Smart home automation acts as a preventive and protective approach for monitoring and recognizing activities, known as human activity recognition, in a non-intrusive manner through ambient intelligence. During monitoring, ambient devices generate vast amounts of sensor activation data for each activity. Machine learning and graphical models have gained prominence in recent years, delivering promising results for recognizing single-user activities in smart homes. However, smart homes often involve multiple users, including family members, pets, and others. For multi-resident activity recognition, traditional models encounter challenges in feature engineering, resulting in suboptimal recognition rates. Deep learning models have emerged as a robust alternative, outperforming state-of-the-art graphical and machine learning approaches. Despite these advancements, the activity recognition with ambient sensing (ARAS) multi-resident activity recognition dataset poses unique challenges due to conflicting label annotations and data overlapping multiple activity classes, making it more complex than other multi-resident smart home datasets. A generalized model is necessary to address these issues to enhance recognition accuracy and reduce execution time. This paper introduces a custom inception-inspired gated recurrent unit model designed for multi-resident activity recognition. The model incorporates various convolution filters to optimize execution time and improve recognition rates. The proposed approach has been evaluated on the ARAS dataset, comparing its performance against a range of deep learning, graphical, and machine learning models, demonstrating its effectiveness in tackling the complexities of multi-resident activity recognition.
Rajkumar et al. (Fri,) studied this question.