Moving Object Recognition and Classification based on Recursive Shape Parameter
Estimation
Dieter Koller
In Proc. of the 12th Israeli Conf. on Artificial Intelligence, Computer Vision, and
Neural Networks, pp. 359-368, Tel-Aviv, Israel, December 27-28, 1993.
Abstract
We present
an approach for recognizing and classifying moving vehicles in monocular
images sequences of traffic scenes recorded by a stationary camera.
A generic vehicle model,
represented by a 3D polyhedral model described by 12 length parameters, is used to
cover the different shapes of road vehicles.
The object recognition process is initialized by formulating a
model hypothesis using a reference model and initial values
provided by a motion segmentation step from a model-based tracking system described
previously. This
model hypothesis is verified and the shape as well as the pose and motion
parameters of the object are estimated simultaneously.
A recursive
estimator updates the state description of the shape and motion parameters.
In this way all relevant data from the image sequence evaluated so far
are accumulated and used for the shape parameter estimation and classification
of a moving vehicle.
A classification is based on the assumption that differences between
class members can be considered as deformations of the shape of a stored prototype.
Results on real world traffic scenes are presented and some open problems
are outlined.
The document is available online in
application/postscript (1038762 Bytes)
Last modified on Tuesday, November 20, 1996,
Dieter Koller
(koller@vision.caltech.edu)