Research And Implementation Of Online Object Tracking Algrorithm Based On Visual Adaptation

Object tracking is an important research blanch and basic problem in the fields of computer vision. Meanwhile, it’s also the important foundation of many high-level applications, which are based on the analysis of object tracking, such as behavior recognition, behavior analysis, etc. But also for intelligent monitoring, human-computer interaction, transportation management, medical image processing, etc. provide basic data fields, so the issue in-depth study has important significance and practical value. Learning to adapt to the use of the domain theory to solve the target tracking problem in computer vision, because it can use the line of vision of knowledge and a lot of pictures so that the process of updating the classifier single use sample data is no longer current frame, a wide range of domestic and foreign scholars in recent years by attention.This thesis is based on the analysis of traditional object tracking algorithm, and focused on online object tracking based on domain adaption learning theory. The concrete work is as follow:1. We described and analyzed the basic classical core part of the online object tracking algorithm, namely, trained a classifier to distinguish the background and th e foreground, including P-N learning theory of TLD object tracking algorithm, multiple instance learning of MIL tracing algorithm and the online boosting algorithm. Through study of the core of the online target tracking algorithm found its shortcomings, combined with the actual situation of the target tracking based on necessity and feasibility of proposed domain adaptation learning objectives to track.2. We put forward the adaptive online target tracking algorithm based on the instance. In this paper, the target tracking online training learning classifier as domain adaptive learning problems, previous image sequence tracking target instance knowledge transfer to the first t-1 frame image. According to these instances of knowledge with the first t-1 frame of target information online training a classifier, and the method of using tracking- check to target the first t frame.3. We offer online target tracking algorithm based on classifier to adapt. Referring to the methods of target detection based on classifier to adapt to the target tracking algorithm using online by offline training a large number of image data sets a classifier. At the same time, we use the current tracking information online video frames training and update an online classifier. According to the offline classifier and online classifier to detect the image sequence of tracking, and then according to the initialization time on target characteristics of the selected target.4. Based on the modified sparse coding domain adaptive target tracking algorithm. We through offline SIFT feature descriptor, lugging a large amount of images and based on the SIFT feature descriptor training a set of complete visual dictionary characterization of offline image visual information. In the process of tracking, we through the sparse coding with multiple scale max- pooling method using the visual knowledge obtained from the offline learning(complete set of visual dictionary) the characteristics of the construction target expression. Positive samples and negative samples extracted from the first frame initialization classifier, also used in the bayesian framework of target motion state of the target tracking task is derived. In order to overcome in the process of tracking the target and the background of cha nge and the moderate classifier appeared in the process of updating the goal of the drift problem, we timely update the classifier and maintaining the original classifier as a detector. For each video frame, target observation model is by the classifier structure of joint detector and online.In summary, this thesis detailed analysis and implementation of three different types of online-based target tracking algorithm. Meanwhile, we conducted experiments on different database validation and comparison with other methods. Experimental results show that you can achieve visual knowledge of a lot of pictures or non-current video frame to migrate to online object tracking, and this migration is beneficial visual knowledge.