Research On Object Tracking Algorithm Based On Siamese Network

With the continuous development of modern society,computer vision technology is more and more widely used in all aspects of society,playing an increasingly important role in human society.As a research of challenge in the field of computer vision,object tracking has attracted many scholars and achieved a lot of research achievement in recent years.However,due to the influence of some factors such as occlusion and fast motion,etc.,it is still very difficult to design a high-performance and fast tracker that can adapt to all scenarios.However,the emergence of siamese network based object tracking algorithm in recent years has attracted widespread attention because it can achieve higher tracking accuracy while maintaining a higher tracking speed.But compared with other types of tracking algorithms,it obviously shows the problem of insufficient robustness.Therefore,the siamese network based tracking algorithm is studied based on the Siam RPN model.This paper focuses on the robustness of the model to improve the model's ability to cope with complex situations.The work is carried out from the following aspects to achieve this goal.To deal with the problem of insufficient coping ability of the baseline model for complex situations and the problem of insufficient quantity and quality of training data when using deep convolutional neural networks as the backbone network,the training dataset is expanded from the following three aspects: increasing the number of positive samples to improve the generalization ability of the model;taking some data augmentation methods such as color change,random erasure,etc.,so that the feature extraction sub-network can better learn the feature pattern in complex situations;using the bootstrap method to fully excavate the hard samples that easily make the model fail to track or affect the performance of the model,so as to improve the model's ability to cope with complex situations and enhance the model's robustness.Aiming at the problem that the baseline model has good accuracy but poor robustness,Siam DC,the siamese network tracker based on deep convolutional networks,is proposed.The deep residual neural network is used as the backbone network of the siamese network based tracking algorithm.And some adaptive adjustments to the parameters and structure of the residual network are made to make it more suitable for tracking tasks,such as using relatively small network steps and moderately adjusting the receptive field,etc..To improve the poor performance of models using only a single feature,two siamese network based object tracking algorithm are proposed based on two different feature fusion methods.Siam EFF,the siamese network tracker based on early feature fusion,directly merges the features from different levels of the backbone network,specifically,the deep features are taken as the main features and shallow features are fused to improve the accuracy of tracking model.Siam MSF,the siamese network tracker based on multi-solution fusion,fuses tracking results obtained by using multiple levels of features so that multiple tracking results can complement each other to obtain more reliable results.Experiments are conducted on three datasets of VOT2018,VOT2019 and OTB2015 respectively,and the performance of each model was evaluated with corresponding performance indicators.The experimental results show that the performance of Siam EFF and Siam MSF on each data set is significantly improved,the early feature fusion method can improve the tracking accuracy of the model and the multi-solution fusion method can effectively improve the robustness of the model and then improve the performance of the model.