92
INITIALIZING THE RECOGNITION OF MOVING PERSONS
Winfried Kinzel, Reinhold Behringer
Institut für Systemdynamik und Flugmechanik, Universität der Bundeswehr München
Werner-Heisenberg-Weg 39, D-85577 Neubiberg, Germany
Phone + 49-89/6004-3583, Fax + 49-89/6004-2082
Keywords: Fast feature extraction and grouping, object recognition; moving humans recognition.
For road vehicle guidance and driver support a system for recognizing pedestrians has been developed. Data-driven
image processing is connected with model-based inferencing. First, image intensity discontinuities are very effi-
ciently transformed into an abstract description in terms of straight line approximations. From this representation,
sets of regions are segmented exploiting grouping phenomena. To validate relational parameters, statistical data
analysis using decision tree techniques has been employed. After additional tests concerning head regions the seg-
mentation results form an object hypothesis which is further assessed by means of combining all evidences by their
Bayesian probabilities and which is smoothed in the time dimension.
1 Introduction
This work ! presents an application-driven solution to the problem of quickly recognizing a class of complex objects
in image sequences. It deals with detecting and classifying scene contents as human beings under the realistic cir-
cumstances of street traffic in order to initialize a model-based reconstruction of limb angles over time. This implies
that no restrictive prerequisites for a person’s clothing nor the viewing conditions can be presupposed. The camera
will normally be in motion, whereas a person in the vicinity of a road might move or be in a state of rest. The ap-
proach is validated on both synthetic and real image sequences. The system presented has been implemented on a
graphics workstation.
The contents of this contribution are organized as follows. First, the motivation for this study is given, the rec-
ognition task is characterized, and related work on recognition of humans is reviewed. Secondly, the steps for ar-
riving at an abstract image description are explained. Thirdly, the procedure of region identification is given. Both
edge and region descriptions are part of a preattentive image processing phase.
1.1 Motivation
The objective of automatically recognizing humans and their movements from image sequences was conceived in
the frame of general object recognition for the purpose of autonomous road vehicle guidance and driver support
[Dickmanns 92]. Especially detecting human beings within an autopilot is justified by their need of special protec-
tion in road traffic as well as their ability to signal messages by gestures that might be relevant for cooperative ve-
hicle control.
1.2 Task characterization
In recent years the topic of recognizing humans by computer vision applications has received increasing attention.
The difficulties in recovering human figures from image sequences mainly derive from the high degree of shape
variability that is exhibited by articulated bodies in motion. But the limbs the body mainly consists of have almost
constant shape properties and can therefore be regarded as rigid subparts. As limbs have a natural axis of elong-
ation, they can be represented by generalized cylinder primitives within a computer model. Their connections at
the joints are fixed, forming constraints on the grouping of parts. A set of physiologically possible movements forms
another class of constraints. Within an animation tool such a figure model has been realized to simulate model-
based movement reconstruction [Kinzel 92]. The crucial problem is how to initialize model-based movement rec-
onstruction during a ‘start-up phase, i.e. how correct assignments can be established between model entities and
image features. So, efficient procedures for data abstraction, feature selection and model matching have been de-
signed.
1 This research project has been supported by the German Federal Ministery of Research and Technology (BMFT) and
Daimler-Banz AG, Grant Eureka Prometheus III.
IAPRS, Vol. 30, Part 5W1, ISPRS Intercommission Workshop “From Pixels to Sequences’, Zurich, March 22-24 1995
the
ten
