Automatic Meal Intake Monitoring Using Hidden Markov Models

In the latest years, the number of elderly people that has been living alone and need regular support has highly increased. Meal intake monitoring is a well-known strategy that enables premature detection of health problems. There are several attempts to develop automatic meal intake monitoring systems, but they are inadequate to monitor elderly people at home. In this context, we propose an automatic meal intake monitoring system that helps tracking people's eating behaviors, and is adequate for elderly remote monitoring at home due to its nonintrusive features. The system uses the MS Kinect sensor that provides the coordinates of the user's sitting skeleton during his meals. It analyzes the coordinates, detects eating gestures, and classifies them using Hidden Markov Models (HMMs) to estimate the user's eating behavior. A demonstrative prototype for detection and classification of gestures was implemented and tested. The detection module got satisfactory percentages of sensitivity, having a minimum of 72.7% and a maximum of 90%. The Classification module was tested with 3 proposed methods and the best method had a good average percentage of success (approximately 83%) in the classification of Soup and Main dish; regarding the left hand transporting Liquids, the results were less successful.