Human Activity Recognition and Localization in Outdoor Environment Using Hybrid Deep Learning

METADEL, ADDIS ADUGNA

dc.contributor.author	METADEL, ADDIS ADUGNA
dc.date.accessioned	2024-12-05T07:41:27Z
dc.date.available	2024-12-05T07:41:27Z
dc.date.issued	2024-03
dc.identifier.uri	http://ir.bdu.edu.et/handle/123456789/16282
dc.description.abstract	This study introduces a Hybrid deep learning approach to Joint Human Activity Recognition and Localization in outdoor environments, leveraging WiFi signal information. The core of our methodology revolves around the implementation of a Hybrid Deep Neural Network, specifically the LSTM-BIGRU architecture, aimed at elevating accuracy in recognizing human activities and estimating their locations. Experiments were precisely conducted on a real-world dataset obtained through the PicoScenes WiFi sensing platform, encompassing both magnitude and phase information. These investigation shown a remarkable improvement in accuracy across both activity recognition and localization tasks. In response to data scarcity challenges, our study employed the Conditional Tabular Generative Adversarial Network (CTGAN) for syntactic data generation, strategically preventing potential leakage of Channel State Information (CSI) data. Moreover, to mitigate the impact of phase fluctuation, we introduced Carrier Frequency Offset (CFO) and Cyclic Shift Delay (CSD) preprocessing techniques. The experimentation was conducted in a Line of Sight (LoS) outdoor setting, where CSI data was collected using the PicoScenes WiFi sensing platform tool, encompassing 4 distinct activities at 10 outdoor locations. WiFi sense, grounded in the notion that human motion influences WiFi signal propagation, leading to unique patterns known as WiFi fingerprints, forms the core of our exploration. These patterns are feed through deep learning algorithms for the identification of human activities and locations. Comparative analyses of experimental results shown our proposed LSTM-BIGRU hybrid deep learning model, achieved a remarkable 99.81% accuracy for activity recognition and 98.93% accuracy for location estimation. While our study contributes valuable insights into human activity recognition and localization it is essential to acknowledge certain limitations. Including, limited transmitter-receiver pair, and different layout impact the generalizability of our findings. Keywords: outdoor environment, human activity, deep learning, WiFi-signals	en_US
dc.language.iso	en_US	en_US
dc.subject	Computer Science	en_US
dc.title	Human Activity Recognition and Localization in Outdoor Environment Using Hybrid Deep Learning	en_US
dc.type	Thesis	en_US