Research And Implementation Of Single Target Tracking Algorithm Based On Deformable Convolution

Single target tracking has always been a hot topic in the field of machine vision.Single target tracking mainly studies how to use related algorithms to track specific targets for a frame of video image input by the model.The traditional single target tracking algorithm adopts the method of feature extraction,such as HOG feature,LBP feature,histogram feature and so on,and then uses the model matching or correlation filtering algorithm to track the target.In recent years,with the development of deep learning theory and algorithm,there are more and more single target tracking algorithms based on deep learning.Many of them have exceeded the traditional target tracking methods based on feature extraction in performance,but they still face many difficulties and challenges in real life.For example,the results of target tracking are easily affected by the size,rotation and angle of the object,which makes the neural network model less robust to the deformation of the target,and even the tracking disappears.Therefore,although the deep learning model has achieved high accuracy in the standard data set,it is still difficult to apply in real life.In order to solve this problem,on the basis of MDNet target tracking model,the traditional convolution network is transformed into deformable convolution neural network to extract the features of the tracking target,and then a single target tracking algorithm based on deformable convolution is proposed.Based on MDNet single target tracking model and deformable convolution neural network,this paper studied the single target tracking algorithm,and then solved the problem that the tracking accuracy and robustness are not high due to the scale change,rotation,deformation and other phenomena of the single target in the process of tracking.The main contents of this paper are as follows:(1)Problem research and analysis.In this paper,a large number of existing single target tracking algorithms are investigated,and the deformation problems of existing single target tracking are analyzed.(2)A single target tracking algorithm based on deformable convolution is proposed.The proposed algorithm uses the deformable convolution neural network architecture to realize the end-to-end tracking training process.The overall structure is based on the MDNet target tracking algorithm,using the deformable convolution operation instead of the traditional convolution,learning the offset value and sampling in the end-to-end way according to the offset position.At the same time,deformable Ro I pooling is added in order to better select candidate areas and more accurately locate the target location.(3)Experimental verification of the proposed algorithm.In this paper,based on Python and Pytorch deep learning framework,a single target tracking algorithm based on deformable convolution is implemented in Linux system,and the proposed algorithm is verified by using the classic target tracking standard data sets OTB50,OTB100 and VOT2016.The experimental results fully reflect the effectiveness of the single target tracking algorithm based on deformable convolution.The algorithm in this paper has significantly improved the precision,success rate,accuracy and robustness of the test data set.Compared with the single target tracking algorithm based on the traditional convolutional neural network,the algorithm in this paper can solve the problem of deformation easily in the process of single target tracking and show better tracking performance.