Inter
VIDEO SEGMENTATION WITH SUPERIMPOSED MOBILE MAPS OF DISTANCES dun
com
A.Viloria , J. Finat ^ , and M.Gonzalo-Tasis " cam
toa
aviloria@lpi.tel.uva.es, PMoBiVA Group, Lab.2.2, Edificio [+D, Univ. of Valladolid, Spain jfinat(Zagt.uva.es, the c
marga@infor.uva.es tools
cont
Commission V, WG V/3 com
tools
KEY WORDS: Vision Sciences, Segmentation, Sequences Application, Surveying x
Fast
ABSTRACT with
spec
A real-time and reliable automatic video segmentation is one of the outstanding problems in Computer Vision with large applications patie
to compression, transmission and motion analysis. In this paper, we show a novel approach based on the superposition of distance and
maps linked to centroids of mobile regions acting as attractors of homogenous regions to which different thresholds are applied. The shop
homogeneity of each region is characterized by colour characteristics. The number of colors and the extremal values allowed for loc
parameters corresponding to the shape of regions can be previously configured or learned through an unsupervised training. Our | nec X
real-time processing does not depend on the scene complexity and it is compatible with egomotion, i.e. it is not necessary to Seid,
discriminate beforehand between foreground and background. Compatibility of our segmentation algorithms with egomotion allows wit
the design of on-line tracking and shots identification for automatic segmentation of video sequences by using a low-level meh
topological representation, which is symbolically represented by means of a kinematic mobile graph. pen
Suge
events, i.e., (dis)apparition of larges regions besides a critical Supp
size. A recurrent problem ([Alt00]) is the automatic selection of Som
1. INTRODUCTION thresholding criteria to identify the critical phenomena for an adja
: ; ; h | automatic segmentation. The introduction of metric information fS
The increasing power of personal computers and a higher provides an objective criterion to select critical thresholds by et s
performance of algorithms allow to extend the analysis of adding a spatial information to the viewpoint of [Alt00]. seen
mobile data to currently available digital libraries, including sequ
video files in different formats. Main problems relative to Temporal segmentation ([Kop01]) is a mobile segmentation disc
image processing concern to the computer implementation of which is focused toward the identification of “shots”. A "shot" fen
segmentation and matching algorithms for mobile data. The in a video sequence means a set of image frames with similar andi
extension of video devices requires the design of computer background and continuous motion. A typical example is Qur
tools able of supporting user interaction, eventually based on a provided by a fixed camera in an indoor scene (underground > i
graphics interface to refine interaction. To satisfy different surveillance, e.g.) or in an outdoor scene (traffic surveillance, ee 2
user’s needs relative to different search procedures, it is e.g). Hence, the analysis concerns mainly to temporal ater
commonly accepted that a hierarchised, hybrid and segmentation, including a kinematic information about mobile dd
multilayered approach is required. Hierarchies concern to the objects, eventually. bour
identification of events allowing to separate units of analysis pixe
(isolated events, shots, scenes) along a video sequence. Hybrid High-level dynamic segmentation concerns to the analysis of prop
character must include aspects relative to low-level image “scenes”. A “scene” consists of several consecutive shots that hom
features (colour, textures, e.g.) and high-level image features are “semantically” correlated. prop
(shapes, geometric primitives, e.g.). Multilayered approach is com
translated to several levels of analysis going from low-level Along a "scene", the background is not necessarily the same, com;
contents retrieval in video sequences to the high level and includes some motion of camera, usually. Hence, the expa
interpretation of scenes. analysis concerns to spatio-temporal segmentation, including extre
the estimation of kinematic characteristics of mobile objects in only
Following an increasing complexity order, we can consider image and the egomotion of camera. prim
spatial, temporal and spatio-temporal segmentation. Spatial the
segmentation is a decomposition of a view in static Segmentation techniques must be applicable to static in an =
homogeneous regions. Homogeneity depends on the chosen accurate way and able of a reliable discrimination between ego- he :
threshold for meaningful characteristics (colour, texture) from and external motion. We have obtained meaningful results in all
the information arising from histograms. of them [Vil02], with a real-time and accurate results for Our
images of arbitrary complexity in the static case, and a fast but :
It is difficult to give an objective evaluation of the *goodness" coarser results for mobile regions, including the on-line is ba
of a segmentation ([Pal93]). An evaluation and comparison of capability of discrimination between camera motions and cent
static segmentation to the mid of nineties can be read in external movements (human bodies in TV scenes, e.g.). In this Two
[Zha97]. In this work we are more interested about some work, we extend above results by superimposing metric an e
general problems concerning to mobile segmentation linked to information in different views which is automatically generated com
temporal and spatio-temporal modelling. Hence, from the static from the construction of distance maps linked to segmented he |
sible no
viewpoint our emphasis is put on the description of individual
regions. A real-time update of iconic information is pos
924
