Bite detection and differentiation using templates of wrist motion

We introduce a new algorithm of bite detection during an eating activity based on template matching. The algorithm uses a template to model the motion of the wrist over a 6-second window centered on the time when a person takes a bite. We also determine if different types of bites (for example food vs. drink, or using different types of utensils) have different wrist motion templates. This method is implemented on 22,383 bites and 5 different types of templates are built. We then describe a method to recognize different types of bites using the set of templates. The obtained accuracy was 46%. Finally, we describe a method to detect bites using the set of templates and compare its accuracy to the original threshold-based algorithm. We get positive predictive value of 75 % and true positive rate of 47% found across all bites.