Human action recognition and segmentation are important tasks in the context of assistive robotics and active assisted living. Multiple datasets of human actions are available and cover a large set of general or daily activities, but they less cover the case of abnormal events. When some specific actions are available, they are not adapted for the context of assistive robotics. In this paper we will describe a new dataset that covers some specific events which are useful in the context of active assisted living. We will also describe an approach for human action recognition based on human body semantic segmentation, to focus on the human body and deal with partial or slow movements and immobility.