(Invited) Mobile Activity Recognition and Healthcare Application

Sozo Inoue,
Proceedings of International Conference on Informatics, Electronics & Vision
(Not Available)
(Not Available)
Dhaka, Bangladesh
Recent deployment of smart phones equipped with accelerometers will make it possible to recognize activities of the users. If human activity can be objectively measured, we can expect various applications. For example,
lifestyle aspects can be quantified and used for prevention of lifestyle-related diseases. In that of agriculture, farmers can improve efficiency by automatically obtaining their own activity record.
Moreover, in more domain specific application such as nursing management, nursing activity will be quantified and optimized for various type of process in hospitals.

However, existing work for activity recognition has the following problems. 1) it does not consider sequential activities, thereby the accuracy of recognition will decrease around the period of transferring one activity to another. 2) It only targets on simple types of activities. More complex activities, which are sometimes more abstracted or composed of simple activities, are not tried to be recognized.

In our research, we take the approach of dividing the activity recognition problem into: 1) firstly applying a segmentation method which outputs multi-level segments, and, 2) upon them, applying a traditional activity recognition method, adjusting the level of segmentation adaptively.

To confirm the validity of our approach. we pursued an experiment of gathering accelerometer data of real nursing in a hospital.
To collect activity data efficiently, we used a large-scale activity gathering system named ALKAN with smart phones.
We asked nurses working at a large hospital to place iPodTouches in to their breast pockets with a roughly fixed direction.
Large-scale nursing data were collected using ALKAN with either annotated rehearsal activities or real nursing actives, for 41 activity classes.
We have gathered 2808 activity instances with 1.9GB for the former, and 282 days by 24 hours with 17.4GB for the latter.
During the measurement of accelerometer data, we also recorded the sound by the same iPodTouches.

In this talk, we demonstrate several application of the method to automatically segment real nursing data, to help annotation with a suggestion from the history, and to improve activity recognition accuracy.

Data Files