HRI30: An Action Recognition Dataset for Industrial Human-Robot Interaction

Over the past years, action recognition techniques have gained significant attention in computer vision and robotics research. Nevertheless, their performances in realistic applications, despite dedicated efforts to collect and annotate medium/large datasets, remain far from satisfactory, especially when it comes to applications in the field of human-robot collaboration. In response to this shortfall, we create a dataset not dispersive in its classes but sectoral, i.e., dedicated exclusively to the industrial environment and human-robot collaboration. Specifically, we describe our ongoing collection of the 'HRI30' database for industrial action recognition from videos, containing 30 categories of industrial-like actions and 2940 manually annotated clips. We test our dataset on multiple action detection approaches and compare it with the HMDB51 and UCF101 public datasets using the best-performing approach. We define a baseline of 86.55% Top-1 accuracy and 99.76% Top-5 accuracy, hoping that this dataset will encourage research towards understanding actions in collaborative industrial scenarios. The dataset can be downloaded at the following link: 10.5281/zenodo.5833411