A study on the Effect of Occlusion in Human Activity Recognition

Abstract

The problem of occlusion plays a crucial role in real-life human activity recognition applications. However, most research works either underestimate it, or base their training solely on datasets shot under laboratory conditions, i.e., without any partly or full occlusion. In this work we perform a study on the effect of occlusion in the task of human activity recognition and the domains of the recognition of a) activities of daily living; and b) medical conditions. Throughout our experiments we use a convolutional neural network that has been trained using a 2D representation of skeleton motion for all available joints, i.e., without using any occluded samples. We evaluate our approach using two challenging, publicly available human motion datasets upon removing one or more body parts. © 2021 ACM.