Human action recognition dataset for ergonomic risk assessment applications

Marker-less sensors are widely used for body-tracking in different applications, as Human Action Recognition (HAR) and ergonomic risk assessment. For HAR different datasets were developed to train Neural Networks (NN) able to recognize actions based on labeled skeleton videos. However, state-of-the-art datasets include many tasks but rarely there are work activities like lifting of loads or manual handling. This paper presents an expansion of Northwestern-UCLA dataset with manual handling and lifting of loads. Our contribution includes the addition of 540 acquisitions to the existing dataset where 3 different tasks are performed by 5 different subjects. The first task is not new but useful to generalize task recognition and regards picking up a box from different height and surfaces; these activities were classified as an existing label of Northwestern-UCLA dataset. While two completely new labels were added to the existing dataset, and they are: leaving a box on a table and leaving a box inside a container. In this way, the final dataset contains 12 labels and 2034 acquisitions, and it was used to train and to test a graph convolutional neural network. Results from the testing of the NN reveal a high accuracy in the recognition of the newly added actions. The work here presented aims at supporting automatic ergonomic evaluation in picking activities.