O. Amft - Activity and context recognition: challenges towards truly smart health assistant systems
A recent uptake of human activity recognition methods in various research fields demonstrates the increasing interest in sensor-based context/situation awareness. Nevertheless fundamental challenges remain yet to be solved, which are imperative to address complex real-world problems. In particular, these issues are related to the diversity of human activities and pattern variability, as it is observed in behaviour monitoring of patients or health-at-risk individuals. Behaviour monitoring is key to understand a patient's situation and to behavioural medicine, where coaching plays an essential role to complement interventions and rehabilitation.
In this talk, I will introduce current research towards personalised assistant systems that use ubiquitous distributed sensing and processing based on embedded and wireless technologies, as well as machine learning and recognition. The latter serves to identify situations and user activities automatically from multi-modal sensor data. Techniques will be discussed to derive behavioural and context knowledge, including recognition stacks and their semantic context abstraction and context-dependent reconfiguration. Application examples will be discussed to illustrate technology and algorithms developments, including assistants for physical rehabilitation, daily routine assessment, and automatic dietary monitoring. Finally, I will provide an outlook on future developments in this area and currently starting projects at the ACTLab at TU Eindhoven, which is concerned with multi-modal activity and context recognition.