Research On Object Tracking Algorithms Based On Siamese Deep Networks

Visual object tracking technology can analyze the motion parameters and trajectories of specific targets from the video sequence.It provides an important basis and basis for the analysis and understanding of video sequences.It is an important part of computer vision.It has a wide range of applications such as aircraft navigation,automotive automation or assisted driving.However,in practical applications,the tracker is easily affected by a large number of uncertain factors,such as fast motion and interference problems of similar targets,etc.,which makes the accuracy of the algorithm face a huge challenge.Based on the popular deep model tracking framework at this stage,this thesis improves the SiamRPN(Siamese Region Proposal Network)tracking algorithm for these challenges and practical applications.The main contributions of this thesis are as follows:First,a tracking algorithm based on lightweight twin networks and kinetic energy analysis is proposed for the problem of incomplete feature extraction and fast motion of the target when the feature is extracted by the feature extraction module of the tracking algorithm.The algorithm has been researched and innovative in the model construction phase and the algorithm tracking phase.In the model construction phase,the feature extraction module of SiamRPN was replaced with a lightweight network MobileNetV2 with similar parameters but better performance,which improved the tracking algorithm feature extraction The ability of the module to extract features to improve the accuracy of the algorithm;in the algorithm tracking phase,by using the principle of kinetic energy to adjust the search box in the tracking phase,the performance of the tracking algorithm when responding to fast motion challenges is improved.Secondly,in order to solve the problem of network model cannot effectively distinguish similar targets and inaccurate labeling methods,a tracking algorithm based on sample enhancement and target tilt angle prediction was proposed.This method has been researched and innovative in the network training phase and the network model construction phase.During the network training phase,the COCO data set was added to improve the completeness of the data set and improve the ability of the network model to distinguish between similar targets and background targets.During the network model construction phase,this article adds an output to the network model to achieve the target tilt Prediction.Improve the accuracy of the tracking algorithm.In terms of experiments,this thesis has performed experiments and verified the above contributions through well-known and recognized standard datasets in the field of target tracking such as OTB-50,OTB-100,VOT2016,and VOT2018,and made a detailed comparison with the current mainstream algorithms.The algorithm proposed in this thesis has better tracking performance,especially compared with the basic algorithm SiamRPN,which has been improved to varying degrees.