Sistema para detecção de pessoas usando sinal do WiFi para ambientes internos

Abstract

Detecting people has various applications across multiple sectors. Imagine a smart home that adjusts lighting, heating, and air conditioning based on the number of occupants in a room. Shopping centers can analyze customer behavior by tracking how many people visit specific areas and how long they stay. As an alternative to traditional detection methods, techniques based on radio frequency (RF) have emerged. However, some require individuals to carry active devices that emit RF signals, which can be inconvenient. Others involve the deployment of dedicated wireless sensor networks, which are expensive and difficult to maintain. Fortunately, the widespread presence of WiFi networks in homes and offices offers a new approach. WiFi signals propagate throughout the environment, reflecting and scattering off objects and people. This inherent interaction carries information about the surroundings and their occupants, incorporating relevant data on geometry and electromagnetic characteristics that can be leveraged for sensing purposes. This is possible because WiFi communication involves dynamic signal adjustments and transmission/reception methods through metrics obtained from the Channel State Information (CSI), which captures characteristics of the communication channel. WiFi Sensing has gained significant attention from both the scientific community and industry in recent years, mainly due to its strong potential for extracting information about indoor environments without relying on physical sensors or additional technologies. In this context, the present undergraduate thesis focused on leveraging WiFi signals for human detection. A new method was proposed for detecting people in indoor environments using WiFi signals and CSI-derived information, combined with supervised machine learning techniques, without requiring additional sensors or wearable devices. Validations were conducted using both a publicly available dataset widely used in the literature and a custom dataset built in controlled environments, using ESP32 prototyping boards with the ESP32 CSI Tool project. The results demonstrated the feasibility of the proposed method, achieving accuracies above 97% in the best evaluated scenarios.