Sensor-Based Human Activity Recognition Using Image-Based 2D Discrete Wavelet Transformation and Deep Learning

The recent development in the wearable devices domain raised interest in wearable sensor-based human activity recognition systems research. This paper introduces a novel human activity recognition system that uses a unique feature extraction method combined with a deep learning network associated with an attention layer. One of the main contributions of this work is the time-wise division of sensor sequences into non-overlapping segments. Each segment is treated as a 2D image, and consequently, we apply image-based feature extraction techniques to it. This includes the use of two-dimensional discrete wavelet transformation to extract useful features from the sensor readings. Experimental results showed that the system outperforms the recognition rates and f1-scores reported in most recent studies in the literature. Specifically, we report recognition rates of 99.44% for the UCI-HAR dataset, respectively.