Unsupervised Long-Term Routine Modelling Using Dynamic Bayesian Networks

Abstract

Routine can be defined as the frequent and regular activity patterns over a specified timescale (e.g. daily/weekly routine). In this work, we capture routine patterns for a single person from long- term visual data using a Dynamic Bayesian Network (DBN). Assuming a person always performs purposeful activities at corresponding locations; spatial, pose and time-of-day information are used as sources of input for routine modelling. We assess variations of the independence assumptions within the DBN model among selected features. Unlike traditional models that are supervisedly trained, we automatically select the number of hidden states for fully unsupervised discovery of a single person's indoor routine. We emphasize unsupervised learning as it is practically unrealistic to obtain ground-truth labels for long term behaviours. The datasets used in this work are long term recordings of non-scripted activities in their native environments, each lasting for six days. The first captures the routine of three individuals in an office kitchen; the second is recorded in a residential kitchen. We evaluate the routine by comparing to ground-truth when present, using exhaustive search to relate discovered patterns to ground-truth ones. We also propose a graphical visualisation to represent and qualitatively evaluate the discovered routine.