| The development of target tracking technology is of great significance to the fields of intelligent driving,intelligent monitoring and human-computer interaction.This thesis introduces the background of target tracking technology and the principles and implementation of several typical trackers based on deep learning algorithms,focuses on new algorithms based on existing deep learning target tracking algorithms and verifies their effectiveness through experiments,and finally briefly discusses the future development trend of target tracking algorithms.Although the field of target tracking is now in a mature stage of development and existing algorithms have achieved good performance in major international test sets,the shortcoming is that existing tracking algorithms still face great challenges in terms of accuracy and precision when faced with interference factors such as occlusion,deformation,insufficient illumination and complex backgrounds.In the field of intelligence,deep learning is evolving rapidly and its role in tracking is becoming increasingly evident.With the help of machine learning algorithms,target tracking algorithms that combine deep learning and twin network structures are emerging,offering significant performance advantages.They offer significant improvements in tracking performance compared to traditional methods.With the wind of deep learning,domestic and international research on deep learningbased target tracking has been going on non-stop and is gradually solving existing challenges.The target tracking algorithm based on twin network structure has a long term development potential,and it can achieve an effective balance between effectiveness and accuracy in the tracking process,which can improve the accuracy and precision of the algorithm in target tracking to a certain extent.By combining existing deep learning models with existing techniques,this thesis aims to improve the accuracy and precision of target tracking algorithms on generic datasets and enhance the ability of tracking models to cope with complex situations by addressing the problems of complex backgrounds,targets that are not easily identified,low feature relevance and the perceived field size of network models in target tracking.Based on this,this thesis focuses on target tracking algorithms based on deep learning and twin network structure,with the following main research results.(1)In this thesis,we propose a new PSF(Point Spread Function)image enhancement algorithm to address the problems of complex and similar backgrounds in target tracking.By weakening the weight of the background in the network training,we optimize the tracking process in which the tracker tends to treat similar backgrounds as tracking targets and other phenomena,thus improving the tracking accuracy of the tracker when dealing with complex backgrounds.Firstly,two copies of the input from the target and search branches of the twin network structure are prepared,one is passed through the PSF image enhancement algorithm,and this part of the image will be turned into a grey-scale map;the other is normally entered into the first layer of convolution for feature extraction,and then the first image is also passed through the first layer of convolution of the backbone network for feature extraction,at this time there will be four feature maps in the first layer of the backbone network respectively.The four feature maps are then concatenated on the same branch,i.e.feature fusion.The fused features are then fed into the second layer of the network for the next step of feature extraction and training.This thesis also addresses the problem that existing deep learning algorithms commonly use SGD(Stochastic Gradient Descent)as a gradient descent optimiser.Although SGD optimisers have advantages in terms of convergence,they are less generalisable and require the researcher to give the appropriate momentum for the model.The aim of this thesis is to fuse a new gradient descent algorithm,the AdaBelief optimiser,which has the characteristics of fast convergence,good stability and high generalisability,and the optimiser adaptively adjusts the learning rate,which uses Belief to adjust the step size,and flexibly adjusts the step size by the difference between the actual value and the predicted value This helps the model to be more stable during the training process.(2)The attention mechanism is a commonly used method to enhance image weights and adjust the focus of network features.In this thesis,ABD-Net attention is added to the twin tracking model to enhance the learning of the network model for targets with low feature relevance.In particular,the ABD-Net attention contains the Attention Module and the singular value decomposition regularisation as well as the spatial and channel information of the features,while the Mask about the application can be obtained by the learning of the context.This module also proposes an orthogonal regularisation approach that incorporates spatial regularisation and weight regularisation,aiming to reduce the relevance of features that can be matched by direct matching.Subsequently,this thesis uses null convolution in the RPN(Region Proposal Network)module,which aims to improve the perceptual field in the RPN module and help the regression and classification layers to deeply mine the relevance of low features.A total of two feasible solutions are proposed,combining a target tracking model via the PSF image enhancement algorithm and the AdaBelief optimiser,and the second one combining the ABD-Net attention mechanism and the null convolution to address low feature relevance and enrich contextual information.It is shown through experiments that both tracking algorithms perform well in the OTB and VOT datasets. |
