Human Activity Recognition via Time Series Imaging of Wearable Sensors and Fusion of Multiple Deep CNN Models

Human Activity Recognition (HAR) plays a vital role in healthcare, smart home, and security applications. Recent advances in sensor technology and machine learning have substantially enhanced HAR capabilities, particularly for continuous monitoring of elderly individuals. Time-series imaging and convolutional neural networks (CNNs) have been effectively applied to this task. In this study, we adapt these techniques to the underexplored and challenging GOTOV dataset and further develop a novel framework, PWMF-ResMiniNet, a lightweight yet highly accurate solution for HAR optimized for resource-constrained edge computing environments. The framework introduces two key innovations. First, the Precision-Weighted Model Fusion (PWMF) algorithm performs decision-level fusion of multiple sensor-specific CNNs by assigning adaptive per-class weights based on each model’s validation precision, achieving state-of-the-art accuracy with zero additional computational cost while remaining interpretable and theoretically sound. Second, the ResMiniNet CNN architecture, inspired by the ResNet model, reduces trainable parameters by approximately 75% while maintaining competitive accuracy. Combined with Gramian angular field imaging, PWMF-ResMiniNet achieves over 90% F1-score on the GOTOV dataset—the highest reported accuracy on this dataset to date—and demonstrates consistent superiority over state-of-the-art methods such as HARDenseCNN and HART on the PAMAP2 and WISDM 2019 datasets. Moreover, the study discusses the limitations of the proposed framework and current HAR approaches, identifies open research challenges, and highlights promising directions for future investigation.