Tracking Module Theory

Recovery of visual motion

Moving objects of our three-dimensional euclideian world are projecting on the two-dimensional image plane of our camera (the device which we use in order to observe our world). One of the most important problems in Computer Vision is the recovery of the motion of the moving objects in our 3-D world. The first step is the recovery of the motion of the projected 2-D images of those objects. We distingish between two types of motion processing (image-plane motion):

Full-field motion processing
We use global information (full frame) and perform optimization (optical flow), applying iterative techniques. This approach is unsatisfactory for real-time applications, since it requires specialized hardware.

Feature-based motion processing
We conctrate on spatially localized areas, reach in geometrical and physical information (i.e. edges, corners, etc.). Our aim is to recover the motion of those areas. These methods are less computationally intensive, because we use only a part of the information which a full frame carries. Those methods are appropriate for real time applications and can be implemented in software (more flexible solution).

Image Level Features

The first step of Feature-Based algorithms is the extraction of image-level features . A feature is an amount of information extracted from an image region of interest. Some examples of features are:

Edges
Spatial discontinuities in the image plane. These discontinuities are caused by depth or normal discontinuity of a scene surface, or by illumination discontinuities (shadows) on our world, or by occlusion . Some of the attributes of an edge are its location and orientation on the image-plane as well as its magnitude (strong or weak edges).

Corners
Interecting edges. An attribute of a corner is the angle between the two interecting edges. Of course the corner inherits the attributes of its edges.

Gray-level image region
A feature could be an image region of a specified gray-level pattern (this is the kind of feature used in our application, see setting tracking parameters.

Our Implementation

Our implementation of tracking uses the XVision package developed by Greg Hager at Yale.

Up to Tracking Contents