| Object tracking is an indispensable technology in the life of modern technological applications,which requires tracker that can continuously and accurately estimate the position and size of a target.Therefore,it is of great importance in the analysis of anomalous behavior and national defense security.However,it is still difficult to perform effective tracking due to many factors such as occlusion and background interference.With the increasing maturity of deep learning techniques,single-target tracking methods based on deep siamese neural networks have been greatly improved.Due to its excellent performance,it is also gradually gaining attention in tracking technology.This technology can help workers identify specific targets in various scenes,which is important for reducing eye strain and work stress.Deep siamese networks obtain results by performing similarity measures on two inputs,and applying them to the field of single-target tracking can effectively improve the accuracy and success rate of tracking.However,to address the problem of its insufficient ability to characterize the semantic information of extracted features,this paper investigates and improves the SiamRPN algorithm in the direction of improving feature extraction by introducing two modules of deep residual network and attention mechanism,as follows:(1)SiamFC,SiamRPN and other algorithms use shallow networks as backbone networks to extract target features,in order to feel deeper semantic information and achieve the purpose of distinguishing background and target,the original backbone network is replaced by deep residual network ResNet-18,and in order to reduce the computational complexity of the model and better match the subsequent regional candidate networks,the last rollup group and fully connected layer are fully dropped;(2)ECA module focusing on channel attention is introduced to suppress noise and useless information and improve the quality of network feature extraction and the discriminative ability of the tracker;(3)the contribution of replacing the feature extraction backbone and adding ECA module to improve the algorithm is verified through a set of comparison experiments and two sets of ablation experiments,and the performance of SiamRPN-ER algorithm is verified.In this paper,we use OTB2015,VOT2016 and VOT2018 public competition datasets to validate the SiamRPN-ER algorithm,and prove that the SiamRPN-ER algorithm in tracking accuracy and success rate by comparing and analyzing with five deep siamese network-based single object tracking algorithms such as SiamFC and SiamRPN on various evaluation metrics performance has been improved. |
