Título de la tesis


Feature-based Human Tracking: From coarse to fine

Doctorando


Jesús Martínez del Rincón

Director


Carlos Orrite

Descripción


A robust tracking algorithm is the cornerstone of human motion analysis applications. This
thesis describes the tracking problem from three points of view: punctual, regional and human
pose modelling. In the first one, the subject is considered as a compact object, small and rigid.
In the second one, it is possible to model it as a set of connected regions, each one with its
own properties, and which junction identifies the subject. The last point of view allows us to
identify the person as a non-rigid object, with morphologic and dynamic properties which are
intrinsically human. Our goal consists in the study and understanding of these three different
stages in which the human being can be modelled, as well as the selection of the most suitable
stage depending on the application and the problem to be solved. The problem will set the
necessities and impose the limitations in each time, due to internal factors (multiple target,
occlusions, camera location) or external factors (cost, technical feasibility). Therefore, the
emphasis in this work lies on producing a system capable of tracking one or multiple people in
video sequences, with different degrees of understanding depending on the specific application,
the necessities and the available means. In order to achieve this goal, the project will focus on
human tracking, human modelling and feature extraction. The system should be able to work
in real life scenarios, such as video surveillance, sport analysis or medical diagnosis.

In punctual domain, a system capable of tracking people successfully has been developed,
in spite of bad measurements or poor quality sensors, thanks to the combination of a static
object detector, a height estimator and a multicamera conjugation algorithm. The system has
been designed for surveillance applications and even the camera calibration has been simplified
as much as possible to fulfil this requirement.

Regarding the feature extraction field, special emphasis on feature extraction has been done.
Colour modelling, using parametric and non-parametric methodologies, has been the basic clue
to track the targets due to the generality and invariance that this features provides. In addition,
a robust colour update technique has been presented, which is able to adapt itself to both fast
and slow changing illumination conditions.

An efficient colour tracking algorithm, based on particle filter, has been proposed to speed up
the computation of the conventional version of this multi-hypothesis algorithm. The inclusion of
techniques such a partitioned or importance sampling reduces the number of samples since they
discard those hypotheses with low probability. Furthermore, the usage of the integral image in
the evaluation procedure minimises significatively the computational time of evaluating each
hypothesis.

An incursion in tracking of multiple identical targets has been done. Association techniques
and interaction modelling have been proposed to deal with the coalescence and help the tracking
to solve ambiguities. The high complexity of multi-target tracking has also demanded the
creation of a integrated framework where multiple sensor information is conjugated.

Finally, articulated models have been employed to track not only the global motion of the
target but also the relative motion of the limbs. In this field, 2d models have been proposed
due to the fact that they are much more adequate for surveillance purposes, being able to work
in monocular sequences, have a lighter computational load and require a simple initialistaion.
The main drawback that has relegated this techniques, i.e the viewpoint dependence, has been
tackled in depth. Morphologic and biomechanical information, introduced as part of the model
or by means of constraints, allows the achievement of this goal. The two possible methodologies,
discriminative and generative approaches, have been tested and compared.


Estado


Defendida - 2008

Calificación


Sobresaliente Cum Laude



Validación de código - XHTML 1.1
Validación de CSS
Validación de Accesibilidad Web