ENHANCING HUMAN ACTIVITY RECOGNITION THROUGH SENSOR FUSION AND HYBRID DEEP LEARNING MODEL

Wearable-based human activity recognition (HAR) is essential for several applications, such as health monitoring, physical training, and rehabilitation. However, most HAR systems presently depend on a single sensor, typically a smartphone, due to its widespread use. To improve performance and adapt to various scenarios, this study focuses on a smart belt equipped with acceleration and gyroscope sensors for detecting activities of daily living (ADLs). The collected data was pre-processed, fused and used to train a hybrid deep learning model incorporating a CNN and BiLSTM network. We evaluated the effect of window length on recognition accuracy and conducted a performance analysis of the proposed model. Our framework achieved an overall accuracy of 96% at a window length of 5 seconds, demonstrating its effectiveness in recognizing ADLs. The results show that belt sensor fusion for HAR provides valuable insights into human behaviour and could enhance applications such as healthcare, fitness, and sports training.