A multi-environment dataset for activity of daily living recognition in video streams

Public datasets played a key role in the increasing level of interest that vision-based human action recognition has attracted in last years. While the production of such datasets has been influenced by the variability introduced by various actors performing the actions, the different modalities of interactions with the environment introduced by the variation of the scenes around the actors has been scarcely took into account. As a consequence, public datasets do not provide a proper test-bed for recognition algorithms that aim at achieving high accuracy, irrespective of the environment where actions are performed. This is all the more so, when systems are designed to recognize activities of daily living (ADL), which are characterized by a high level of human-environment interaction. For that reason, we present in this manuscript the MEA dataset, a new multi-environment ADL dataset, which permitted us to show how the change of scenario can affect the performances of state-of-the-art approaches for action recognition.