[Research.001] Practice makes Perfect? The Role of Place-Dependent Expertise in Mobile Robotics - IROS2015 - Talk - Paul Newman

1. Shall we trade everywhere mediocrity for place-specific excellence

  • wherever, whever, whatever the weather
  • Experience-based Navigation (IJRR2013)
  • Experience diversity makes localisation robust
[Research.001] Practice makes Perfect? The Role of Place-Dependent Expertise in Mobile Robotics - IROS2015 - Talk - Paul Newman_第1张图片
Paste_Image.png
  • 三组试验中,系统产生不同的匹配结果。第一组与第三组可以完全匹配,但第二组只有两幅场景完成了匹配

2. Learning never stops but life becomes less surprising

[Research.001] Practice makes Perfect? The Role of Place-Dependent Expertise in Mobile Robotics - IROS2015 - Talk - Paul Newman_第2张图片
Paste_Image.png
  • 新知识在最初被一一记录

  • 随着系统的发展和变化,我们不再面对大量新的知识,而此时模型也就逐渐建立。

  • 中间的那个波动,可以理解为天气的剧烈变化导致场景的不同

3. We have a memory management problem

  • Linegar, Churchill, Newman ICRA2015
[Research.001] Practice makes Perfect? The Role of Place-Dependent Expertise in Mobile Robotics - IROS2015 - Talk - Paul Newman_第3张图片
Paste_Image.png

对阳光的处理

[Research.001] Practice makes Perfect? The Role of Place-Dependent Expertise in Mobile Robotics - IROS2015 - Talk - Paul Newman_第4张图片
Paste_Image.png
  • Colour Explained
  • 用公式描述颜色在相机中的成像
[Research.001] Practice makes Perfect? The Role of Place-Dependent Expertise in Mobile Robotics - IROS2015 - Talk - Paul Newman_第5张图片
Paste_Image.png
[Research.001] Practice makes Perfect? The Role of Place-Dependent Expertise in Mobile Robotics - IROS2015 - Talk - Paul Newman_第6张图片
Paste_Image.png
  • McManus, ICRA2014

对场景大范围变化的处理

[Research.001] Practice makes Perfect? The Role of Place-Dependent Expertise in Mobile Robotics - IROS2015 - Talk - Paul Newman_第7张图片
Paste_Image.png
  • 上图中,同一个场景,发生了很大的变化
[Research.001] Practice makes Perfect? The Role of Place-Dependent Expertise in Mobile Robotics - IROS2015 - Talk - Paul Newman_第8张图片
Paste_Image.png
  • 选择大片特殊的区域,来进行描述(McManus, Newman, RSS2014)
[Research.001] Practice makes Perfect? The Role of Place-Dependent Expertise in Mobile Robotics - IROS2015 - Talk - Paul Newman_第9张图片
Paste_Image.png
  • 将scene signatures用SVM进行分类,因此在offline得到classifiers bank,而这个过程是unsupervised
  • 基本步骤
    • collect images under varying appearances at same location

    • deterministically sample patches of various shapes

    • HOG features -> SVM classifier

具体测试

[Research.001] Practice makes Perfect? The Role of Place-Dependent Expertise in Mobile Robotics - IROS2015 - Talk - Paul Newman_第10张图片
Paste_Image.png
  • 红色和蓝色的训练效果和测试效果很好,偏差不大
  • 但绿色的bounding box出现严重偏差

统计位置进一步筛选

[Research.001] Practice makes Perfect? The Role of Place-Dependent Expertise in Mobile Robotics - IROS2015 - Talk - Paul Newman_第11张图片
Paste_Image.png

一些有意思的结果

[Research.001] Practice makes Perfect? The Role of Place-Dependent Expertise in Mobile Robotics - IROS2015 - Talk - Paul Newman_第12张图片
Paste_Image.png
[Research.001] Practice makes Perfect? The Role of Place-Dependent Expertise in Mobile Robotics - IROS2015 - Talk - Paul Newman_第13张图片
Paste_Image.png

对路灯的处理

  • Nelson Churchill, Newman ICRA2015

Training: Hard Negative Mining

[Research.001] Practice makes Perfect? The Role of Place-Dependent Expertise in Mobile Robotics - IROS2015 - Talk - Paul Newman_第14张图片
Paste_Image.png
  • Object Detector Traning (N Dalal and B. Triggs CVPR20015)
  • 不成功的例子主要是因为对场景信息的不知道

对场景的滤波

  • Hawke, Gurau, Tong, Posner FSR2015
[Research.001] Practice makes Perfect? The Role of Place-Dependent Expertise in Mobile Robotics - IROS2015 - Talk - Paul Newman_第15张图片
Paste_Image.png
  • 在Detector之后,引入场景Filter。例如,在这个场景下,是否会有行人目标的存在。
[Research.001] Practice makes Perfect? The Role of Place-Dependent Expertise in Mobile Robotics - IROS2015 - Talk - Paul Newman_第16张图片
Paste_Image.png

![Upload Paste_Image.png failed. Please try again.]

  • EBC is Handsfree, locally specific hard negative mining

Original Video: https://lecture2go.uni-hamburg.de/veranstaltungen/-/v/18334
转载请注明出处。
联系方式:kongww.nudt [at] gmail [dot] com.

你可能感兴趣的:([Research.001] Practice makes Perfect? The Role of Place-Dependent Expertise in Mobile Robotics - IROS2015 - Talk - Paul Newman)