Feasibility Study for Food Intake Tasks Recognition Based on Smart Glasses

In this exploratory study 13 adult test subjects have performed different food intake tasks while wearing a three axis accelerometer mounted at a temple of glasses. Two different algorithms for task recognition have been applied and compared. The retrospective data processing leads to better task recognition results when the frequency range of 50 Hz to 100 Hz is analysed within accelerometer signal recordings. A straightforward variance threshold algorithm is able to detect the intake of crunchy food with a sensitivity of 0.923 and a specificity of 0.991. Furthermore it identifies calm behaviour of test subjects with a sensitivity of 0.923 and a specificity of 0.914. Drinking from a cup can be detected with a sensitivity of 0.846 and a specificity of 0.986 by application of a k-nearest neighbour classification approach. By demonstrating the feasibility to recognise different food intake tasks based on analysing the acceleration of glasses, the door for employing accelerometer related data from smart glasses also in specific domains, such as dietary profiling, has been opened.