Real-life Human Activity Recognition Based on Accelerometer Data from Smartphone Sensors Using CNN-BiGRU Model

Abstract

Human activity recognition using smartphone sensors is essential for health monitoring and context-aware applications. This study presents a hybrid deep learning model that combines convolutional neural networks (CNNs) and bidirectional gated recurrent units (BiGRUs) to recognize daily activities from tri-axial accelerometer data. The proposed approach is evaluated on the Real-Life Human Activity Recognition (RL-HAR) dataset, which includes data from 19 participants performing four everyday activities under real-world conditions. Raw sensor signals are pre-processed using noise filtering and overlapping sliding-window segmentation. The CNN component extracts spatial features, while the BiGRU component captures bidirectional temporal dependencies. Experimental results from subject-independent 5-fold cross-validation demonstrate that the proposed CNN-BiGRU model achieves 95.84% accuracy and 95.65% F1-score, outperforming both CNN-only and recurrent baseline models.