An Improved Algorithm For Object Tracking Based On And-Or Graph

Object tracking has always been an important subject in the field of computer vision.The problem of object tracking widely exists in many disciplines such as imaging,computer science,criminal investigation,artificial intelligence and so on.It has great research value and significance.This thesis proposes an object tracking algorithm,which is based on the AOG object tracking algorithm proposed by Tianfu Wu et al.It improves the original algorithm in many ways,improving the accuracy and success rate of the original algorithm on the TB50 / 100 dataset.The original algorithm has great advantages in object tracking.It uses AOG to model the structural relationship between the object and its components,combines the modeling and training ideas of DPM,and uses a series of multiple image description features at groundtruth image features as the final object feature,the experiment of the original algorithm also proves that it has good tracking performance.In addition,because of the characteristics of AOG top-down modeling,the object model has the advantages of interpretation and scalability.However,this article found in the experiments that due to the original algorithm's search range is inflexible and the model is not updated in time,it will lead to tracking loss or model drift in tracking some complex videos.Therefore,this article flexibly limits the target search range and updates the object in time The model and the process and composition of the training samples are optimized to achieve the purpose of improving the tracking performance of the original algorithm.This thesis aims at improving the performance of the algorithm by improving the following four different aspects,including(1)using the speed control improvement module to dynamically control the search area,and limiting the search range according to the speed of the object to make the model more flexible.And because the change of the search range implicitly excludes the interference object,thereby improving the overall tracking performance;(2)use the score filtering improvement module to actively control the model update process,and analyze the parse tree score to actively control the model update,reducing The disadvantages of passive update are solved;(3)The positive and negative example overlap rate improvement module is used to dynamically control the positive and negative example overlap rates of the training set,so that the model can make corresponding changes to the complexity of the data set,which improves the model for various scenarios.(4)Use the negative sample optimization module to optimize the negative training data of the algorithm,so that the algorithm has better discriminative ability for object tracking in complex scenes,thereby improving the overall tracking performance.In addition,the four improved algorithms are independent of each other and do not interfere with each other,and can be combined to achieve better results.Finally,in this thesis,experiments are performed on the TB50 / 100 data set,and the comparison charts of tracking accuracy and success rates are drawn respectively.The experiments prove that the four different improvements mentioned above can improve the tracking performance of the original algorithm to a certain extent,and ultimately better.To complete the object tracking detection task.