Gomila, Cristina (2001) Mise en correspondance de partitions en vue du suivi d'objets. PhD thesis Morphologie Mathématique, CMM- Centre de morphologie mathématique, ENSMP p.242.

Full text available as: - These_Gomila.pdf ( 27624 Kb )

Licence: Copyright

Abstract

In the field of multimedia applications, the incoming standards promote the creation of

new ways of communication, access and manipulation of audiovisual information that go far

beyond the plain compression obtained by the preceding coding norms. Among the new functionalities,

it is expected that the user will be allowed to access the image content by editing

and manipulating the objects of interest. Nevertheless, standards are restricted to object representation

and coding, leaving opened a large field of development concerning the problem

of object extraction and tracking when they move along a video sequence.

In a first step, we have proceeded to the study and fine tuning of widespread applicated

algorithms for image filtering and segmentation, being these tools at the basis of all contentbased

image and video analysis systems. More particularly, we have focused on a novel class

of morphological filters known as levelings, as well as on a variant of the segmentation algorithms

based on the constrained

ooding of a gradient image. Segmentation techniques aim

at yielding a partition image as close as possible to the one produced by the human eye, with

a view to the later object recognition. Nevertheless, in most cases this last task needs human

interaction. However, when we would like to retrieve an object from large collection of images,

or when we would like to track an object through a long sequence, the surveillance of each

image becomes infeasible. To face these situations, the development of matching algorithms

able to propagate the information through a series of images become essential, human interaction

being limited to a initialization step.

Going from still images to sequences, the core of this thesis is devoted to the study of the

partition matching problem. The method we have developed, named Joint Segmentation and

Matching technique (JSM), can be defined as being of hybrid nature. It combines classical algorithms

of graph matching with new editing techniques based on the hierarchy of partitions

resulting from morphological segmentation. This mix provides a very robust algorithm, in

spite of the instability classically associated to the segmentation processes. The result of segmenting

two images can strongly differ if the segmentation process produces a single partition

image, however we have shown that results are much more stable when producing a hierarchy

of nested partitions, in which all contours are present and ranked through a weighted

value. The JSM technique is considered a very promising approach according to the obtained

results. Being flexible and powerful, it allows the recognition of an object when it reappears

after occlusion thanks to the management of a memory graph. Although we have particularly

focused our interest on the tracking problem, the developed algorithms can be extended to



a large field of applications, being specially suited to perform object retrieval from image or

video sequences databases.

Finally, in the framework of the European project M4M (MPEG f(o)ur mobiles), we have

focused on the development and implementation of a real-time demonstrator for detecting,

segmenting and tracking the speaker in videophone sequences. In the view of this application,

the real-time constraint has become the greatest challenge to overcome, forcing us to simplify

and optimize our algorithms. The main interest in terms of new services is twofold : on one

hand the automatic segmentation of the speaker permits the object-based coding, reducing

the bitrate without loss of quality on the regions of interest ; on the other hand, it allows the

user to edit the sequences by changing the scene composition, for example by introducing a

new background, or grouping several speakers in a virtual meeting room.

Statistiques de consultation

Repository Staff Only: edit this item