A Hybrid Convolution Neural Network with Channel Attention Mechanism for Sensor-Based Human Activity Recognition

Research studies in machine intelligence and ubiquitous computing have become increasingly intrigued by human activity recognition based on wearable sensors. In recent decades, various learning-based approaches have been significantly applied for human activity recognition to determine successful models for identifying human behaviors. Because of the robust feature extraction abilities of deep learning algorithms, their conveniently extracted features acquire excellent detection capability. Nevertheless, most accomplished deep learning approaches have been established for complicated models with several hyperparameters. This paper investigates the present human activity recognition study on deep learning techniques and discusses appropriate recognition strategies. First, we utilized many convolutional neural networks to determine the effective architecture of human activity recognition. Subsequently, we designed a hybrid convolutional neural network that was implanted with a channel attention mechanism to capture deep spatio-temporal characteristics hierarchically and differentiate human movements in everyday life. Investigations on the UCI-HAR and WISDM datasets revealed that our suggested model, which included cross-channel multi-size convolution transformation, surpassed previous deep learning structures with 98.92% and 98.80% accuracy, respectively. The proposed model outperforms state-of-the-art approaches in terms of total accuracy, as indicated by the research findings.