Motion perception is one of the major goals of computer vision research. Motion models span from rigid motion — motion of rigid objects, egomotion (camera motion), and so forth — to complex articulated motion. Deformable objects form an interesting case because the models are closely related to the underlying physical phenomena. Examples of motion perception scenarios of increasing complexity are:

- Find the motion parameters of one or several objects, possibly articulated, such as a person or a group of people, observed by several cameras. The motion parameters thus estimated can then be fed into higher level processes such as gesture recognition, motion analysis, motion animation, interactions with virtual objects and avatars, surveillance and monitoring systems, etc.

- Detect the positions and relative speeds of obstacles in front of a vehicle, whether it is autonomous or equipped with a driving assistant. Motion information about these obstacles can then be used to modify the vehicle’s trajectory or its speed, for example to avoid accidents and crashes.

Generally speaking, the process of motion analysis through visual data can be decomposed into the following tasks. First, an appropriate motion model must be choosen such that it represents the physical properties of the object (rigid, articulated, deformable) and such that the associated motion parameters are observable and hence measurable from visual data. Second, detection of motion must be performed through the extraction of motion cues such as optical flow, points of interest, contours, silhouettes, and so forth. Third, tracking consists in estimating the parameters associated with the motion’s degrees of freedom. Tracking can be performed in image space, in 3-D space, or in motion space.

Currently we investigate the followings: