Research On Target Tracking Algorithms Based On Feature Matching Of Region

As an important research topic in computer vision,target tracking has great practical significance and application value.It is widely used in intelligent surveillance,human-computer interaction,medical treatment,unmanned driving and other fields.Though significant progress has been made in the last few years in target tracking,the diversity and complexity of practical application scenarios,it brings great challenges to target tracking technology.Therefore,the research of target tracking technology is still a topic of great scientific value and challenge.The most challenging thing of target tracking technology is that it has to ensure the adaptability of the algorithm to various complex scenarios and taking into account the performance and realtime of the algorithm.This thesis studies the target features and tracking model on the basis of relevant work,and focuses on developing efficient,accurate,robust tracking algorithms.In this thesis,we study the Mean shift tracking framework in the generative tracking method and the correlation filtering tracking framework in the discriminant tracking method.The main research work and contributions are summarized as follows:(1)In the framework of Mean shift tracking,background interference and occlusion in tracking scenarios can cause performance and robustness degradation of tracking algorithms.Aiming at solving the problem,a new target tracking algorithm based on multi-channel extraction of AHLBP(Adaptive Haar Local Binary Pattern)texture features is proposed.After analyzing the defects of HLBP(Haar Local Binary Pattern)texture in expressing image features,we took into account the local visual information of the image and proposed an Adaptive Haar Local Binary Pattern texture extraction method with adaptive threshold.The method can improve the effective expression of texture information to the internal structure information of the image.At the same time,in order to obtain rich and deep texture information to describe the target,we separated the gray image of each color channel from the color space of the target area,and then extracted the texture information separately.An accurate and proper target model is established by using the texture features of the target.Finally,the target tracking is realized under the framework of Mean shift.Experiments show that the algorithm can make full use of texture features and maintain good tracking performance and robustness in complex scenarios such as background interference and target occlusion.(2)In the KCF(Kernel Correlation Filter)tracking algorithm,the target location method may be disturbed by similar objects or certain noise leading to performance degradation.A novel tracking algorithm is proposed to address the question.Firstly,based on the principle of target location in the tracking model of KCF algorithm,we analyze the cause of target location failure under multi-peak response graph.Then,a local multi-peak location strategy based on the core idea of selecting effective candidate target location and re-detection is proposed to improve the accuracy of target location.In addition,considering that the response map can reflect the correlation between target and candidate targets,we introduce two parameters update criteria and propose a new model parameter update strategy to reduce the corruption of inaccurate positioning results to the model.A large number of experiments compared with the mainstream tracking algorithms show that the proposed algorithm has good adaptability to target tracking tasks in complex scenes,and has better tracking performance and real-time performance.