University of Twente Student Theses

Login

Automatic annotation of the cooking process

Baardewijk, Jan Ubbo van (2018) Automatic annotation of the cooking process.

[img] PDF
1MB
Abstract:Automatic annotation of actions using accelerometer data would make the analysis of everyday, spontaneous activities much easier. It could create a context when trying to relate physiology to experience and it could lead to extracting new features. The goal of this thesis is to examine how different actions can be classified using inertial measurement unit data and to see how well a model that is trained in a structured experiment could be used to classify actions in a more naturalistic setting. This is done in the context of food preparation, because the project ’Quantified consumer’ is used as the framework. Therefore, the data from ’Quantified consumer’ is used as the data set for this experiment. We chose an approach where we split the data in windows of 1 second and classify each window. As the ground truth we used a selection of the data set for which we could be sure about the performed action. We performed three different classification tasks, each with a different number of actions. k-nearest neighbor with k=5 was used as classification method with up to 6 different features. The information from neighboring windows were taken into account for each of the windows. Classification between action and standing still had a high performance, classification with more classes gave the highest results on the dry cooking session data. A model trained on dry cooking data and run on real cooking data performed almost as well as a model trained on real cooking data, provided that the dry cooking set is balanced between classes. This study presented a way to evaluate different action recognition models without requiring manual annotation of videos. To improve the results, we could look at increasing the number of neighboring windows that is taken into account or better look at the used features. A different classification method like deep learning hopefully improves recognition performance as well as a more user-specific approach.
Item Type:Essay (Master)
Clients:
TNO, Soesterberg, Nederland
Faculty:EEMCS: Electrical Engineering, Mathematics and Computer Science
Subject:54 computer science
Programme:Interaction Technology MSc (60030)
Link to this item:http://purl.utwente.nl/essays/74615
Export this item as:BibTeX
EndNote
HTML Citation
Reference Manager

 

Repository Staff Only: item control page