Page - 103 - in Proceedings of the OAGM&ARW Joint Workshop - Vision, Automation and Robotics

Fig. 6. Attain horizontal borders of the box In the third approach that can be seen in fig. 7 advantage is taken of the texture of the level gauge. Besides the metal box and the liquid column in the middle it consists of ten big screws arranged in two vertical lines. A set of screw images is used for template matching and detecting possible screws in the scene. The concept is to deliberately look for more than ten screws and then classify them into so called good screws, that do belong to the gauge, and bad screws, that do not. This is done by creating a new black image, where the center of every found screw is marked as a white pixel. If a pixel is alreadywhite, the one below is made white instead to make it count. Afterwards Hough line detection is applied in this binary image to find lines of screws. The two lines with most participating pixels are used to finally determine the position of the box. Knowledge about the maximum amount of screws or about their similar vertical distance can be used to optimize the result. Fig. 7. Attain the position of the outer box by using template matching to find screws of the box. More than the existing 10 screws are to be found to then form lines of screws. If theheightof thebox isgiven, the fourthapproachcanbe applied. Similar to the third approach white dots are created in a black image for found screw templates. However, the white dots for found screws are made bigger and compared to an image created in the algorithm, consisting out of ten big white filled dots. Those are placed exactly on the spots, where a level gauge of the given size has located its screws. The comparison is done by sliding the artificial ten dot image over the scene image and adding one to the correlation variable for eachpixel that is white in both images. Thepoint with the highest correlation marks the estimated position of the level gauge’s outer box. To combine the strengths of the algorithms mentioned above, the fifth option is based on line detection of box bor- ders and template matching with screws. Instead of getting just the two best vertical lines, more of them are to be found implementing a Canny edge detector and Hough transform. Next screw templates are found in the scene. Lines, as well as screws are then graded identifying their relative horizontal distances. There have to be a certain number of screws in the vicinity of a line, to mark both of them as good. Fig. 7 shows, how finally good screws and lines are marked in green, other discarded ones in blue and bad ones in red. Fig. 8. Combine the use of template screws and line detection to optimize the result. B. Cut-out of Liquid Column The combination of the box detection methods above lays the foundation for localizing the inner liquid column and cutting it out. As the result image of the box detection contains just the box, the position of the inner liquid column is acquired using height and width of the column in respect to the box size. As the appearance of the box is known, the outer edges of the column are searched for in a specific area, to get detailed borders. C. Level Detection Having acquired and cut out the liquid column, the level is obtained. Allthough there are many different kinds of disturbances when detecting the liquid level, reflections and translucence are the ones affecting the reading the most, as described infig.2.Horizontal reflectionsof the sunornearby objects create horizontal lines, that often are even more prominent than the real water level. Pipes or other objects behind the liquid level gauge also create horizontal gradients in the intensity image of the liquid column that is used to obtain the correct level. To overcome these disturbances, images of the gauge are taken from different heights. As the mobile robot’s arm is restricted to five degrees of freedom, the normal pose of the camera mounted on the arm has to be altered. To achieve readings from different heights, 103