Seite - 105 - in Proceedings of the OAGM&ARW Joint Workshop - Vision, Automation and Robotics

level gauge or resolution of the image. As the correctness of the reading and the declaration of the confidence of the final value are of particular importance, performing different approaches and a subsequent comparison are worth the additional time needed. Running tests of the algorithm on the real robot on an oil platform training site it became apparent, that the underlaying algorithm that takes images of the level gauge returns images with low deviations of the box position from the image center. On average the probability of the box being close to the center of the scene image is much higher than it being close to the edge of the image. Taking this into account, the confidence of the detected box being correct is multiplied with an additional function 1− c 100 √ (w/2−x)2+(h/2−y)2) (w/2)2+(h/2)2 where w is the width and h is the height of the scene. x and y define the center-point of the found box. c is a constant giving the percentage of how much the confidence is lowered if the box center is in one of the corners of the image, i.e. the box center-point with the biggest distance to the scene center that is still in the image. Using screw templates worked best for high resolution images and only slight differences in size and illumination. The approach for box detection based on finding the correct border lines outperformed this method when the image had a low resolution. Fig. 12 shows the results of box detection for six different images, that are used to get the correct waterlevel. Fig. 13 shows the cropped and warped boxes, for later getting the column images. B. Cut-out of Liquid Column Tests have shown that the precision and correctness of cutting out the liquid column of the level gauge highly depends on the preceding box detection. Having detected the outer box with an error less than ten percent of its width, ensures a high probability of getting a correctly cut out liquid column as a basis for the following level detection. C. Level Detection After testing with a halogen work light serving as an artificial sun and a model of the level gauge, images taken in the real environment with sunlight show the same results. As images of the level gauge have been taken from five or six different heights, they now have to be arranged in a way they can be compared to each other. Using one column image as reference the other column images are fitted by subtracting intensities while slightly shifting the column image to fit. Multiple checks with small changes in size improve the result. Applying the Histogram-Gauss-Adding method on images that are taken from different heights delivers good results. To verify the algorithm also sets of images with many scenes with high intensity gradients at the same height are tested. Fig. 14 shows the liquid columns cut out of the boxes in Fig. 12. Box detection in multiple images of the scene that are used to obtain the liquid level Fig. 13. Resulting box images for liquid column extraction Fig. 14. Extracted liquid columns with multiple level hypotheses (Despite the fact that four out of six images have similar positions of the reflections the correct level is still found.) 105