Tracking-Learning-Detection TLD解析四 - 扩展及局限

最新推荐文章于 2020-12-03 12:09:30 发布

原创最新推荐文章于 2020-12-03 12:09:30 发布 · 4.8k 阅读

4 ·

CC 4.0 BY-SA版权

文章标签：

#TLD #learning #tracking #detection #论文

论文推演专栏收录该内容

6 篇文章

订阅专栏

本文探讨了TLD算法在人脸跟踪领域的局限性，并介绍了FaceTLD如何通过引入验证器来提升跟踪精度。FaceTLD在检测器基础上增加了验证器以匹配特定人脸，使用2bitBP特征与最近邻分类器，相较于原始TLD，在特定场景下显著提升了跟踪效果。

本篇文章主要基于：Face-tld: Tracking-learning-detection applied to faces. ICIP, 2010

Limitations:

TLD does not perform well in caseof fullout of plane rotation.In that case the, the Median-Flow trackerdrifts away from the target and can bere-initialized onlyif the object reappears withappearance seen/learned before.
TLD trainsonly the detector andthe trackerstay fixed. As a result the tracker makes always the same errors. An interestingextension would be to trainalso the tracking component.
TLD currently tracks a singleobject. Multi-target tracking opensinteresting questions howto jointly train the models and sharefeatures in order toscale.
Current version does not performwell for articulated objects such as pedestrians. In case ofrestricted scenarios, e.g. static camera, an interestingextension of TLD would beto include background subtraction in order to improve the tracking capabilities.

Extensions:

因为TLD是面向未知物体开发的一个长期跟踪系统。而FaceTLD则只跟踪面部，故FaceTLD与TLD相比将原来的检测器扩展为检测器与验证器。通用检测器通过线下训练来检测人脸（frontal face），在线训练的验证器来将检测结果匹配到具体某张脸。Validator以第一帧中的人脸初始化，之后会加入跟踪得到的结果更新。使用的特征叫做 2 bit binary pattern(2bitBP),使用最近邻分类器。 FaceTLD在检测人脸时比TLD效果改善在两个视频上显示提高了20%左右，因其更具有针对性。而且由于现在检测人脸的检测器很多不必再使用随机森林在线训练，送到验证器中的人脸数量也大大减少，所以时间消耗也较少。