Page - 106 - in Proceedings of the OAGM&ARW Joint Workshop - Vision, Automation and Robotics

fig. 13, that originally belong to the scenes in fig. 12. It can be seen that there are a lot of wrong level hypotheses at approximately the same height. However, as more column images contain correct level hypotheses, the correct ones predominate and the real level is returned. Testhaveshownthat theHistogram-Gauss-Addingmethod described above mostly outperforms approaches like creating an average image as in fig. 15 using ix= 1 n (ix1+ix2+...+ixn). On the right side of the original image, images show the same situation photographed from different heights with the column already identified and cropped. In the following images the original columns are added equally weighted with linear blending. First the first two columns are added, then the first three, then the first four and in finally all five of them are put into one single image. The found level is marked in red. This clearly shows, that reflections can be suppressed using multiple images taken from different heights. Further improvements are reached using a guided filter. The middle one of the column images is used as a guidance image for thisedgepreservingfilteringmethod.However, thisapproach only works for perfectly alligned images. Therefore it is just used to raise the confidence of the reading, if it delivers similar results as the method using Gauss functions. Fig. 15. Scene image with reflections in the liquid column, extracted liquid columns from images taken from different heights and addition of 2,3,4 and 5 column images Doing tests to compare the different level detection algo- rithms it canbeseen that everyoneof themhas its advantages that make it reasonably useful for a reliable detection. Line detecting algorithms have their strengths in transparent liquids similar to water. The integral over the intensity array performs best for liquids that give a big intensity difference compared to the empty part of the liquid column. IV. CONCLUSIONS The initial task of detecting the liquid level of an analog gauge was reached using an algorithm for locating the outer box in the image, based on canny edge detection, hough line detectionand templatematching.The levelwas thenobtained identifying the horizontal gradients standing out most. The crucial enhancement of the reliability of the process was achieved using multiple images and creating a sum of Gauss functions, each at the position of a level hypothesis. Disruptive effects of sunlight, rain and even objects like pipes shining through can be handled. Despite being cheaper and easier to implement than solutions with flashlight, polar filters or spectral filters, the reached confidence value of the reading can be increased drastically by a small additional arm movement of the robot, taking multiple images. Future detection algorithms may base on this approach to detect other kinds of reflective objects in outdoor conditions. REFERENCES [1] E. Musayev, S. E. Karlik, “A novel liquid level detection method and its implementation” Sensors and Actuators A: Physical, vol. 109, pp. 21–24, Dez. 2003. [2] K. J. Pithadiya, C. K. Modi, J. D. Chauhan, “Machine Vision Based Liquid Level Inspection System using ISEF Edge detection Technique” in Proc. International Conference andWorkshop on Emerging Trends in Technology ((ICWET’10), Mumbai, India, Feb. 2010, pp. 601–605. [3] S. Park; N. Lee; Y. Han; H. Hahn, “The Water Level Detection Algorithm using the Accumulated Histogram with Band Pass Filter” World Academy of Science, Engineering & Technology, issue 32, p. 193, Aug. 2009. [4] T. Hies, P. S. Babu, Y. Wang, R. Duester, H. S. Eikaas, T. K. Meng, “Enhanced water-level detection by image processing” in Proc. 10th International Conference on Hydroinformatics, Hamburg, Germany, Jan. 2012. [5] M. Iwahashi, S. Udomsiri, “Water Level Detection from Video with Fir Filtering” in Proc. IEEE 16th International Conference on Computer Communications and Networks (ICCCN’07), Honolulu, Hawaii, USA, Aug. 2007, pp. 826–831. [6] L. Ding, A. Goshtasby, “On the Canny edge detector”Pattern Recog- nition, vol. 34, pp. 721–725, Mar. 2001. [7] R. O. Duda, P. E. Hart, “Use of the Hough transformation to detect lines and curves in pictures”Communications of the ACM, vol. 15, pp. 11–15, New York, USA, Jan. 1972. 106