**1. Introduction**

Recently, with the rapid development of wearable sensor environments, a human activity recognition (HAR) with consistently collected daily data and various learning classifiers has become a popular issue: a vision-based recognition using a camera [1], recognition of five daily activities with acceleration data from a mobile phone and vital signs [2], and recognition with acceleration data from a chest-wearable device [3], and so on. However, despite mature studies and analyses on simple actions, like walking, standing, or sitting, complex activities that are composed of many low-level contexts and show various sensor patterns with respect to the background contexts have not been deeply studied yet [4].

In this paper, we propose a method which recognizes the eating activities in real life. Providing automatically information related with eating activities, such as the time and duration of eating activities, is crucial for healthcare management systems for people, in general, automatic monitoring for patients, such as diabetics, whose eating activities should be carefully managed, or the elderly who live alone, and so on. Although there are already plentiful studies recognizing simple eating and other daily activities, their approach did not catch the very large variety of activities in real life and are, therefore, difficult to extend to real situations. Eating activities could be a very complicated activity

to recognize using sensors, especially with limited low-power sensors, as it could have different sensor patterns with respect to different backgrounds and spatial/temporal contexts. In this paper, we propose a probabilistic method, especially the Bayesian network, which is based on the idea that those complexities might be handled better with a probabilistic approach.

The paper is organized as follows: In Section 2, we provide some analyses to show the complexity of eating activities based on the real-life logging, and specify requirements to deal with those issues. In Section 3, we explore HAR-related works using low-level sensor data, and related theories analyzing components of human activity. In Section 4, we explain how to construct Bayesian networks in further detail, and verify their realistic usefulness in a variety of angles in Section 5. Finally, Section 6 concludes the paper and discusses future works.
