Sensor Fusion-Based Deep Learning Models for Human Activity Classification

Wearable sensors have been widely deployed for human activity recognition (HAR) in various sectors, including health monitoring, medical treatment, and motion analysis.Deep-learning-based HAR models have obtained enhanced accuracy rates over traditional machine learning, but there is still a gap in the acceptable recognition accuracy for HAR models.In this regard, this paper aims to propose sensor-fusion-based deep learning models for human activity classification for enhanced accuracy.Experimental validations are conducted on wearable sensor datasets such as ScientISST MOVE, containing various bio-signals, collected during daily human activities using chest, forearm, and wrist sensors.Experimental results show that our proposed sensor fusion-based deep learning model that systematically fuses signals from different sensors at one of the intermediate layers of the model, obtained enhanced performance in HAR.On average, our proposed sensor fusion-based models obtained an increment in HAR accuracy of about 9.14%, Precision of 7.30%, Recall of 8.70%, and F1 score of 9.0% for all deep learning models based on single sensor data (such as wrist, chest, and forearm individually) for HAR.