Enhancing Sensor-Based Human Activity Recognition using Efficient Channel Attention
Anuchit Jitpattanakul, Sakorn Mekruksavanich · 2023
Sensor-based human activity recognition (S-HAR) has found applications in various domains such as medical care, sports player tracking, and health monitoring using smart wearable devices. Among the different types of sensors used in S-HAR, inertial sensors have gained popularity. Researchers have made significant progress in S- HAR systems over the past decade by employing machine learning (ML) techniques. However, the effectiveness of ML approaches was limited because they relied on manually designed feature extraction methods. To overcome this limitation, researchers have developed deep learning (DL) methodologies that automate the feature extraction process from sensor data. This study aims to investigate the effectiveness of an attention mechanism in improving the identification capabilities of DL models. We introduced an efficient method called efficient channel attention (ECA) to enhance the performance of advanced DL models in HAR. To evaluate the progress, we utilized the UCI-HAR dataset, a publicly available benchmark dataset that records six distinct human behaviors in daily living. Five baseline DL models, namely CNN, LSTM, BiLSTM, GRU, and BiGRU, were employed. The ECA mechanism was then incorporated into these models. The experimental findings demonstrate that the proposed ECA mechanism has the potential to improve the recognition capabilities of S- HAR.