Human activity recognition for pervasive interaction

Abstract

This thesis addresses the challenge of computing food preparation context in the kitchen. The automatic recognition of fine-grained human activities and food ingredients is realized through pervasive sensing which we achieve by instrumenting kitchen objects such as knives, spoons, and chopping boards with sensors. Context recognition in the kitchen lies at the heart of a broad range of real-world applications. In particular, activity and food ingredient recognition in the kitchen is an essential component for situated services such as automatic prompting services for cognitively impaired kitchen users and digital situated support for healthier eating interventions. Previous works, however, have addressed the activity recognition problem by exploring high-level-human activities using wearable sensing (i.e. worn sensors on human body) or using technologies that raise privacy concerns (i.e. computer vision). Although such approaches have yielded significant results for a number of activity recognition problems, they are not applicable to our domain of investigation, for which we argue that the technology itself must be genuinely “invisible”, thereby allowing users to perform their activities in a completely natural manner. In this thesis we describe the development of pervasive sensing technologies and algorithms for finegrained human activity and food ingredient recognition in the kitchen. After reviewing previous work on food and activity recognition we present three systems that constitute increasingly sophisticated approaches to the challenge of kitchen context recognition. Two of these systems, Slice&Dice and Classbased Threshold Dynamic Time Warping (CBT-DTW), recognize fine-grained food preparation activities. Slice&Dice is a proof-of-concept application, whereas CBT-DTW is a real-time application that also addresses the problem of recognising unknown activities. The final system, KitchenSense is a real-time context recognition framework that deals with the recognition of a more complex set of activities, and includes the recognition of food ingredients and events in the kitchen. For each system, we describe the prototyping of pervasive sensing technologies, algorithms, as well as real-world experiments and empirical evaluations that validate the proposed solutions.