User:KYN/Motion analysis in computer vision

In computer vision motion analysis refers to the processing of image data in order to produce information about motion in the images or in the 3D scene depicted in the images.

Motion analysis can be applied to a sequence of images (video) from a single camera or to only two images from this camera and the motion which is measued in the images refers to moving objects in the depicted scene, or to the motion of the camera relative the scene, or to both. Alternatively, the images may come from two distinct cameras and are taken at the same time. In this case, the differences in the two images is an apparent motion due to the transformation from the coordinate system of one camera to that of the other. Even though there is no real motion here, the estimation of this transformation can be made in the same way as for a single moving camera.

There are two conceptually different approaches for dealing with the motion analysis problem:


 * Measurements of changes in the positions of corresponding points (or other geometric structures such as lines) in two or more images
 * Measurements of changes in image intensities at the same points in two or more images

The two approaches differ mainly in the underlying theory for determining the motion parametersand the type of processing which is applied to the image data.

The first approach assumes that a set of points can be correctly identified and brought into correspondence in at lest two images. Depending on whether the camera in stationary or moving relative to the scene, various types of analysis of the image coordinates of the points can be made in order to determine either the 3D motion of the corresponding 3D points, of the camera, or both. This approach does often provide 3D motion information more or less directly.

The second approach models the image data as a two or three-dimensional signal, with two spatial and possibly one additional temporal coordinate, i.e., it can be described as a function of two or three continuous variables. By considering, for example, derivatives of various orders of the variables, information about the motion components in the image plane can be determined. Provided that additional information, for example, about the camera setup, is available it is then possible to transform the 2D motion information into 3D motion components.

In its simplest form motion analysis means that motion is only determined as occuring or not, i.e., the analysis is focused on detecting whether there is a difference between two images, for example, indicating that something in the scene has moved or that a person has entered into the image. More complex motion analysis provides detailed information about, for example, how fast and in which direction a point or an object is moving.

Motion detection versus motion estimation

 * Maybe move this to a separate article Motion or change detection

Two frame analysis versus sequence analysis

 * Define motion stereo


 * Define optic flow

Two frame analysis of points

 * Ref to epipolar geometry and matching constraints

Two frame analysis of a continuous signal

 * Ref to displacement estimation

Sequence analysis of points

 * Tracking

Sequence analysis of a continuous signal

 * List of methods

Applications

 * Tracking
 * Surveillance
 * Egomotion estimation
 * Video coding