Automated dual-stream deep network design for activity recognition.

Abstract

This article presents an automated deep learning (AutoDL) framework for multivariate time-series sensor data tailored to the needs of biomechanists. Four public activity recognition datasets featuring different Inertial Measurement Unit sensor placements and class imbalance ratios were selected. Time- and frequency-domain features were automatically extracted, class imbalance was addressed, and classification was performed using a dual-stream convolutional neural network, with a customized genetic algorithm for optimization. Two hypotheses were evaluated: (1) whether the ablation of frequency- or time-domain features results in a significant reduction in model performance, and (2) whether the proposed AutoDL achieves higher F1-score performance than an existing AutoDL system or a model designed under human supervision. A pairwise Wilcoxon test revealed a significant reduction in model F1-score performance when time-domain or frequency-domain features were removed in three of the four datasets (p<0.05). In addition, the proposed AutoDL framework achieved higher F1-scores than both an existing AutoDL system and human-designed models in three of the four datasets (p<0.05). The final models were deployed on a Raspberry Pi 3. The findings underscore the significance of automation and frequency-domain representations in capturing the rhythmic biomechanical characteristics of cyclic movements. The release of the open-source code is expected to accelerate research in biomechanics and facilitate the development of health monitoring tools (https://github.com/MojtabaMohasel/AutoDL_DualStream).