Vjacheslav Listsyn 
  
determined range. In this case, both binary images of positive and negative contours are being processed 
simultaneously. 
The filter operates as follows. The image of positive contours is scanned using the semi-mask detecting positive 
contours of the required inclination. In points where output response of “positive” semi-mask exceeds the threshold, the 
calculation of “negative” semi-mask response on the negative image is performed. Horizontal displacement of this 
semi-mask in some minor neighborhood area relative to the position set by the “positive” semi-mask is performed. The 
maximal response value within the above neighborhood area is found, and this value is summed with the “positive” 
semi-mask response value. If the total filter response exceeds the pre-determined threshold, then it is assumed that a 
contour belonging to an 3D object is present in the given point. If not, then scanning of the positive contour images is 
continued. 
The displacement of the “negative” semi-mask in some region is intended to diminish the effects of TV-cameras’ pitch 
misalignment and inaccuracies of images’ transformation to the common observation point on the operation of the 
algorithm. In such cases, on the difference image there appear not the typical facedown triangles, but facedown 
trapeziums, which are successfully detected by this procedure. 
4 MODELING RESULTS 
Figs. 2-6 illustrate the algorithm operation. 
Fig. 2 shows the initial stereoscopic image with long landmarks (road marking lines in this case) that were determined 
using some procedure. This procedure is not considered in this paper. Landmarks positions are defined in an image co- 
ordinates by equations of straight lines approximating road marking inner contours. Object detection is performed 
within the rectangular area. 
Fig. 3 illustrates the results of image transformations to the common observation point. This transformation is 
performed for the object search areas only. The image of a 3D-object has got certain inclination: to the right on the left 
image, and to the left on the right image. 
Fig. 4 represents the results of subtraction of the image fragments corresponding to the object search areas. Here, pixels 
that have positive values on the difference image form the right image and pixels that have negative values on the 
difference image form the left image. The images incorporate four types of pixels: values of pixels derived by 
subtracting background from background, object from object, object from background and background from object. The 
values of pixels obtained by subtracting background from background and object from object are close to zero, while 
the values of pixels obtained by subtracting background from object and object from background, are significant and 
proportional to the object contrast relative to the background. The difference picture incorporates road marking contour 
lines resulting from the imprecise matching of images. In points of the object boundaries typical facedown triangle 
patterns are present on the difference picture. 
Fig. 5 shows the results of the inclined contours selection on the difference images (positive and negative), threshold 
processing and suppression of isolated points. Contours of corresponding inclination have appeared in place of 
triangles. Each triangle is transformed into a pair of contours: one positive, and the other negative. The presence of such 
a pair is the sign that an 3D-object is detected in the given point. 
Fig. 6 represents the results of the matched filter operation. The detected object is marked by bright dots on its edges. 
Investigations of the algorithm sensitivity to the TV-cameras misalignment were performed. To do it, one image of a 
stereo pair (left or right) was displaced by software. The modeling has shown that the algorithm operation is stable (i.e. 
the algorithm detects 3D-objects on images where they are present and does not give false alarms in case of images with 
no 3D-objects) for images with vertical misalignment up to 5-6 pixels. The algorithm’s sensitivity to a horizontal 
camera misalignment is even lower, as it affects primarily the preciseness of determination of the distance to the object. 
As a whole, the performed modeling has proved stable operation of the algorithm. 
  
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 537