Clustering of Trajectories in Video Surveillance Using Growing Neural Gas - Foundations on Natural and Artificial Computation - page 460

Clustering of Trajectories in Video Surveillance
Using Growing Neural Gas

Javier Acevedo-Rodr ıguez 1 , Saturnino Maldonado-Bascon 1 ,
Roberto Lopez-Sastre 1 , Pedro Gil-Jimenez 1 ,
and Antonio Fernandez-Caballero 2 , 3

1 University of Alcala, Teorıa de la senal y Comunicaciones, Alcala de Henares, Spain
http://agamenon.tsc.uah.es/Investigacion/gram/index.html
2 Instituto de Investigacion en Informatica de Albacete (I3A), Universidad de
Castilla-La Mancha, 02071 Albacete, Spain
3 Departamento de Sistemas Informaticos, Universidad de Castilla-La Mancha, 02071
Albacete, Spain

Abstract. One of the more important issues in intelligent video surveil-
lance systems is the ability to handle events from the motion of objects.
Thus, the classification of the trajectory of an object of interest in a
scene can give important information to higher levels of recognition. In
this context, it is crucial to know what trajectories are commonly given
in a model in order to detect suspect ones. This implies the study of a set
of trajectories and grouping them into different categories. In this paper,
we propose to adapt a bioinspired clustering algorithm, growing neural
gas, that has been tested in other fields with high level of success due to
its nice properties of being unnecessary to know a priori the number of
clusters, robustness and that it can be adapted to different distributions.
Due to the fact that human perception is based on atomic events, a seg-
mentation of the trajectories is proposed. Finally, the obtained prototype
sub-trajectories are grouped according to the sequence of the observed
data to feed the model.

Keywords: Trajectory clustering, growing neural gas, high-level video
surveillance.

1 Introduction

The research on intelligent video surveillance systems has been intensive for the
last years with emphasis on obtaining high-level information for interpreting a
scene in order to give some kind of alarms when an event detection has been
triggered. One of the issues that has gained attention is to automatically recog-
nize behaviors, where the movement followed by the objects of interest can give
rich information about specific or suspicious behaviors. Trajectory classification
is then placed between low level stages such as segmentation, object recognition
or tracking and high level interpretation of the scene. The trajectories followed
by objects of interest in the image is a valuable source of information to auto-
matically detect these suspicious behaviors or generate some kind of alarms [13].