Page - 102 - in Proceedings of the OAGM&ARW Joint Workshop - Vision, Automation and Robotics

Fig. 3. Structure of level detection approximatedquitewellusingsensordata.Therefore the size of the box can be estimated and used for further processing. A. BoxDetection As the outer box of the level gauge has hardly any unique features, tests applying feature-based algorithms like SIFT did not bring the results intended. To robustly find the correct location of the box a combination of five approaches is used. For implementation on the robot all five are combined comparing their returned edge points of the box as well as their confidence values. Thereby maximum knowledge of the accuracy of the detected box position in the image can be achieved, that is crucial for further processing and for a final output of a confidence value of the reading. The straight forward approach is template matching, using an image of the whole outer box as a template. The box template is resized using the height of the box in the scene image. Fig. 4 shows, that this works nicely for scenes, where the expected height has just minor deviations from the real one and the image is taken frontal or is made looking like a frontal image by warping in detection preprocessing. This keeps the influences of distortion and rotation low. Furthermore lighting conditions should be similar. The second approach focuses on detecting the outer lines of the box based on Canny edge detection [6] and Hough line detection [7]. Varying the parameters of edge detection and just taking the hough lines that are roughly vertical, Fig. 4. Detecting the outer box with basic template matching works for ideal conditions. i.e. within a certain angle threshold, gives possible lines for the left and right border of the outer box. Using further knowledge about the structure of the level gauge the final vertical border lines are found. The upper and lower line of the box are not as present in the image. The horizontal lines containing the most points in the Canny image rarely concur with those vertical box borders. To overcome this, the scene image is cropped on the left and right side using the found vertical borders. As the grayscale image recieved from prepocessing steps is often warped, there sometimes occurs a black part at the top and bottom of the scene image. If that is the case, the horizontal lines standing out most are the transitions between the real image and the black parts. To solve this, the black parts at the top are filled with the same intensity as the uppermost pixels of the real scene, that can be seen in fig. 6. The black parts at the bottom are filled with thesameintensityas thepixelsat thebottomof the real scene image. In the new image the protruding horizontal lines are the upper and lower box borders. To make sure to correctly distinguish the horizontal borderlines from other remaining horizontal lines within the cropped image, again knowledge about the structure of the liquid level gauge is used. The four intersections of the vertical and horizontal border lines are returned as box edge points. If the box dimensions are given, they provide a further constraint to reliably detect the vertical as well as the horizontal lines and find the correct borders. Fig. 5. Attain vertical borders of the box 102