RECOGNITION OF PARTIALLY OCCLUDED MOVING OBJECTS
Michael Schmid
Universität der Bundeswehr München
Fakultät für Luft- und Raumfahrttechnik
Werner-Heisenberg-Weg 39, D-8014 Neubiberg, Germany
Abstract
A computer vision system in an autonomous vehicle
guidance application is presented for interpreting
image sequences acquired by a camera moving relative
to the environment. Objects with different shapes and
changing positions as well as motion parameters in the
perceived scene have to be recognized even if they
occlude each other. The approach described is based
on checking hypotheses by a combination of methods
from knowledge representation and from control
theory, e.g. recursive estimation. Hypothesis verifica-
tion is done by analysing the estimated motion parame-
ters using methods from statistics. These algorithms
have been implemented and tested on synthetic images.
Tests using noise corrupted measurements from a
CCD-camera are currently performed.
Keywords
Computer vision, 3D-object recognition, occlusion, hy-
pothesis generation and verification, recursive estima-
tion
I INTRODUCTION
Recognizing shape and position of three-dimensional
(3D) rigid objects of a given scene is regarded as one of
the main research fields in computer vision. There exist
many different techniques to handle this task in mod-
erately complex situations successfully; to get an over-
view see e.g. [Brady 81], [Chelappa et al. 90] and [Enkel-
mann 90]. But increasing complexity of the scene
observed causes significant problems in identifying and
locating the objects of a given situation. In most cases
multiple objects with different shape and motion may
appear or. disappear, and probably they may partially
occlude each other. This fact complicates the task of
object recognition, but it is an essential feature of a
computer vision system to be able to deal with a wide
range of everyday situations including partially oc-
cluded objects.
Occlusions occur usually in every kind of image pro-
cessing application by different reasons. By using a
CCD-camera the viewing angle onto the environment is
restricted. This fact causes a clipping of the observed
objects, if they are moving at the verge of the perceiving
CCD-chip. Also occlusions may result from the move-
ment of the camera relative to the surrounding environ-
ment or from autonomous moving objects, e.g. cars
overtaking each other. The research work discussed
here deals with occlusions arising from situations of
overtaking cars on German motorways. But it should be
no problem to adapt the algorithms to different situa-
tions. Figure 1 shows a synthetic image of a German
standard "Autobahn" scene generated by a graphic-
workstation with two cars (similar to trucks) driving in
front of the ego-car causing occlusions.
Free |
Es
Traffic situation with an occluded object
Figure 1
Section II starts with a short introduction ofthe machine
vision system for autonomous vehicle guidance on mo-
torways developped at the ’Universität der Bundeswehr
München’ (UniBwM) by the group of Prof. Dickmanns.
Section III gives an overview of the prerequisites for the
internal adaptive model of the real world inside the
image processing system, which is necessary to compare
the measurements from the camera with the internal
description of the tracked objects for updating the esti-
mated parameters of these objects. Section IV de-
scribes the process of initializing an object hypothesis
by the assumption of an existing occlusion. À method
how to assess the generated hypothesis is represented
in Section V. The requirements for the implementation
and some practical results are pointed out in Section VI.
Finally Section VII summerizes the results and gives an
outlook on future research works.
II. SYSTEM OVERVIEW
The structure of the object recognition module will be
described now to give an overview (figure 2).
