Abstract:
The rapid growth of the elderly population has underscored the importance of accurately
recognizing Activities of Daily Living (ADLs) for effective health monitoring and timely
interventions, particularly in regions such as Korea, where the aging demographic presents distinct challenges. Traditional methods often struggle with processing complex sensor data and
optimizing model performance, especially when dealing with varied activities captured from
multiple body locations. To address these limitations, we propose an advanced deep learning
framework that utilises Long Short-Term Memory (LSTM) networks to analyze time-series
sensor data, enabling the model to capture temporal dependencies and patterns within the
signals. Additionally, we transform these time-series signals into Short-Time Fourier Transform (STFT) spectrogram images, which are subsequently processed using Convolutional
Neural Network (CNN), EfficientNet_B0, and Vision Transformer (ViT) models through
transfer learning techniques. To mitigate the class imbalance inherent in the dataset, the
Synthetic Minority Over-sampling Technique (SMOTE) is employed to generate synthetic
samples for underrepresented classes. Furthermore, a Butterworth low-pass filter is applied to
remove background noise, ensuring higher quality data for model training. The proposed models
exhibited robust performance across multiple sensor placements, with the Efficient- Net_B0
model demonstrating superior accuracy, achieving 99%, 98%, 99%, and 97% for sensor
placements on the bag, belly, hand, and thigh, respectively. These results highlight the potential
of our approach for enhancing ADL recognition in health monitoring systems
