Automatically Segmenting LifeLog Data into Events

Abstract

A personal lifelog of visual information can be very helpful as a human memory aid. The SenseCam, a passively capturing wearable camera, captures an average of 1785 images per day, which equates to over 600000 images per year. So as not to overwhelm users it is necessary to deconstruct this substantial collection of images into digestable chunks of information, i.e. into distinct events or activities. This paper improves on previous work on automatic segmentation of SenseCam images into events by up to 29.2%, primarily through the introduction of intelligent threshold selection techniques, but also through improvements in the selection of normalisation, fusion, and vector distance techniques. Here we use the most extensive dataset ever used in this domain, 271163 images collected by 5 users over a time period of one month with manually groundtruthed events.