metric model. 
  
the strip of line 
:d range points, 
] (viewpoint on 
CAR COLLISION AVOIDANCE SYSTEM 
BASED ON ORTHOPHOTO TRANSFORMATION 
S. Yu. Zheltov, A. V. Sybiryakov, O. V. Vygolov 
State Research Institute of Aviation Systems (GosNIIAS), 7, Victorenko str., Moscow, Russia 
zhl@gosniias.msk.ru 
Commission V, WG V/1 
KEY WORDS: Image Processing, Object detection, Three-dimensional scene, Stereoscopic, Model, Real-time. 
ABSTRACT: 
In this paper we propose the complex method of 3D-objects detection based on special orthophoto transformation. Three- 
dimensional scene is registered by calibrated stereo-system with known relative and interior orientation parameters. The objects of 
interest are located on the surface of known analytical model. The paper shows that orthophoto transformation represents each 3D- 
object raised above the surface as a 2D-structure with the predicted properties on the resulted orthophoto images. To find these 
structures the correlation based approach is used in common with statistical analysis of special vertical projections of orthophoto 
images. The method provides simple description of the detected objects such as distance, width and height above the surface. In the 
paper implementation of the method is considered for car collision avoidance system. 
1. INTRODUCTION 
The problem of 3D-objects detection comprised in observation 
scene often occurs in various machine vision applications. 
Typical example of the problem is obstacle detection for car 
safety systems (Bertozzi and Broggi, 1977; Zheltov, 
Sybiryakov, 2000) that is the objective of this paper. 
In the considered work the car-based calibrated stereo-system, 
with known relative and interior orientation parameters, is used 
for real-time detection of obstacles that are located on a road 
ahead of the car and in the same road lane. The road model 
acquisition is performed by computer vision methods followed 
by 3D-reconstruction. 
The obstacle detection method is based on orthogonal 
projections (orthophotos) of the road to some convenient plane 
with use of left and right stereo images. 
Orthophoto represents the 3D-object raised above the road as a 
2D-structure with the predicted properties on the resulted 
images. Transformation of orthophoto image into polar 
coordinate system is suggested to simplify the detection method 
by introducing hardware supported projection technique. The 
3D-object corresponds to simple 2D-clusters of vertical 
straight-line edges on the transformed orthophoto images. 
These clusters are found by implementing correlation based 
approach and statistical analysis of special vertical projections 
of orthophoto images. 
The obstacle detection method consists of number of simple 
single-pass image processing operations such as convolution, 
projections computation and LUT-based transformations. To 
obtain orthophotos in polar coordinate system a piece-wise 
bilinear transformation is applied. The paper shows that 
transformed orthophotos have common properties of both 
orthophotos and stereo images. 
The method provides simple description of each obstacle such 
as distance, width, position in a lane and height above the road. 
Kalman filtering gives relative speed of the obstacle. 
In the paper the examples of moving car detection based on the 
developed method are shown. 
2. OBTAINING A ROAD MODEL 
In this work the situation is assumed that 3D-object, registered 
by car-based calibrated stereo system with known relative and 
interior orientation parameters, stands on smooth road ahead of 
the car (Figure 1). 
  
Figure 1. Example of 3D-scene (left image of stereo-pair) 
Usual way of background surface model acquisition is the least 
square method of fitting to the set of points, which are exactly 
not belonging to the detectable object. In this work the 3D 
marking points are considered as such a set. 
Let the relative coordinate system of left and right cameras be 
defined as follows: the system origin is in projection center of 
the left camera, the X axis coincides with the photography basis 
direction, the Y axis is parallel to the CCD-matrix column 
accepted as initial (average CCD-matrix column) and is 
directed to the sky, the Z axis coincides to the main optical axis 
of the left camera and directed opposite to car move. 
To obtain a road model the road-based exterior coordinate 
system (XYZ) is introduced so that the surface has 2.5D-view 
—125—