Device-free Human Activity Recognition for the Kitchen

Abstract

Device-free human activity recognition (HAR) is a promising approach for smart environments, but existing methods often experience performance drift with context-specific tasks, particularly in complex settings like kitchens. While prior work explored coarse activities (e.g., walking, sitting), fine-grained, context-specific tasks, such as kitchen activities involving object interactions, remain widely underexplored and are challenging due to signal noise, environmental variability, and limited datasets. In this thesis, we begin by evaluating three HAR approaches for kitchen tasks: deep learning (43% accuracy, limited by small datasets), transfer learning (39%, hindered by domain mismatch), and classical machine learning (68%, which proved best suited for the context-specific data). Given the superior performance of classical learning, its reliance on effective feature selection becomes a critical challenge. Current methods often rely on heuristics to select features, which can limit performance and require domain knowledge. To address this, we propose an optimized method that combines multivariate analysis of variance (MANOVA) with genetic algorithms for discriminative feature selection. This approach improves accuracy to 80% for coarse-grained tasks such as storing food and washing pots. However, fine-grained motions (chopping and slicing potatoes) remain challenging (with average recognition accuracy of 37%), highlighting the limitations of WiFi sensing for subtle activities. The key contributions of this thesis include: (1) an evaluation of HAR methods for kitchen tasks, (2) a novel approach to feature selection based on MANOVA-genetic algorithm to generalize the feature selection process, and (3) a public WiFi-based dataset to advance research in context-aware HAR. These findings highlight the potential of classical learning in context-specific tasks while revealing open challenges in fine-grained activity recognition.