Target Tracking Research Based On Deep Siamese Convolutional Neural Networks

Moving object detection and tracking is currently one of the important research topics in the field of computer vision.It has a wide range of applications in fields such as video surveillance,autonomous driving,human-computer interaction,and air defense early warning.At present,many research results have been achieved in object tracking.However,in some complex and changeable scenes,due to the partial occlusion of the object,geometric deformation,rapid motion,scale transformation and other factors,the accuracy and robustness of the existing algorithms for tracking the object are not good.Therefore,object tracking is still a very challenging task.In recent years,with the hot development of deep learning in computer vision,deep learning has gradually become the mainstream method in the field of object tracking.With the powerful feature extraction capabilities of convolutional neural networks and the introduction of various optimization algorithms on the neural network,the target tracking method based on deep learning has also improved the tracking performance to a new level.However,on the one hand,the existing networks are trained using shallow networks such as AlexNet,and do not fully utilize the advantages of deep convolutional neural networks.On the other hand,for complex scenes such as occlusion and background interference,the network does not highlight important regional information,resulting in the tracker is not robust.This article improves on this,the main research content and innovation are as follows:(1)This paper proposes a object tracking algorithm based on improved CIResNet and feature fusion.Aiming at the insufficient discriminative ability of the existing siamese network to extract features from the shallow network and the tracking drift problem of the deep neural network,a deep convolutional neural network based on CIResNet-34 was constructed as a method for extracting features of the backbone network through experimental comparison of the settings of the CIR unit in the backbone network structure and the control of the network receptive field and step size;This paper also uses two feature fusion methods to solve the problem of insufficient apparent information in the deep network,thereby further improving the performance of the tracker.(2)This paper proposes a object tracking algorithm in complex scenes based on attention mechanism.In view of the problem that the tracker designed in this article does not work well for some complex scenes,by introducing the spatial attention mechanism module and the channel attention mechanism module to assign different weights to the convolution features,the network model has a higher ability to discriminate the object;Aiming at the problem of unbalanced positive and negative samples in the generation of sample candidate frames in the object tracking process,by using the method of improving the loss function,on the basis of the original loss function,the balance factor is introduced to improve the loss function,reducing the influence of the negative sample training process.In the end,the algorithm proposed in this paper was verified on the OTB100,VOT2016 and 2017 data sets.In the OTB100 data set,the success rate of the algorithm in this paper ranked first among all comparison algorithms,which was 4.4% higher than Siam RPN.In the VOT2016 and VOT2017 data sets,the average overlap expectation of the algorithm in this paper also achieved the first place among all the comparison algorithms,which is 1.5% and 1.9% higher than the second place Siam RPN,respectively.