2016-GRSL-Infrared Target Tracking Based on Robust Low-Rank Sparse Learning

He2015InfraredTT.md

2016 - GRSL - Infrared Target Tracking Based on Robust Low-Rank Sparse Learning

Abstract

最近很流行的 low-rank sparse tracker

• Assumption：
  • exploiting low-rank constraints to capture the underlying structure of candidate particles.
    • 注意哦，是用 low-rank constraints 刻画 candidate particles
  • uses simple sparse error to account for occlusion and noise measured by the L1-norm， which is assumed to be following the Laplacian distribution
    • sparse error 是用来刻画 occlusion 和 noise
• 不足：
  • However, this Laplacian assumption may not be accurate to describe complex corruptions
• 本文的改进：
  • seek for the maximum-likelihood estimation solution of the residuals in the tracking framework

Introduction

Related Work

• AO - 2014 - Real-time infrared target tracking based on l1 minimization and compressive features
  • combine the real-time advantages of the compressive tracking and the robustness of the $L_1$ tracker
• IPT - 2013 - Track infrared point targets based on projection coefficient templates and non-linear correlation combined with Kalman prediction
  • a tracking framework based on template matching combined with Kalman prediction

Low-Rank Sparse Tracker (LRST)

• 提出论文
  • IJCV - 2015 - Robust visual tracking via consistent low-rank sparse learning
  • ECCV - 2012 - Low-rank sparse learning for robust visual tracking
    • 前者应该是后者的期刊扩展版
• Method
  • formulates object tracking as a sparse and low-rank representation problem in particle filter framework
• 不足
  • simply assuming experientially that the errors follow the Laplacian distribution and employs $L_1$-norm regularization to specify the errors.
  • Obviously, noise in most cases is much more complicated when severe occlusion, heavy clutter, and distortion occur during tracking process
  • 简单的说就是 1 范数不足以刻画复杂的杂波干扰
  • this assumptions would lead to tracking drift and target loss.

本文 Motivation

• 受 CVPR - 2011 - Robust sparse coding for face recognition 启发
• employs the maximum-likelihood estimation (MLE) principle to deal with complex noise by estimating the real distribution of noise.
• Subsequently, an IR target tracking method based on the robust low-rank sparse representation is presented under particle filter framework, namely, a robust LRST (RLRST), which could address the tracking drift problem and be robust
  • 也就是说，本文方法叫做 RLRST 咯
  • 无非就是用了 MLE 来估计噪声，然后再嵌到原来的 LRST framework 里

Robust low-rank sparse representation

LRST framework

• transforms the tracking task into a sparse and low-rank representation problem under particle filter framework,

• aims to seek sparse and low-rank coefficients $S$ of the object and the background templates dictionary $D$ in

  (1)

  

  • S 既是 sparse 也是 low-rank 的啊，有意思

• allows for a sparse error $E$ to contaminate S in order to deal with noise and occlusion.

• adding the low-rank constraint to the coefficients of candidate particles 的好处

  • because the representations with regard to the dictionary of different particle samples are expected to be similar.
  • 意思是说稀疏表示系数相似，所以他们组成的矩阵就是低秩的了，然后本来就是稀疏表示系数自然也是稀疏的，所以最后的情况是既低秩又稀疏
  • 因为 the property of nonlocal self-correlation of candidate particles， this similarity is even more obvious in IR images

本文的 Motivation

• 作者认为 From the viewpoint of maximum likelihood, 公式（1）中的 noise E 被认定是 sparse 的或者说是 following Laplacian distribution，不符合实际情况，比如图 1 中树的遮挡
  • 这里我就纳闷呢？ 人脸识别的那些论文里，不是认为 1 范数已经够鲁棒能够来处理遮挡了么？为什么作者还这么 1 范数不够呢？
  • ​

@article{He2015InfraredTT,
  title={Infrared target tracking based on robust low-rank sparse learning},
  author={He, Yujie and Li, Min and Zhang, Jinli and Yao, Junping},
  journal={IEEE Geoscience and Remote Sensing Letters},
  volume={13},
  number={2},
  pages={232--236},
  year={2015},
  publisher={IEEE}
}