RADAR-BASED HUMAN ACTIVITY RECOGNITION USING DEEP LEARNING AND ITS IMPLEMENTATION ON EDGE DEVICES

• Main Author: Restu Triani, Listi
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/79115
• Institution: Institut Teknologi Bandung

Radar-based Human Activity Recognition explores a wide range of applications, including healthcare, security, and smart home automation. This is due to its non- invasive nature and resilience to environmental conditions. The research compares traditional machine learning methods (K-Nearest Neighbor and Support Vector Machine) with deep learning approaches using Convolutional Neural Network (CNN) and transfer learning. The main goal is to develop an algorithm for radar- based human activity recognition. The proposed approach utilizes CNN, specifically the VGG-19 architecture with transfer learning. This approach allows the integration of feature extraction and classification from radar data in a single learning phase, enhancing learning efficiency. Additionally, the study considers the resource limitations in implementing deep learning models for Human Activity Recognition on edge devices. The proposed solution is an optimized deep learning model, leveraging transfer learning, pruning, and quantization. Evaluation on a public HAR dataset shows that the proposed model significantly reduces the model size (5.30–6.99 times smaller) while maintaining competitive accuracy levels (92.57–94.28%). Reducing computational complexity and storage requirements ensures the feasibility of implementing the model on edge devices with limited resources. The optimized model has also been implemented on the edge device Raspberry Pi 4 Model B, achieving a model accuracy of 94.86% with an inference time of approximately 2 seconds. Furthermore, the Google Coral USB Accelerator has been employed, resulting in improved inference time.