Kilby, JeffKhan, Aamer2025-02-142025-02-142024http://hdl.handle.net/10292/18653Each year, about a third of people over 65 years old experience falls, with one out of five sustaining significant injuries such as head trauma or fractures. Even those who do not suffer injuries often struggle to get up unassisted, leading to a fear of falling, loss of confidence, reduced physical activity, poor social interaction, and depression. Monitoring the movements of older adults is crucial to address these issues without compromising their privacy or hindering daily activities. Device-free sensing has emerged as a reliable method for monitoring presence, location, motion, activity, and gestures without requiring attached devices. This technique leverages wireless signals such as Wi-Fi and 4G/5G to detect movements, offering a dependable alternative in conditions where other technologies like vision-based cameras might fail, such as in smoke or darkness. Additionally, device-free sensing respects privacy, making it ideal for pervasive sensing applications, particularly fall detection. This research aims to develop a prototype for elderly care that is effective and affordable. The objective is to create a system that does not require frequent recharging, can promptly alert caregivers about falls, and is within the financial reach of most elderly care centers. The study explores the effectiveness of different machine learning models—Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM) networks, and a hybrid Convolutional LSTM (Conv-LSTM) model—in detecting activities using Channel State Information (CSI) data, both with and without automatic data labeling. Experiments conducted without data labeling revealed that the CNN model, trained on pre-processed CSI data, achieved a validation accuracy of 85.33% and a testing accuracy of 84.52%, though it displayed instability during training. The LSTM model performed slightly better, with a validation accuracy of 87.12% and a testing accuracy of 86.87%, benefiting from its suitability for time-series data. However, the Conv-LSTM model, which combines CNN and LSTM layers, demonstrated superior stability and performance, with a validation accuracy of 89.34% and a testing accuracy of 90.16%, effectively capturing spatial and temporal information from the data. This reliability of the Conv-LSTM model provides a strong foundation for the research findings. When automatic data labeling was applied, all models showed significant performance improvements. The CNN model achieved a validation accuracy of approximately 93% and a testing accuracy of around 92%. The LSTM model reached a validation and testing accuracy of approximately 94%, demonstrating robust training behavior. The Conv-LSTM model outperformed both, achieving exceptional results with a validation accuracy of around 97% and a testing accuracy of approximately 98%. These findings underscore the significant impact of data preprocessing and labeling on the performance of machine learning models in activity detection using CSI data. The proposed automatic data labeling method proved particularly effective, removing extraneous frames and applying heuristic-based labels. This method enhances the models’ ability to interpret data and significantly reduces the complexity and noise within the dataset, leading to more accurate and reliable activity detection. In conclusion, developing and applying effective data labelling and preprocessing techniques are crucial for improving the accuracy and reliability of device-free sensing systems for elderly care. This research demonstrates that with proper data handling, machine learning models can significantly enhance fall detection and other monitoring capabilities, providing a valuable tool for ensuring the safety and well-being of older adults.enThesisOpenAccess