Organic Computing Techniques for Run-Time Self-Adaptation of Ubiquitous, Multi-Modal Activity Recognition Systems

无处不在的活动和上下文识别(AR)旨在将简单传感器提供的信息转化为关于人类活动和环境状况的高级知识。在过去的十年里，研究人员已经证明了识别活动和情况的原则上的可行性，从维护任务的步骤到家庭中的日常活动，再到运动和社交互动。当今最先进的方法的一个主要局限性是，它们大多假定在系统设计时准确定义的系统配置，这些配置在运行时保持固定。因此，对于每种应用，用户需要将特定的传感器放置在环境中某些明确定义的位置和他的身体上。然后，信号处理链的所有阶段(从信号调节到特征选择到分类)都针对具体的配置和任务进行定制设计。虽然这样的静态运行时间设置可以在受控的实验室条件下得到保证，但在现实世界的设置中必须考虑传感器退出和出现新传感器的可能性。在这个提案中，我们将开发新的有机计算(OC)技术，以促进无处不在的活动和上下文识别系统的自我修复(当传感器退出时)和自我改进(当新传感器出现时)。具体地说，我们将开发一个分层的观察者/控制器体系结构，其中被观测和控制(SuOC)的系统是人类(S)在一个智能的、传感器启用的环境中。底层(反应层)可以看作是标准的基于增强现实的上下文敏感系统的蓝图。适配层使系统能够使用新的传感器信息自主地--或通过零星的人类反馈半自主地--改进反应层的分类器，或者使传感器丢失。一般来说，自主适应方法不能保证总是导致改进，在特殊情况下，它们甚至可能导致性能下降。因此，不仅在适配层而且在对长期系统演变建模的反射层(顶层)估计和考虑可能的适应的潜在收益和风险，以确保系统配置的持续修改导致长期改进，而不是整个系统的无限性能降低。在我们的方法中，我们结合和扩展了机器学习、模式识别以及相关领域(特别是产生式和判别式建模、半监督学习、主动学习和非线性动态系统理论)的方法，开发了新的增强现实技术。我们将在现有的大规模AR数据集上对我们的方法进行评估。

Ubiquitous activity and context recognition (AR) aims at translating information provided by simple sensors into high level knowledge about human activities and the situation in the environment. Over the last decade researchers have shown the principle feasibility of recognizing activities and situations ranging from the steps of a maintenance task, through every day activities at home, to sport and social interactions. A major limitation of today's state of the art approaches is that they mostly assume system configurations exactly defined at the system's design-time that remain fixed at run-time. Thus, for each application, the user needs to place specific sensors at certain well-defined locations in the environment and on his body. All stages of the signal processing chain (from signal conditioning through feature selection to classification) are then custom-designed for the concrete configuration and task. While such static runtime setups can be guaranteed under controlled laboratory conditions, the possibility of sensors dropping out and new sensors appearing must be taken into account in real world settings. In this proposal we will develop new Organic Computing (OC) techniques to facilitate self-healing (when a sensor drops out) and self-improvement (when a new sensor appears) for ubiquitous activity and context recognition systems. Specifically, we will develop a layered Observer/Controller architecture where the System under Observation and Control (SuOC) is (are) human(s) in an intelligent, sensor enabled environment. The bottom layer (reaction layer) can be seen as a blueprint of standard AR based context sensitive systems. The adaptation layer enables the system to improve autonomously -- or semi-autonomously with sporadic human feedback -- the classifier at the reaction layer using the new sensor information or to adapt it a sensor drops out. In general, autonomous adaptation methods cannot guarantee to always lead to an improvement and, in special cases, they can even result in performance degradation. Thus, the potential gains and the risks of a possible adaptation are estimated and considered not only at the adaptation layer, but also at the reflection layer (top layer) that models the long term system evolution to ensure that continuous modifications of the system configuration lead to long term improvement and not to un-bounded performance degradation of the overall system. In our approach we develop new OC techniques for AR by combining and extending methods from Machine Leaning, Pattern Recognition, and related fields (in particular generative and discriminative modeling, semi-supervised learning, active learning, and nonlinear dynamic systems theory). We will evaluate our methods on existing large scale AR data sets.