Light-PTNet: A lightweight parallel temporal network for smartphone-based human motion classification

The increased popularity of smartphone-based human activity recognition (HAR) in recent decades has been driven by its low computational requirements and user privacy protection. Yet, developing a reliable smartphone-based HAR still presents several challenges. For example, handcrafted feature-based approaches highly depend on laborious feature engineering/selection techniques that require human intervention. Implementing conventional Convolutional Neural Networks may result in unsatisfactory performance in time series classification as they cannot effectively extract time-dependent features. Although recurrent models excel at extracting temporal information, they require extensive computational resources to attain high performance, limiting their practicality for real-time applications. Thus, we propose a lightweight smartphone-based HAR architecture called Lightweight Parallel Temporal Network (Light-PTNet) for reliable classification. Light-PTNet comprises parallelly organised Light Spatial-Temporal Heads (LSTC Heads) that capture underlying patterns at various scales of the inertial signals. These heads utilise dilations and residual connections to preserve longer-term dependencies without increasing the model parameters. This work assesses the proposed Light-PTNet's performance on open-access HAR datasets: UCI HAR, WISDM V1, and UniMiB SHAR, following a user-independent protocol. The results reveal that our proposed Light-PTNet achieves 98.03% accuracy on UCI HAR, 81.58% on UniMiB SHAR and 97.02% on WISDM V1 with fewer model parameters (lower than 0.1 million parameters).