Culture as a Sensor? A Novel Perspective on Human Activity Recognition
Culture as a Sensor? A Novel Perspective on Human Activity Recognition
复制标题
DOI：
10.1007/s12369-019-00590-3
复制
发表时间：
2019-12-01
影响因子：
4.7
通讯作者：
Sgorbissa, Antonio
中科院分区：
文献类型：
--
作者：
Chiang, Ting-Chia;Bruno, Barbara;Sgorbissa, Antonio
人类活动识别（HAR）系统致力于在由一个或多个传感器提供的感觉流中识别与多个先验定义的感兴趣活动的执行相关的部分，所述传感器被定位成使得它们可以监视人的动作。提高人类活动识别系统的性能是一个长期的研究目标：解决方案包括更精确的传感器，更复杂的算法，用于从感官数据中提取和分析相关信息，以及增强感官分析与有关感兴趣活动执行的一般或个人特定知识。后一种趋势，在这篇文章中，我们提出了关联和增强的感官数据分析与文化信息，这可以被看作是一个人的特定信息的估计，减轻了负担的一个漫长/复杂的设置阶段。我们提出了一种文化感知的人类活动识别系统，该系统将最先进的、文化感知的HAR系统提供的识别响应与有关在不同文化中最有可能执行活动的地点和时间的特定文化信息相关联，并编码在本体中。文化信息与文化不知道的反应的合并是由贝叶斯网络，其概率方法允许避免刻板的表示。离线和在线进行的实验，使用公寓中的移动的机器人获得的图像，表明文化感知HAR系统始终优于文化不知道的HAR系统。
Human Activity Recognition (HAR) systems are devoted to identifying, amidst the sensory stream provided by one or more sensors located so that they can monitor the actions of a person, portions related to the execution of a number of a-priori defined activities of interest. Improving the performance of systems for Human Activity Recognition is a long-standing research goal: solutions include more accurate sensors, more sophisticated algorithms for the extraction and analysis of relevant information from the sensory data, and the enhancement of the sensory analysis with general or person-specific knowledge about the execution of the activities of interest. Following the latter trend, in this article we propose the association and enhancement of the sensory data analysis with cultural information, that can be seen as an estimate of person-specific information, relieved of the burden of a long/complex setup phase. We propose a culture-aware Human Activity Recognition system which associates the recognition response provided by a state-of-the-art, culture-unaware HAR system with culture-specific information about where and when activities are most likely performed in different cultures, encoded in an ontology. The merging of the cultural information with the culture-unaware responses is done by a Bayesian Network, whose probabilistic approach allows for avoiding stereotypical representations. Experiments performed offline and online, using images acquired by a mobile robot in an apartment, show that the culture-aware HAR system consistently outperforms the culture-unaware HAR system.