<![CDATA[Smartphone-based human activity recognition (HAR) is an important research area due to its wide-ranging applications in health, security, gaming, etc. Existing HAR models face challenges such as tedious manual feature extraction/selection techniques, limited model generalisation, high computational cost, and inability to retain longer-term dependencies. This work aims to overcome the issues by proposing a lightweight, homogenous stacked deep ensemble model, termed Homogenous Stacking Temporal Convolutional Network with Nu-Support Vector Classifier (HSTCN-NuSVC), for activity classification. In this model, multiple enhanced TCN networks with diverse architectures are organised parallelly to capture hierarchical spatial-temporal patterns from raw inertial signals. Each base model (i.e., TCN) incorporates dilations and residual connections to preserve longer effective histories, allowing the model to retain longer-term dependencies. Additionally, dilations can diminish the number of trainable parameters, reducing the model complexity and computational cost. The base models’ predictions are concatenated and fed into a meta-learner (i.e., Nu-SVC) for final classification. The proposed HSTCN-NuSVC is evaluated using a publicly available database, i.e., UCI HAR, and a subject-independent protocol is implemented. The empirical results demonstrate that HSTCN-NuSVC achieves 97.25% accuracy with only 0.51 million parameters. The results exhibit the model’s effectiveness in enhancing generalisation across individuals with better accuracy and computational efficiency. Doi: 10.28991/ESJ-2025-09-01-026 Full Text: PDF]]>
Copyrights © 2025
