Year: 2024
Type: dissertation
Abstract: Human activity recognition through passive Radio Frequency (RF) sensing systems has gained attention in recent research due to its promising applications across various fields. These passive RF-sensing systems are often characterized by their cost-effectiveness, high portability, reliability, and non-intrusive nature. In this thesis, we present a passive RF-sensing system that utilizes two Software Defined Radio (SDR) devices operating within the Very High Frequency (VHF) range, specifically at a frequency of 104.3 MHz. To examine and analyze human postures, we generated two experimental datasets. The first dataset encompasses the entire frequency range of an ??transmitter, spanning 0–100 kHz, while the second dataset focuses solely on the stereo pilot frequency of 19 kHz. These datasets were utilized to automatically identify three categories: standing, sitting, and a "none" class, which indicates the absence of a person. A Variational Autoencoder (VAE) was employed to extract features from both datasets, alongside the use of statistical features. These features were subsequently classified using a neural network, achieving the highest accuracy of 92% in the offline scenario for the statistical features derived from the complete frequency band dataset, followed by 67% for the VAE features from the same dataset. For the dataset extracted for the stereo pilot, the maximum accuracy achieved was 46%. It was concluded that the information available solely from the stereo pilot is insufficient for effective classification in this context. Additionally, a prototype was developed for real-time human posture classification using the complete frequency band dataset, yielding promising results. The results obtained validate the selected features and underscore the effectiveness of the classification techniques applied in this domain, highlighting the significant potential of passive RF-sensing systems for human posture recognition, which encourages further research in this area.
Author: Oliveira, Ricardo José Batista
Institution: Universidade Nova de Lisboa
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7038638309
