Target Tracking Method Based On Convolution Neural Network

Target tracking is an important research direction in the field of computer vision.Due to target tracking technology has important application value in many fields,target tracking is a hot spot in computer vision research at present.Based on the analysis and research of the target tracking methods in recent years,this thesis comes to two key research ideas:Firstly,the new target tracking algorithm generally focuses on how to extract the robust features of the target through the convolutional neural network and focuses on improving the utilization of the feature mapping of each layer of the network.Secondly,the new target tracking algorithm can focus on the subsequent feature processing and target location selection,with the emphasis on improving the use of network output features.In this thesis,the multi-layer convolution network is used as the feature extractor for tracking,and Siamese network is used as the basic network framework for tracking.Aiming at the current multi-layer convolution network structure,residual network block is used as the basic structure of feature extraction network,and then the input of target feature and search area feature is divided into high-frequency feature and low-frequency feature in different channels for processing,and the low-frequency feature output results of each layer are exchanged.Finally,the tracking performance has inproved.In this thesis,we propose a target tracking method of mesh feature point fusion.This method is based on Siam RPN.Aiming at the offset regression branch of Siam RPN,the localization ability of the regression branch is optimized by introducing the method of mesh point feature fusion and by means of full convolution localization.The new tracking method is more powerful and has better performance.The proposed method in this thesis are used to test the tracking effect on the mainstream tracking data sets OTB and VOT.The performance of two proposed tracking algorithms have improved on these data sets.