We are investigating the problem of recognizing different types of human gaits (i.e. walking, running, jumping...) in the space of dynamical systems where each gait is represented. We start from a collection of trajectories of joint angles/positions obtained from visual tracking or motion capture. From this data we learn a dynamical system representing the gait. Then we define a distance between models that allows discrimination between different classes of gaits.

We are studying the problem of modeling human gaits for the purpose of synthesis. Our approach is based on representing the trajectories of a certain number of salient features on the human body as the output of a linear dynamical system driven by white noise. We are investigating the problem of recognizing different types of human gaits (i.e. walking, running, jumping...) in the space of dynamical systems where each gait is represented. We start from a collection of trajectories of joint angles/positions obtained from visual tracking or motion capture. From this data we learn a dynamical system representing the gait. Then we define a distance between models that allows discrimination between different classes of gaits.

We are developing techniques for tracking human body in video sequences. We model the human body as a kinematic chain of body parts which undergo a transformation composed of rigid motion and shape variation. The tracking problem is posed as the estimation of unknown position, orientation and shape of the body parts.

We propose models and learning algorithms for synthesis of human facial motion driven by a speech signal. We collect trajectories of a collection of feature points for an individual and the associated speech waveform, an from these data build a model that can be used to generate novel synthetic facial motions associated with novel speech segments.

We are investigating a characterization of "visual action" that allows for detection in presence of distractors. To obtain this goal we proposed models which have a compositional property, where a simple action (e.g. foreground action) can be detected within a more complex one (e.g. foreground and background action).

We developed a fast visual feature tracking system which takes advantage of dedicated hardware to perform the computationally intensive step of selection. A software system uses the output of the hardware selector to develop tracks using filtering and data association techniques, and image-based validation.