Human activity recognition using deep learning

Details

Supervisors

Pecheur, Charles

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil en informatique, à finalité spécialisée

Abstract

This thesis explores the detection of "Freezing of Gait" (FOG) events within the broader domain of Human Activity Recognition (HAR), with a focus on the application of deep learning models to process time series data from accelerometer sensors. The primary objective was to develop a robust model capable of accurately detecting FOG events in Parkinson's patients, a critical challenge in the field. To address this, a novel hybrid model integrating Convolutional Neural Networks (CNNs) for spatial feature extraction and Gated Recurrent Units (GRUs) for temporal sequence modeling was proposed. The model was trained and evaluated on a Kaggle competition dataset specifically designed for FOG detection. The CNN-GRU model achieved a notable mAP score of 0.4017, demonstrating superiority over standalone CNN and GRU models. However, it exhibited limitations in detecting minority events and did not fully match the performance of transformer-based models, which are better suited for capturing complex temporal dependencies. The hybrid model's generalizability was also assessed using the UCI-HAR dataset, where it showed reasonable performance, reinforcing its potential applicability in other HAR tasks. These findings suggest that while the CNN-GRU model is promising, further refinement and exploration may be necessary to fully address the challenges of FOG detection.