Reconhecimento de atividades humanas em dispositivos móveis utilizando sensores inerciais e aprendizado de máquina: desenvolvimento de um aplicativo para coleta de dados e classificação embarcada

Abstract

Human activity recognition using inertial sensors embedded in smartphones is an area with direct applications in health monitoring, wellness tracking, and indoor navigation. This work presents the development of a complete activity recognition system for Android devices, covering all stages from data collection to real-time on-device classification. An Android application called DataCollector was developed for the synchronized acquisition of accelerometer and gyroscope data during the performance of four activities: walking, standing still, climbing stairs, and descending stairs. Data were collected using five dif- ferent smartphones (SMA556E, SMS911B, SMS926B, Motorola Edge 40, and Motorola Edge 50 Pro), introducing inter-device variability into the training set. The raw signals were segmented into sliding windows of 150 samples with 50% overlap, and 42 statistical time-domain features were extracted from each window. Two classifiers were trained and compared: Random Forest (RF) and Support Vector Machine (SVM), with hyperparame- ter optimization via RandomizedSearchCV and 5-fold cross-validation. Initial experiments with a publicly available dataset revealed domain shift between the external data and the target devices, motivating the collection of a custom dataset. The models trained on the custom data were exported in ONNX format and integrated into the application using the ONNX Runtime for Android library, enabling on-device activity classification without relying on external server connectivity. The Random Forest achieved 88.40% accuracy on the test set, while the SVM reached 86.42%, both evaluated on a total of 2,021 windows extracted from data collected across five distinct devices.