Reconhecimento de gestos de ingestão de líquidos utilizando redes convolucionais profundas

Abstract

Adequate fluid intake is essential for health, yet continuous monitoring still relies primarily on self-reporting, which is prone to errors and poorly scalable. Automatically recognizing drinking gestures is the first step toward passively monitoring hydration in real-world settings. However, this problem faces two main bottlenecks: first, the scarcity of public datasets with sufficient size and diversity to train models; second, the need for lightweight models that generalize across individuals and can run on wearable devices. To address these challenges, this work presents a wrist-worn inertial signal dataset (\textit{InGesture}) for liquid intake gesture recognition. \textit{InGesture} comprises recordings from 50 participants performing drinking gestures and analogous actions, totaling 663 minutes of labeled accelerometer and gyroscope data across eight classes. This dataset fills a critical gap in the literature, the lack of large-scale, diverse data needed to train and fairly, reproducibly compare models, thereby promoting cross-subject generalization. Building on this foundation, the dissertation proposes a lightweight 1D CNN designed to capture temporal patterns and optimized for execution on wearable devices. Validation emphasizes cross-subject generalization and covers both binary (drinking vs.\ non-drinking) and multiclass scenarios, with direct comparisons against classical \textit{baselines} and state-of-the-art deep architectures. In experiments, SVM achieved an F1-score of 78.12\% in the binary scenario, while the 1D CNN reached 90.79\% (drinking vs.\ non-drinking) and 87.50\% in the multiclass setting (eight classes). Although data were collected in a controlled environment and the current focus is on event detection (not volume estimation), results indicate the feasibility of continuous, privacy-preserving monitoring using smartwatches, establishing a solid basis for future studies in free-living conditions and for coupling gesture detection with volume estimation.