Hand-Crafted Features With a Simple Deep Learning Architecture for Sensor-Based Human Activity Recognition

• Published in: IEEE sensors journal 2024-09, Vol.24 (17), p.28300-28313
• Main Authors: Albadawi, Yaman, Shanableh, Tamer
• Format: Article
• Language: English
• Subjects: Accuracy; Attention; bidirectional long short-term memory (BiLSTM); Data models; Datasets; Deep learning; Dynamic structural analysis; Feature extraction; Feature recognition; Human activity recognition; Long short term memory; long short-term memory (LSTM); Machine learning; Segments; Sensor phenomena and characterization; sensor-based human activity recognition (HAR); Sensors; Sequences; Telemedicine; Time series; time-series differencing; Two dimensional analysis; Ubiquitous computing; Wearable technology
• ISSN: 1530-437X; 1558-1748
• DOI: 10.1109/JSEN.2024.3422272

With the growth in the wearable device market, wearable sensor-based human activity recognition (HAR) systems have been gaining increasing interest in research because of their rising demands in many areas. This article presents a novel sensor-based HAR system that utilizes a unique feature extraction technique associated with a deep learning (DL) method for classification. One of the main contributions of this work is dividing the sensor sequences timewise into nonoverlapping 2-D segments. Then, statistical features are computed from each 2-D segment using two approaches; the first approach computes features from the raw sensor readings, while the second approach applies time-series differencing to sensor readings prior to feature calculations. Applying time-series differencing to 2-D segments helps in identifying the underlying structure and dynamics of the sensor reading across time. This work experiments with different numbers of 2-D segments of sensor reading sequences. Also, it reports results with and without the use of different components of the proposed system. Additionally, it analyses the best-performing models' complexity, comparing them with other models trained by integrating the proposed method with an existing transformer network. All of these arrangements are tested with different DL architectures supported by an attention layer to enhance the model. Four benchmark datasets are used to perform several experiments, namely, Mobile Health (mHealth), ubiquitous computing human activity dataset (USC-HAD), University of California Irvine HAR (UCI-HAR) dataset, and daily and sports activities (DSA). The experimental results revealed that the proposed system outperforms HAR rates reported in the most recent studies. Specifically, this work reports recognition rates of 99.17%, 81.07%, 99.44%, and 94.03% for the four datasets, respectively.