convolution nerve structure on the basis of ascending into the study of the convolution nerve structure,adds the normalization layer,and adopts the maximum pooling mode instead of the average pooling mode in the initial structure,which effectively enhances the generalization ability of the network model and effectively improves the operation speed.Based on the above improvements,an ICNN-PF algorithm is proposed.The experimental results show that the ICNN-PF algorithm achieves better tracking effect than the traditional convolutional neural structure algorithm.2.The siamese network is introduced: A target tracking framework based on twin network and improved convolutional neural structure is proposed.It is a symmetrical model for matching in image recognition.Because the number of sample sets that can be used for network training in target tracking is small,the effect is not ideal.The siamese network can learn from the samples to measure the similarity between the data.This similarity judgment can be used to complete the construction of tracking target feature model.The siamese network maps the input,replaces the original minimum variance to measure the input similarity by distance calculation in the mapped target space,takes the predicted position with the highest similarity as the target position tracked in the subsequent frame,and then measures the similarity of the two in the target space by loss function to get the target position.3.Modification of template updating based on weight coefficients: Fixed and intermittent updating of template may skip the best template and lead to inaccurate tracking results;Frame-by-frame updating may alleviate the problem of missing the best template to a certain extent,but the current template may also be disturbed,and frequent updating will take up more resources and increase the amount of computation.Aiming at the problems of accuracy and computational complexity,this paper updates the template by combining different video frame information on the basis of tracking algorithm flow.A set of coefficients is pre-set as weighting factors,and the weights are assigned to different sequences to form a set.The elements in the set participate in the calculation and update of the template.It improves the accuracy and real-time performance.4.On the basis of the two improvements,the DST algorithm is put forward.Compared with the traditional particle filter algorithm and several typical deep learning algorithms,the DST algorithm has better accuracy and robustness,and has higher application value.