基于语义分层的行为推理框架

PDF下载HTML

Title:

Inference framework for activity recognition based on multiple semantic layers

作者:

聂慧饶, 陶霖密

清华大学 计算机科学与技术系, 北京 100084

Author(s):

NIE Huirao, TAO Linmi

Department of Computer Science and Technology, Tsinghua University, Beijing 100084, China

关键词:

行为理解; 特征行为关系; 环境上下文; 语义分层; 分层推理框架

Keywords:

activity recognition; feature activity relation; environment context; semantic layer; multilayer inference framework

分类号:

TP301.6

DOI:

10.3969/j.issn.1673-4785.201407009

文献标志码:

A

摘要:

人类行为理解是实现“人本计算”模式的基础,其本质在于获取行为的语义,即由动作特征推导人体的行为,需要跨越两者之间的语义鸿沟;为此提出了环境上下文进行隐式建模的方法,并基于此提出了语义分层的行为推理框架,该框架使用了从模糊语义到确定语义的渐近式推理。根据知识将特征合理地分为多个层次,系统则根据当前状态去提取所需要的特征,推理当前可能的候选行为集;并由该候选行为集指导处理模块,更新特征集并进行新一轮的推理,反复迭代至推理完成。 应用提出的环境建模方法和渐近推理框架可以有效地实现行为理解。 使用隐式环境方法可以提高行为理解的准确率;渐近式推理框架可以避免传统推理方法无差别地提取所有特征,从而提升了推理效率。

Abstract:

Human activity recognition is the core of the implementation of human-centered computing(HCC), whose nature is to acquire activities’ semanteme. The basic problem is the semantic gap between observable actions and human activities. They should be bridged by environment context based inference. In this paper, a method is proposed to model the environment context implicitly. Further, a novel semanteme multilayered activity inference framework was presented, which divided the inferring process into 2 stages. One stage used to acquire fuzzy semanteme and another one to acquire accurate semanteme. The feature set was divided into different subsets according to knowledge. The system extracts the corresponding features according to the current state and obtains the possible set of candidate activities that can instruct the system to update the current feature set. Update the features set and infer it, the process continues until the inference is completed. The modeling method and progressive inference framework proposed could handle the activity-recognition problem well. Implicitly modeling the environment context could improve the accuracy of activity recognition. The progressive framework can improve the efficiency by avoiding extracting all features indistinguishably, whose validity was proven in the data set.