Performing Activity Recognition on Unscripted Culinary Activities in a Living Lab

点击翻译按钮获取中文摘要

Culinary activities, including the preparation and consumption of sustenance, have thepotential to provide valuable insights into the mental and physical health of individuals.Behavioural habits or variances which may become evident through monitoring of theseprocesses can be symptomatic of conditions such as diabetes, obesity, depression anddementia. Therefore, autonomous monitoring and understanding of activities that take placein the kitchen is an important goal that if achieved could improve the lives of many people inthe comfort of their own homes.A key part of autonomous home monitoring systems is activity recognition, as this allows formachine understanding of human situations and actions. This project focuses onknowledge-based activity recognition, since this has more flexibility for smaller data sets.The sequences used to create and test our model will be recorded in the Bristol SPHEREhouse, which already has an assortment of cameras and sensor installed. Sequences willbe unscripted, allowing for a wide range of interesting and challenging behaviours.A Computational State Space Model will form the foundation of the recognition system, withmarginal filtering used to determine the most likely state at each moment in time. During thecourse of the project, elements of this will be enhanced or upgraded in order to improvemodel performance. Action recognition will be improved through the integration of visualdata through Convolutional Neural Networks, which will be compared to current featurebased approaches in terms of speed and accuracy of classification. In addition, recognitionwill be improved to ensure that classification is invariant to different locations and systemconfigurations. Activity recognition will be upgraded to allow the model to process activitiesperformed by multiple individuals in the same environments. Additional improvements willbe made to improve efficiency by reducing the overall state space through simple logicalchanges. Data permitting, data-driven approaches to activity recognition will also be trialledand compared to previous methods.The implications of this project are wider reaching than it's use in the kitchen. Applications toother environments or situations could have a wide range of different uses in many differentsectors. Mainly, this project also builds upon and enhances the work being done by theSPHERE project.