Location-Based Activity
Recognition using Relational Markov Networks
Proc. of the International Joint Conference on
Artificial Intelligence (IJCAI-05).
Abstract
In this paper we define a general framework for activity
recognition by building upon and extending Relational Markov
Networks. Using the example of activity recognition from
location data, we show that our model can represent a variety
of features including temporal information such as time of
day, spatial information extracted from geographic databases,
and global constraints such as the number of homes or
workplaces of a person. We develop an efficient inference and
learning technique based on MCMC. Using GPS location data
collected by multiple people we show that the technique can
accurately label a person's activity locations. Furthermore,
we show that it is possible to learn good models from less
data by using priors extracted from other people's data.