TY  - BOOK
AU  - Armenta García,Jesús Albany
AU  - González Navarro,Félix Fernando
AU  - Caro Gutiérrez,Jesús
ED  - Universidad Autónoma de Baja California.
TI  - A system for human activity recognition using WI-FI CSI on embedded device 
AV  - TK5103.2 A75 2025
PY  - 2025///
CY  - Mexicali, Baja California
KW  - Sistemas de comunicación inalámbrica
KW  - Tesis y disertaciones académicas
KW  - Sistemas de comunicación inalámbrica: innovaciones tecnológicas
KW  - Sistemas de comunicación inalámbrica: medidas de seguridad
N1  - Maestría y Doctorado en Ciencias e Ingeniería; Tesis (Doctorado)--Universidad Autónoma de Baja California, Instituto de Ingeniería, Mexicali, 2025.; Incluye referencias bibliográficas
N2  - This thesis addresses the limitations of current safety alert systems
that rely on wearable sensors by proposing a novel Human Activity
Recognition system based on Wi-Fi Channel State Information (CSI)
and designed following the Edge Computing paradigm, where data
processing occurs directly on the device responsible for data collection.
The research is structured in three phases.
First, an Structured Literature Review (SLR) was conducted to assess

state-of-the-art methods for Wi-Fi CSI collection and processing, re-
vealing two critical gaps: the lack of a user-friendly, high-performance

CSI collection tool and the absence of edge-based implementations for
near real-time Wi-Fi sensing with low-cost devices.
To address these gaps, this work first introduces the ESP32 CSI Web
Collecting Tool, a tool for collecting Wi-Fi CSI from ESP32 devices.
Experimental results demonstrate that the proposed tool outperforms
existing ESP32-based solutions, achieving a packet rate of up to 85
packets per second (compared to 30 packets per second in prior tools)

when transmitting CSI data to a computer via a USB port. Addition-
ally, the tool supports local data storage via an SD card and real-time

transmission to external devices through GPIO pins, enhancing its
versatility for edge applications.

Leveraging this tool, a lightweight Deep Learning model was opti-
mized for deployment on the resource-constrained ESP32 microcon-
troller. By considering the hardware limitations of the implemen-
tation device, the model achieves an overall classification accuracy

of 90.65% with an inference time of 232 ms, enabling near real-time
functioning at the edge.

The contributions of this work include a high-performance, open-
source CSI collection tool for the ESP32 platform and the first known

implementation of an embedded system leveraging Deep Learning
model for Wi-Fi HAR on an ESP32 device, which demonstrates the
feasibility of near real-time, device-free safety monitoring
UR  - https://drive.google.com/file/d/1h_rEWMm1V4xU7wOIjI7De-fvtkvrteeu/view?usp=sharing
ER  -