Für eine korrekte Darstellung dieser Seite benötigen Sie einen XHTML-standardkonformen Browser, der die Darstellung von CSS-Dateien zulässt.


Recognition of Activities of Daily Living: How to Extract High-level Activities from Raw Sensor Data

Prof. Dr. Kimiaki ShirahamaKindai University

Date: 21/06/2018

Time: 11:00

Room: H-F 010 

Person's capability of Activities of Daily Living (ADLs, e.g., "Dressing", "Preparing foods", "Cleaning a room" etc.) is a useful medical measurement to evaluate his/her functional decline as well as the possibility of his/her independent living. Considering the ageing population and declining birthrate in the current society, there is a great demand to develop a system that utilises wearable and environmental sensors to continuously monitor ADLs of an elderly person.  However, each ADL occurs in a long time interval including a lot of irrelevant movements. In other words, different occurrences of the same ADL are characterised by a huge intra-class variance of sensor data. Thus, it is difficult to directly recognise ADLs from raw sensor data.

To overcome this, we have developed a system that recognises ADLs based on the following two layers: The first layer targets at recognising short-term elemental activities, called atomic activities, like “Walking”, “Bending”, “Cleaning a floor”, “Squatting” and “Cleaning a surface” etc. Compared to ADLs, atomic activities happen in much shorter time intervals and have a much smaller intra-class variance, so they are much easier to recognise. The second layer deduces ADLs (also called composite activities) by considering temporal relations among recognised atomic activities. Although occurrences of an ADL are dissimilar in terms of raw sensor data, we can assume that they are similar in terms of atomic activities. For example, when cleaning a room, a person is mainly walking, bending and squatting to clean floors using a vacuum cleaner, and sometimes standing to cleaning surfaces like windows and tables. The talk presents our development of atomic and composite activity recognition methods in the above-mentioned two-layers framework.