Skeleton-based human action and gesture recognition for human-robot collaboration

The continuous development of robotic and sensing technologies has led in recent years to an increased interest in human-robot collaborative systems, in which humans and robots perform tasks in shared spaces and interact with close and direct contacts. In these scenarios, it is fundamental for the robot to be aware of the behaviour that a person in its proximity has, to ensure their safety and anticipate their actions in performing a shared and collaborative task. To this end, human activity recognition (HAR) techniques have been often applied in human-robot collaboration (HRC) settings. The works in this field usually focus on case-specific applications. Instead, in this thesis we propose a general framework for human action and gesture recognition in a HRC scenario. In particular, a transfer learning enabled skeleton-based approach that employs as backbone the Shift-GCN architecture is used to classify general actions related to HRC scenarios. Pose-based body and hands features are exploited to recognise actions in a way that is independent from the environment in which these are performed and from the tools and objects involved in their execution. The fusion of small network modules, each dedicated to the recognition of either the body or hands movements, is then explored. This allows to better understand the importance of different body parts in the recognition of the actions as well as to improve the classification outcomes. For our experiments, we used the large-scale NTU RGB+D dataset to pre-train the networks. Moreover, a new HAR dataset, named IAS-Lab Collaborative HAR dataset, was collected, containing general actions and gestures related to HRC contexts. On this dataset, our approach reaches a 76.54% accuracy.