Improving activity recognition without sensor data: a comparison study of time use surveys

Abstract

Wearable sensing systems, through their proximity with their user, can be used to automatically infer the wearer's activity to obtain detailed information on availability, behavioural patterns and health. For this purpose, classifiers need to be designed and evaluated with sufficient training data from these sensors and from a representative set of users, which requires starting this procedure from scratch for every new sensing system and set of activities. To alleviate this procedure and optimize classification performance, the use of time use surveys has been suggested: These large databases contain typically several days worth of detailed activity information from a large population of hundreds of thousands of participants. This paper uses a strategy first suggested by [16] that utilizes time use diaries in an activity recognition method. We offer a comparison of the aforementioned North-American data with a large European database, showing that although there are several cultural differences, certain important features are shared between both regions. By cross-validating across the 5160 households in this new data with activity episodes of 13798 individuals, especially distinctive features turn out to be time and participant's location. Additionally, we identify for 11 different activities which features are most suited to be used for later on activity recognition.