A Hybrid Approach for Smart Home Activity Recognition Using Sensor Data and Deep Learning Techniques
K.V.N. Kavitha, Shashank Pandey, Piyali Saha, Atul R. Patel · Advances in systems analysis, software engineering, and high performance computing book series · 2025
Recognizing human activities in smart homes poses challenges due to environmental variability, sensor systems, and sparse signals. Deep learning models struggle to extract meaningful features autonomously, necessitating additional context. This study proposes a novel hybrid approach merging natural language processing and time series classification techniques to address feature extraction for activity recognition. Sensor events are encoded into frequency-based terms, generating embeddings capturing semantic relationships. Evaluation on two smart home datasets demonstrates the effectiveness of encoding-based embeddings for improving automatic feature learning. Comparisons with Cascade LSTM and other models show the superiority of the proposed approach. The hybrid technique, centered around Cascade LSTM, effectively leverages contextual information from sensor event sequences for recognizing complex human activities in smart homes.