Distinguishing Between Human Activities in Real-Time Based on Wearable Sensor Data Using a Low-dimensional Model of Human Movement

人体活动识别是根据佩戴在人体上的加速计和陀螺仪等运动传感器来推断人类活动的过程。这些运动传感器嵌入到身体活动跟踪设备和智能手表中。对人类活动的准确推断为健康监测和促进个人的整体健康提供了许多好处。有许多基于机器学习的方法可以从传感器数据中识别人类活动。然而，在实践中，由于人体类型的广泛、传感器在身体上的位置以及诸如活动速度和强度的变化、传感器噪声、分组丢失、传感器故障等实时不确定性来源的不确定性，当前的方法经常受到精度低下的影响。该奖项支持基础研究，为开发稳健的活动识别方法和算法提供所需的知识。据预测，到2020年，联网的人和设备中将有1万亿个嵌入式传感器。该项目的算法和软件工具可能直接应用于健康监测、老年人护理支持、长期预防和慢性护理以及康复。因此，这项研究的结果将有利于美国的经济和社会。为了促进转型，计划了几项旨在让本科生参与创业的教育举措。这项研究的一个主要目标是开发方法和算法来减少机器学习问题中的不确定性，例如涉及动态数据集的活动识别问题。设计了一个控制理论框架，不仅解决了由于不确定性引起的鲁棒性问题，而且实现了从传感器数据中学习模式的某些统一架构。如果成功，这项工作可以导致新的算法方法来表示、学习和识别非结构化动态数据集中隐藏的低阶模式。除了方法学上的发展，该项目还将推动其他更切实的成果，例如开发反馈粒子滤波的算法和软件，阐明用于表示数据中复杂模式的控制架构和算法，以及开发用于人类活动识别系统的软件工具。

Human Activity Recognition is the process of inferring human activity from motion sensors such as accelerometer and gyroscopes worn on the human body. These motion sensors are embedded in physical activity tracking devices and SmartWatches. Accurate inference of human activity offers many benefits for health monitoring and for promoting overall wellness of an individual. There are many machine learning-based approaches to human activity recognition from sensor data. However, in practice, the current approaches often suffer from poor accuracy on account of uncertainty due to wide range of human body types, sensor locations on the body and real-time sources of uncertainty such as changes in activity speed and intensity, sensor noise, packet drops, sensor failure etc. This award supports fundamental research to provide needed knowledge for the development of robust activity recognition methodology and algorithms. It is projected that there will be a trillion embedded sensors in connected people and devices by 2020. This project's algorithmic and software tools can potentially be directly applied to fitness monitoring, eldercare support, long-term preventive and chronic care, and rehabilitation. Therefore, results from this research will benefit the US economy and society. To promote transitions, several educational initiatives are planned that seek to engage undergraduate students in entrepreneurship. A major objective of the research concerns development of methods and algorithms to mitigate uncertainty in machine learning problems, such as the activity recognition problem, involving dynamic data sets. A control-theoretic framework is planned to not only address the robustness issues due to uncertainty, but also enable certain unified architectures for learning patterns from sensor data. If successful, the work can lead to novel algorithmic approaches to represent, learn and recognize hidden low order patterns in unstructured dynamic data sets. Besides methodological developments, this project will engineer other more tangible outcomes such as the development of algorithms and software for feedback particle filter, enunciation of control architectures and algorithms for representation of complex patterns in data, and development of software tools for the human activity recognition system.