Research On The Improvement Of Struck Object Tracking Algorithm

Vision is the most important way for people to obtain external information.With the sustained and rapid development of machine learning technology,computer vision plays a more and more i mportant role in the military field,video surveillance,traffic safety,human-computer interaction,s ports and other fields.The final purpose of the development of computer vision is to use all the fea tures of the human eye andcombine some of the advantages of the computer itself to assist human.As an important research direction of computer vision,target tracking attracts a lot of scholars to study it.This paper will first divide the process of tracking algorithm into four modules: Feature Extractor,Motion Model,Observation Model and Update of Model,and then introduce each module in detail.The following part introduced the Struck tracking algorithm and test database TB.Through the analysis of the Struck tracking algorithm,we try to improve the two modules of it: Motion Model and Feature Extraction.The different Motion Models and Feature Extractions are used to replace the corresponding modules of the Struck tracking algorithm to do experiments,and the experimental results are analyzed and compared.The specific research contents are as follows:(1)Firstly,this paper introduced several commonly used feature extraction algorithms and feature fusion,then substitute a total of 6 feature extraction algorithms into the feature extraction module of Stuck tracking algorithm to do simulation experiment,and the experimental results was compared and analyzed,which showed that all kinds of feature extraction methods have their own advantages and disadvantages.They are suitable for dealing with different situations.In addition,the feature fusion algorithm can achieve the complementary effect in some cases to get better results than single feature.(2)On the basis of previous experiments,this paper proposed an Adaptive Discontinuous Frame Multi-scale Sliding Window search algorithm for the problem that the Sliding Window,the motion model of Struck tracking algorithm,cannot respond to the changes of the target size,which always causing the failure of tracking.The algorithm solves the problem of the response to the change of the target size in the tracking process.In the simulate experiment,with the Haar feature,which has the best performance in the previous experiments,this paper used every single frame,every N frames and Adaptive Discontinuous Frame to do Multi-scale Sliding Window searching in Motion Model part,and compared them with origin Struck and Particle Filter Struck.The qualitative and quantitative analysis of the experiment results shows the effectiveness and robustness of the proposed algorithm.Finally,the 6 kind of features with the 2 kind of motion models are tested,and the conclusion is drawn that the best performance of the tracking algorithm is Struck tracking algorithm based on the Haar feature and the Adaptive Discontinuous Frame Multi-scale Sliding Window.