Action Recognition in Meeting Videos Using Head Trajectories and Fuzzy Color Histogram

People attending teleconference meetings usually follow specific trajectories corresponding to their intentions. In most situations the meeting video content can by sufficiently characterized by capturing the head trajectories. The tracking of the head is done using a particle filter built on cues such as color, gradient and shape. The head is represented by an ellipse with fuzzy color histogram in its interior and an intensity gradient along the ellipse boundary. By comparing pixels in entry zones to a model of the background we can detect the entry of the person quickly and reliable. The fuzzy color is constructed then in the interior of an ellipse fitting best the oval shape of the head. When a new person appears in the scene a creation of new trajectory is initialized. The recognition of actions is performed using kernel histograms built on head positions as well as segmented trajectories that are related to the layout of the room.