Towards the automatic data annotation for human activity recognition based on wearables and BLE beacons

In machine learning, the data annotation process is an essential, but error-prone and time-consuming manual activity, which associates metadata to the samples of a dataset. In the context of Human Activity Recognition (HAR) this generally reflects in a human watching the video recordings of the activities carried out by the target user to assign a label to each video frame. The label can refer, for example, to the nature of the performed activity, or to the time series collected through sensors worn by the user or present in the environment. This paper deals with the automation of the data annotation process in the HAR context by presenting a methodology that (i) maps Bluetooth Low Energy (BLE) beacons distributed in the environment to the locations where a human typically performs activities like eating, cooking, working, and resting, and (ii) associates the data collected by sensors embedded in the smartwatch worn by the user (i.e., acceleration, angular velocity, and magnetometer) to the nearest BLE beacon. In this way, data gathered through the smartwatch are automatically annotated with the human activity associated to the nearest beacon.