1.2. Previous work

An extensive body of literature has been accumulated in  the  computer   vision   community   regarding  the

study of motion. Most of  this  literature  focuses on the structure-from-motion  proble    2 which  involves

the computation of camera, object andJor environ- mental parameters based on relative motion between these entities. Often, assumptions are made in visual motion research that prohibit the use of the proposed techniques in applications such as robotic visual servoing. A large number of previously proposed systems exhibit one or more of the following characteristics:

• Systems which avoid the visual detection issue altogether. They assume that their methods are applied on an image after the presence of moving objects has been  identified  and measured.

• Systems which are applied to artificially trivial conditions  that  do  not  occur  in natural settings.

In our work, we have tried to avoid these assumptions, specifically focusing on addressing the 8°11l of visually detecting objects  of  interest  for robotic visual servoing.

The majority of existing attempts at detecting moving objects has employed either optical flow or frame-differencing techniques. Optical flow methods are interesting, since they naturally encompass ego- motion of the camera (although some optical flow methods    have    the    equalizing    disadvantage   of

actually requiri •8 é o-motion). For instance, Jain,’ Nelson,”’  and  Thompson  and  Pon  6 have compiled

a collection of optical flow-based motion detection algorithms which detect a moving object as an inconsistency in some constraint on the optical flow field. Some of these optical flow-based algorithms use a constraint that is based  on  the  orientation of motion vectors away from a focus of expansion (FOE). However, algorithms using the FOE con- straint are not reliable when the distance between a moving object and the FOE is small. Another common optical flow constraint is the assumed relation between optical flow gradients and corre- sponding depth disparities  (typically computed  with a stereo vision system). Instead of using a  stereo vision system in our research, we have restricted ourselves to monocular systems that can  acquire visual information with relatively unsophisticated off-the-shelf sensor devices.

In contrast to detection based on optical flow, our framework shares many characteristics with other frame-differencing techniques. An example of these is the system developed by Anderson ei at.’ They detect motion through the use of the Gaussian/Laplacian pyramid which Burt and others have used in a variety of computer vision systems."" The  pyramid  is applied to the difference between a current  input frame and the previously input frame. This has the disadvantage of only signaling appearing and disappearing edges of a moving  object.  Moreover, the difference responds similarly to large objects as it does  to  fast  ones.  Both  of  these  situations  do  not

occur  in our approach.  The Anderson  method 7 uses

another  Gaussian  pyramid  to  facilitate subsampling

down  to a level  where  motion  segmentation  can be 1.3.  Structure of the paper

performed by a general purpose computer. However, This  paper  incrementally  presents  each  one  of the subsampling  is  restricted  to  the  logarithmic   levels