Human activity recognition algorithms for manual material handling activities

Human Activity Recognition (HAR) using wearable sensors has prompted substantial interest in recent years due to the availability and low cost of Inertial Measurement Units (IMUs). HAR using IMUs can aid both the ergonomic evaluation of the performed activities and, more recently, with the development of exoskeleton technologies, can assist with the selection of precisely tailored assisting strategies. However, there needs to be more research regarding the identification of diverse lifting styles, which requires appropriate datasets and the proper selection of hyperparameters for the employed classification algorithms. This paper offers insight into the effect of sensor placement, number of sensors, time window, classifier complexity, and IMU data types used in the classification of lifting styles. The analyzed classifiers are feedforward neural networks, 1-D convolutional neural networks, and recurrent neural networks, standard architectures in time series classification but offer different classification capabilities and computational complexity. This is of the utmost importance when inference is expected to occur in an embedded platform such as an occupational exoskeleton. It is shown that accurate lifting style detection requires multiple sensors, sufficiently long time windows, and classifier architectures able to leverage the temporal nature of the data since the differences are subtle from a kinematic point of view but significantly impact the possibility of injuries.