计算机工程与应用 ›› 2015, Vol. 51 ›› Issue (13): 15-24.

• 博士论坛 • 上一篇    下一篇

基于多任务混合噪声分布模型表示的视频跟踪

亚森江·木沙1,2,赵春霞1   

  1. 1.南京理工大学 计算机科学与工程学院,南京 210094
    2.新疆大学 机械工程学院,乌鲁木齐 830046
  • 出版日期:2015-07-01 发布日期:2015-06-30

Multi-task mixed noise distributed model representation for visual object tracking

Yasin Musa1,2, ZHAO Chunxia1   

  1. 1.School of Computer Science and Engineering, Nanjing University of Science and Technology, Nanjing 210094, China
    2.School of Mechanical Engineering, Xinjiang University, Urumqi 830046, China
  • Online:2015-07-01 Published:2015-06-30

摘要: 视觉跟踪中,目标信息是不确定的非线性变化过程。随时间和空间而变化的复杂动态数据中学习出较为精确的目标模板并用它来线性表示候选样本外观模型,从而使跟踪器较好地适应跟踪作业中内在或外在因素所引起的目标外观变化是视觉目标跟踪研究的重点。提出一种新颖的多任务混合噪声分布模型表示的视频跟踪算法,将候选样本外观模型假设为由一组目标模板和最小重构误差组成的多任务线性回归问题。利用经典的增量主成分分析法从高维数据中学习出一组低维子空间基向量(模板正样本),并在线实时采样一些特殊的负样本加以扩充目标模板,再利用扩充后的新模板和独立同分布的高斯-拉普拉斯混合噪声来线性拟合当前时刻的候选目标外观模型,最后计算候选样本和真实目标之间的最大似然度,从而准确捕捉当前时刻的真实目标。在一些公认测试视频上的实验结果表明,该算法将能够在线学习较为精准的目标模板并定期更新目标在不同状态时的特殊信息,使得跟踪器始终保持最佳的状态,从而良好地适应不断发生变化的视觉信息(姿态、光照、遮挡、尺度、背景扰乱及运动模糊等),表现出更好的鲁棒性能。

关键词: 目标跟踪, 多任务, 增量子空间学习, 实时字典扩充

Abstract: In object tracking, target state is an uncertain nonlinear task and it needs to process some large-scale dynamic data. The object representation needs to learn more precise models from the complicated data in order to adopt the variations caused by intrinsic or extrinsic factors during tracking. This paper proposes a novel object template learning scheme and efficient multi-task structured sparse representation method. In this framework, firstly it learns some low dimensional subspace basis vectors(positive template) from tracked data using Incremental Principal Component Analysis and augments them with some instantaneous negative samples generated from every frame. Then the observation model can be represented linearly by using object templates with some additive i.i.d.(independently and identically distributed) Gaussian-
Laplacian noise assumption. Finally, the tracker estimates the most accurate object state among particles based on the highest likelihood between candidate samples and the real object information. Meanwhile an efficient template update rule is used for adopting the variations of object such that caused by illumination, occlusion, pose, background cluttering and motion blur etc. Extensive experiments have been carried out to validate the new algorithm that each state covers a specific property of the object in every task, then the tracker is designed by associating them and well represents the observation model. Experimental results show that the proposed algorithm performs favorably against several state-of-the-art trackers.

Key words: object tracking, multi-task, incremental subspace learning, template expansion