Recognising Activities at Home: Digital and Human Sensors
Recognising Activities at Home: Digital and Human Sensors
复制标题
DOI：
10.1145/3102304.3102321
复制
发表时间：
2017-07
期刊：
影响因子：
--
通讯作者：
Jie Jiang;Riccardo Pozza;Kristrún Gunnarsdóttir;Nigel Gilbert;K. Moessner
中科院分区：
文献类型：
--
作者：
Jie Jiang;Riccardo Pozza;Kristrún Gunnarsdóttir;Nigel Gilbert;K. Moessner
家里有什么活动？什么时候发生，持续多久，谁参与其中？在关于家庭的社会研究中，提出这些问题是很重要的，例如，研究与能源有关的做法、辅助生活安排以及家庭和家庭生活的各个方面。寻求答案的常见方法是自我报告，这是由研究人员引起的（访谈，问卷调查，调查）或非引起的（时间使用日记）。纵向观察也很常见，但所有这些方法对参与者和研究人员来说都是昂贵和耗时的。数字传感器的进步可能提供另一种选择。例如，温度、湿度和光传感器报告活动发生的物理环境，而能源监视器报告用于协助活动的电气设备的信息。使用传感器生成的数据来进行活动识别是研究家庭活动的一种潜在的非常强大的手段。然而，我们如何量化我们在传感器生成的数据中检测到的数据与我们从自我报告的数据中了解到的数据之间的一致性，特别是非挑衅数据？为了给出部分答案，我们在一个家庭中进行了一项试验，我们从一套传感器中收集数据，以及从两个居住者之一完成的时间使用日记中收集数据。对于使用传感器生成数据的活动识别，我们研究了均值移位聚类和变化点检测的应用，用于构建用于训练隐马尔可夫模型的特征。此外，我们提出了一种基于Levenshtein距离的方法来评估传感器数据中检测到的活动与参与者报告的活动之间的一致性。最后，我们分析了识别不同类型活动的不同特征的使用。
What activities take place at home? When do they occur, for how long do they last and who is involved? Asking such questions is important in social research on households, e.g., to study energy-related practices, assisted living arrangements and various aspects of family and home life. Common ways of seeking the answers rest on self-reporting which is provoked by researchers (interviews, questionnaires, surveys) or non-provoked (time use diaries). Longitudinal observations are also common, but all of these methods are expensive and time-consuming for both the participants and the researchers. The advances of digital sensors may provide an alternative. For example, temperature, humidity and light sensors report on the physical environment where activities occur, while energy monitors report information on the electrical devices that are used to assist the activities. Using sensor-generated data for the purposes of activity recognition is potentially a very powerful means to study activities at home. However, how can we quantify the agreement between what we detect in sensor-generated data and what we know from self-reported data, especially non-provoked data? To give a partial answer, we conduct a trial in a household in which we collect data from a suite of sensors, as well as from a time use diary completed by one of the two occupants. For activity recognition using sensor-generated data, we investigate the application of mean shift clustering and change points detection for constructing features that are used to train a Hidden Markov Model. Furthermore, we propose a method for agreement evaluation between the activities detected in the sensor data and that reported by the participants based on the Levenshtein distance. Finally, we analyse the use of different features for recognising different types of activities.