OLMM: An Optimized Lightweight Multiscale Model for Efficient Human Activity Recognition in IoT Consumer Applications

Internet of Things (IoT) technologies enable human-centered applications, enhancing the quality of life across various domains. Human Activity Recognition (HAR) plays a crucial role in IoT by analyzing sensor data to infer user states. While recent deep learning approaches show promising results, they demand models that balance accuracy and resource efficiency for consumer devices in IoT environments. Thus, we propose the Optimized Lightweight Multiscale HAR Model (OLMM), a lightweight robust model for HAR, combining three key modules: a Multi-Scale Attention-Based Contextual Feature Extraction Module (MACFEM), a lightweight Gated Recurrent Unit (GRU) module, and a fused center-loss-based classification module. The MACFEM utilizes attention-based depthwise separable convolutions for efficient contextual feature extraction. The second module is a lightweight GRU-based module whose parameters are optimized using our proposed Hiking Optimization Algorithm (HOA). Finally, the classification module integrates Softmax and Center-loss to enhance inter-class separation and minimize intra-class variance. The OLMM achieves F-scores of 99.23 on WISDM, 96.60 on Daphnet, 96.41 on UCI-HAR, and 91.04 on PAMAP2 compared to state-of-the-art works. This model addresses key HAR challenges such as real-time processing and resource constraints by integrating lightweight architecture and optimized GRU configurations, ensuring reduced computational complexity and suitability for consumer devices.