基于梯度纹理特征的稀疏跟踪方法  

Visual Tracking with Sparse Representation Based on Gradient Texture Feature

作　　者：胡昭华[1,2] 鞠蓉 李容月 何军[1] 

机构地区：[1]南京信息工程大学电子与信息工程学院,南京210044 [2]南京信息工程大学江苏省大气环境与装备技术协同创新中心,南京210044

出　　处：《小型微型计算机系统》2016年第8期1857-1861,共5页Journal of Chinese Computer Systems

基　　金：国家自然科学青年基金项目(61203273)资助;江苏省自然科学基金青年基金项目(BK20141004)资助;江苏省普通高校自然科学研究项目(11KJB510009;14KJB510019)资助;江苏高校"信息与通信工程"优势学科Ⅱ期建设工程项目资助;江苏省大学生实践创新训练计划项目(201410300021)资助

摘　　要：现有大多稀疏跟踪方法往往采用基于灰度信息的特征描述目标,但灰度特征一般无法处理目标的旋转、姿态变化等问题.针对此问题,本文采用一种新的基于平均曲率和高斯曲率的梯度纹理特征的稀疏跟踪方法.梯度纹理特征具有旋转不变性,能很好地解决图像的旋转以及光照变化等问题.在粒子滤波框架下,提取目标模板和候选粒子的梯度纹理特征,用字典模板对每个候选粒子进行稀疏表示,求解稀疏系数并实现重构,将具有最小重构误差的粒子作为跟踪结果.同时采用增量子空间学习的方法对模板进行更新,有效抑制跟踪漂移并能处理目标遮挡问题.与现有跟踪器的大量对比实验表明,本文算法在各种挑战因素下有明显优势.Most available sparse tracking methods applied gray information to describe objects which can＇t handle the problems caused by rotation and pose variation. In order to solve the problems,a novel sparse tracking method is presented to solve the problem caused by rotation and illumination variation, which is based on the gradient texture feature from mean and Gaussian curvature of the image. Firstly, the gradient texture features of the templates and candidates are extracted. Secondly, each candidate is represented sparsely as a linear combination of the dictionary template atoms and is reconstructed by the sparse coefficient. Finally, the candidate with the mini- mum reconstruction error is considered as the tracking result in the particle filter framework. Furthermore, the dictionary template is updated dynamically by incremental subspace learning method, which can effectively adapt the templates to the appearance changes of the object with less drift and reduce occlusion influence. Numerous comprehensive experimental evaluations with state-of-the-art algorithms on challenging sequences demonstrate the superiority of the proposed tracking algorithm.

关 键 词：梯度纹理特征 目标跟踪 高斯曲率 平均曲率 稀疏表示 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]