|M.Sc Student||Sheinkin Arkady|
|Subject||Automatic Segmentation of Moving Objects in|
Videoconferencing Sequences for Video Object Plane
|Department||Department of Electrical and Computers Engineering||Supervisor||PROF. Moshe Israeli (Deceased)|
In the present thesis an algorithm was developed for automatic segmentation of moving objects in videoconferencing sequences based on the detection of their outer boundaries. The following assumptions relevant to videoconferencing sequences are made: (a) The camera is assumed to be motionless, so no global motion estimation and compensation is needed; (b) The displacement of an object between successive frames is assumed to be a small fraction of its size.
The difference between gray levels in consecutive frames (FD) is used in order to incorporate the motion information into the segmentation algorithm . Subsequently the FD is thresholded in order to detect a moving object.
In the thesis the solution of three problems is considered: outliers, shadows, and missing data. We call outliers the pixels of the background whose absolute value in the FD is above a chosen threshold . We call missing data those pixels of the moving object whose absolute value in the FD is below a chosen threshold.
At the starting step of our algorithm, a local block matching motion compensation is used in order to balance the motion in different parts of the frame and therefore enable the choice of an appropriate value for the threshold.
For outliers removal a new method based on multiscale frame representation and multiplication is proposed.
For shadows removal a new procedure is used, based on an analysis of the shape of the shadows.
Concerning the missing data, we propose a three-step scheme. Initially a multiscale version of the tensor voting is used. At the second step of the scheme, a tracking algorithm is used, in order to obtain the parts of the object that have completed their motion. At the third step of the scheme we analyze the outside boundary of the object in order to locate discontinuities in it, then we smooth them by cubic spline interpolation.
The designed algorithm obtains a moving object by analyzing four frames only and tries to obtain the real contour of the object. The algorithm was successfully applied to five standard videoconferencing sequences.