Temporal convolution neural network for food and drink intake recognition

Eating difficulties are a prevalent issue within the elderly population, leading to weight loss and malnutrition. Likewise a poor diet is considered a confounding factor for developing chronic diseases and functional limitations. Given the above issues, alongside the current advances in computational intelligence achieved with the use Convolutional Neural Networks (CNNs), this paper proposes a wrist-worn tri-axial accelerometer-based food and drink intake monitoring system by combining an adaptive segmentation technique and a CNN model using 1-dimensional (1D) temporal convolutions. First, potential eating or drinking gestures are identified by the use of the adaptive segmentation technique. Once identified, the resultant gesture set is used to train the network for the recognition of four commonly occurring dietary gestures (drinking, using a spoon, using a fork and using the hand to take a bite). The problem is tackled as a 5-class classification model where the remaining class is composed by all the irrelevant gestures. The results reported, with an average per-class classification accuracy of 97.15%, suggest the system proposed is a viable solution for food and drink intake monitoring as well as a great contribution to the field of pervasive computing in support of independent living.