HUMAN ACTIVITY RECOGNITION USING WEARABLE DEVICES AND DEEP LEARNING

• Main Author: Parluhutan Hutabarat, James
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/86059
• Institution: Institut Teknologi Bandung

Human Activity Recognition (HAR) based on motion sensors, data such as accelerometer, gyroscope, and magnetometer sensors from wearable devices, gives benefits in the healthcare sector, particularly in patient monitoring in indoor environments. This research aims to develop an optimal algorithm for human activity recognition based on motion sensors, such as Emotibit. The deep learning model was conducted, including Multi-Layer Feedforward Neural Network (MFNN), Convolutional Neural Network (CNN), and Long-Short Term Memory (LSTM). These models were configured using the Optuna framework for hyperparameter optimization. The configuration results in an accuracy of 94.62% for MFNN, 92.90% for CNN, and 97.52% for LSTM. The number of sensor input channels leads to more accurate predictions. Additionally, this research implements deep learning models on edge devices, with optimization solutions utilizing the model weight quantization process. The optimization was applied to the LSTM model with the highest accuracy. The experiment showed that the model size was reduced by up to 90.98% of its original size. This size reduction allows the model to be implemented on edge devices like the Raspberry Pi 4B with limited resources. After implementation, the model achieved an accuracy of 94.61% with an inference time of approximately 96 milliseconds. Lastly, the time from data transmission to activity prediction displayed on the user interface was measured, resulting in a duration of 114 milliseconds.