Collecting big datasets of human activity one checkin at a time

A variety of cutting edge applications for mobile phones exploit the availability of phone sensors to accurately infer the user activity and location to offer more effective services. To validate and evaluate these new applications, appropriate and extensive datasets are needed: in particular, large sets of traces of sensor data (accelerometer, GPS, micro- phone, etc.), labelled with corresponding user activities. So far, such traces have only been collected in short-lived, small-scale setups. The primary reason for this is the difficulty in establishing accurate ground truth information outside the laboratory. Here, we present our vision of a system for large-scale sensor data capturing, leveraging all sensors of todays smart phones, with the aim of generating a large dataset that is augmented with appropriate ground-truth information. The primary challenges that we address consider the energy cost on the mobile device and the incentives for users to keep running the system on their device for longer. We argue for leveraging the concept of the checkin - as successfully introduced in online social networks (e.g. Foursquare) - for collecting activity and context related datasets. With a checkin, a user deliberately provides a small piece of data about their behaviour while enabling the system to adjust sensing and data collection around important activities.

In this work we present up2, a mobile app letting users check in to their current activity (e.g., "waiting for the bus", "riding a bicycle", "having dinner"). After a checkin, we use the phone's sensors (GPS, accelerometer, microphone, etc.) to gather data about the user's activity and surrounding. This makes up2 a valuable tool for research in sensor based activity detection.