Hybrid deep learning model for human activity recognition using smartphone and smart home data

Abstract

With the increasing development of applications in smart homes, healthcare, and personal fitness, a broad scientific community has identified human activity recognition (HAR) as a critical study issue. This paper introduces a unique hybrid deep learning architecture for efficient human activity identification that combines multi-layer perceptrons (MLPs) with convolutional neural networks (CNNs). The suggested CNN–MLP model provides a robust solution for HAR problems by leveraging the feature extraction capabilities of CNNs and the classification prowess of MLPs. Four different datasets were employed to comprehensively assess the model’s performance: WISDM, PAMAP2, and UCI HAR datasets for smartphone-based HAR, and the CASAS Aruba dataset that provided a novel perspective on activity detection in a home context, based on smart homes, while the smartphone-based UCI HAR, WISDM, and PAMAP2 datasets offered a variety of activity data. Across all datasets, our hybrid design outperformed all previous benchmarks, achieving high accuracy rates. These results underscore the model’s adaptability and efficiency in handling diverse sensor data types and activity scenarios. Furthermore, the model’s robustness and generalizability, demonstrated by its consistent performance across multiple datasets, establish it as a significant contribution to the field of HAR.