Método para a identificação de gestos de mão usando Wi-Fi sensing e aprendizado de máquina

Abstract

Human-Computer Interaction (HCI) seeks increasingly intuitive interfaces, and hand gesture recognition is a promising field for this evolution. Traditional approaches, such as those based on computer vision or specialized hardware, present limitations in cost, privacy, and technical dependencies such as line of sight. This work explores WiFi Sensing as a robust and non-invasive alternative, using disturbances in Wi-Fi signals, known as Channel State Information (CSI), to classify gestures. The central hypothesis is that machine learning models can achieve high accuracy in gesture classification from CSI data. To validate this hypothesis, a proof of concept was developed with an evolutionary methodology, structured in two experimental scenarios: the first, using a public dataset to establish statistics in a complex environment, and the second, with a proprietary dataset collected in a controlled environment to isolate hand gestures. Data processing evolved from an initial approach that treated each signal subcarrier in isolation to a final methodology that aggregates the entire gesture capture into a single feature vector, preserving the correlation between channels. Model optimization was performed using the Optuna framework, which allowed not only the search for optimized hyperparameters but also an in-depth analysis of the computational cost (Work Effort) of each model. The results demonstrated the superiority of the aggregated capture approach. In the controlled scenario, the SGDClassifier model achieved the highest accuracy, with 88.14%, also standing out for its exceptional computational efficiency, with the lowest optimization cost and the fastest convergence. This work concludes that WiFi Sensing, combined with machine learning pipelines and intelligent optimization, is a viable and effective solution for gesture recognition, offering a positive balance between accuracy, cost, and respect for privacy.