316
ball 1 ball 2 ball 3 ball 4
A x (cm) | A y (cm) | A x (em) | A y (em) | A x (em) | A y (em) || A x (em) | A y (em)
L.311 [| 105 T 200 | da | 81 | Wy | 530 | 78 |
Table 1: Maximum values of ball movements for the snow fall from February 28, 1994, 1100 to March 1, 1994,
615. Vertical and horizontal movement is increasing with distance from the trunk. Ball 4 is mounted
in a distance of 66 cm from the trunk, Ball 1 is mounted in a distance of 200 cm from the trunk.
Especially o? is a quite high value because this is the angle of the branch but not of the snow surface on the
branch, which is much larger. That is especially important when snow fall continues after a branch is heavily
loaded because falling snow often bounces off when the branch angle is too steep and then no more snow is
accumulated (Kobayashi, 1987).
3 DISCUSSION
As explained in section 2.3 motion of branches during a snow fall event can be observed quite well. How precise
this method is working is determined by the accuracy the ball coordinates can be certained. This is determined
by (i) the precision of the coordinates of the digitized image and (ii) the possibility and precision of the al-
gorithm which detects the balls and calculates the center of mass. The sampling accuracy of digitized video
images is very high. Due to image resolution (736 by 578 pixel) and the size of the ROI (544 cm by 408.5 cm)
one pixel has the size of 7.4 by 7.1 mm. To check the sampling accuracy (noise of position of every image) in
every image the position of a fix well defined point has been determined. In every analysis run (see section 2.3)
the coordinates of such a point were calculated. The range of these values lay between 7.4 mm in the horizontal
and 7.1 mm in the vertical direction with a standard deviation of 2.2 mm in the horizontal and 3.5 mm in the
vertical direction which corresponds to a precision of less than one pixel.
As mentioned in section 2.1 and section 2.3 the changing light conditions in a forest present a major problem
in object recognition. Especially when branches are loaded with snow the contrast between the balls, snow and
the rest in the ROI is low. Although the ROI was chosen as small as possible especially the red balls in the
image which showed a comet-like red shadow could not be recognized. The red color of the balls was too less
clear in the ROI that means the color could be found several times. Tests with image enhancement yielded no
satisfactory results especially for day images. On night images there is a better contrast but it still was not
impossible for the machine to find the red balls automatically.
In winter 1993/94 the red balls were replaced by yellow ones with little bulbs inside. With an alternating
illumination with floodlights every 5 min the contrast could be improved. For day images the problem of object
recognition still exists, because the contrast between snow, ball and the tree is still too small. In an example a
thresholding with values of the balls provided an error of 77%. That means 720 from 933 pixel in the ROI had
the same colors in the RGB color room as the balls. In this case only the analysis with the “search criteria”
brought a success. But there is another problem. The edge of the balls is often not very clear so that there
results an undefined area containing parts of the ball and parts of the snow. Fig. 3a-b shows the histograms
for a night image with illumination and without illumination with yellow balls.
At night and without illumination, most of the image is dark with only some bright values. The balls at night
with illumination also yielded a better contrast than the balls in the day image but automatic recognition still
remains difficult. For day images digitizing had to be mainly done by hand.
4 CONCLUSIONS
The results of this experiment can be summarized as follows: (i) for motion studies of natural processes the
video technology is a useful tool, (ii) to quantify any observation it is necessary that contrast in the image is
very clear. Light conditions in a natural environment are changing very quickly and requires an adapted image
analysis algorithm. Best results can be obtained in a bright-dark environment (iii) the accuracy of video is
quite high for the presented application. Images can be analysed with at least a precision of one pixel.
The automatic recognition of the balls, and therefore the movement of the branch, was not possible using a
standard algorithm. Further experiments will be done using an adaptive algorithm, which uses information .
IAPRS, Vol. 30, Part 5W1, ISPRS Intercommission Workshop “From Pixels to Sequences”, Zurich, March 22-24 1995