Human Activity Recognition Using Sensor Fusion and Deep Learning for Ergonomics in Logistics Applications

The rapid growth within the e-commerce industry and the necessity of hiring more inexperienced labor in logistics made the warehouse safety more critical than ever. Monitoring the workers' activity during warehouse operations is a fundamental task to create an ergonomic workplace and to reduce operational costs. Therefore, an advanced Human Activity Recognition (HAR) solution is needed to monitor such industrial environments. In this paper, the development of an accurate and reliable HAR classifier using sensor fusion methods is investigated. Three Inertial Measurement Units (IMUs) are used to capture body movement while performing typical warehouse tasks. Deep Learning (DL) algorithms are employed to model the raw time series data and two sensor fusion techniques are applied to combine the data from the different sensors. Overall, the optimal HAR classifier was found to be a stacked model, being more practical and less computationally expensive for industrial applications.