Human detection for multiple pose by boosted randomized trees
Human detection for multiple pose by boosted randomized trees
复制标题
DOI：
10.1109/acpr.2011.6166541
复制
发表时间：
2011-11
期刊：
影响因子：
--
通讯作者：
Takayoshi Yamashita;Yuji Yamauchi;H. Fujiyoshi
中科院分区：
文献类型：
--
作者：
Takayoshi Yamashita;Yuji Yamauchi;H. Fujiyoshi
在本文中，我们提出了一种鲁棒的姿势不变人体检测框架。大多数现有的人体检测框架都采取站立姿势，需要单独的检测器来支持其他人体姿势。我们提出了一个具有分层树结构的单一框架，可以检测各种姿势。所提出的方法基于随机树。候选特征的选择如下所示，为了学习高性能决策树，1）决策树的每个节点根据类似然性用类进行约束，2）通过联合Boosting为上述类预先选择有效特征，3）根据这些有效特征随机生成候选特征。根据 1) 和 2)，可以训练根节点来区分人类和背景，并且可以训练叶节点来识别特定姿势。对“购物场景”中出现的各种姿势进行了性能比较，所提出的方法优于其他基于联合提升、随机树和 Adatree 的多类分类器。
In this paper we propose a robust pose invariant human detection framework. Most of the existing human detection frameworks assume a standing posture and needing a separate detectors for supporting other human postures. We propose a single framework with a hierarchical tree structure that can detect various poses. The proposed method is based on Randomized trees. Candidate features are selected as shown below, to learn high performing decision trees, 1)each node of the decision tree is constrained with classes based on class likelihood, 2)effective features are pre-selected with Joint Boosting for the above classes, 3)the candidate features are randomly generated based on these effective features. From 1) and 2), the root nodes can be trained for discriminating the human from the background, and leaf nodes can be trained for specific poses. Performance comparison was performed for various poses that arise for a “shopping scenario”, and the proposed method outperformed other multi-class classifiers based on Joint boosting, Randomized trees and Adatree.