Page - 104 - in Proceedings of the OAGM&ARW Joint Workshop - Vision, Automation and Robotics

the camera has to turn around its lateral axis. Beside this pitchingmovement it needs tomovealong thevertical axisas illustrated in fig. 9. These movements result in images taken from different heights, where the reflections and objects behind the liquid levelgaugemovevertically in respect to the liquid column itself. However, the level of the liquid remains at the same height within the column (see fig. 10). Fig. 9. Robot pose to achieve readings of the liquid level gauge from different heights (Camera has to turn around the lateral axis, i.e. pitch and has to move along the vertical axis.) Fig. 10. Taking images from different heights to make reflections and translucence of objects behind the gauge move vertically while the actual liquid level remains at the same vertical position As for box detection, reliability of the reading is of higher importance than speed. Hence three algorithms for level detection are performed and combined resulting in a final value and confidence. The obvious method is the detection of the most outstand- ing horizontal line. Canny edge detection is used, Hough transform is applied and only lines within a certain angle threshold are considered. However, this first part of level detection is not restricted to find just one line, but multiple ones. The reason for accepting multiple level hypotheses for one column is that there can be found horizontal lines within the column that have nothing to do with the real liquid level. The y-positions of the detected level hypotheses are then normalized between 0 and 100 and subtracted from 100 to get the liquid levels in percent. The whole range from zero to a hundred percent is divided in equally sized intervals and a histogram is created. Each level hypothesis in the histogram is then convoluted with a Gauss function. This takes into accout that there might be slight deviations of the real level position in the liquid column images that are cut out in the box image, that is detected and cut out of the original scene image. Fig. 11 shows that the histograms with Gauss filtering are created out of every column image and summed up. To optain the real liquid level, the position of the maximum of the resulting function is found. Fig. 11. Liquid columns with reflections at different heights. The strongest gradients are represented with Gauss functions and added. The final level is obtained by finding the maximum of the sum of the functions describing the columns. The second option gets the average intensity for every horizontal row of pixels in the column image. This array of intensity values having the size of the number of rows in the column image is smoothed with a filter and the following level detection algorithm is performed. All intensity values are normalized to have a maximum value of 1. Starting from the top, a separator divides the intensity values in two parts. Then two integrals are obtained. On one side the integral above the curve is used, on the other side of the separator the integral below is used. As the lower part of the liquid column contains the liquid, it is expected to have the higher intensity. Nevertheless, it is also done the other way round to achieve safe results. The sum of the two integrals is stored in a new array for each position of the separator. The position, where the sum becomes a maximum is selected as the result for the level detection. III. EXPERIMENTS A. BoxDetection Images are taken by the robot outdoors under very dif- ferent weather conditions. Evaluating the template matching approach, it can be shown, that the more the illumination and weather resemble the conditions on the template image, the better it works. When performing approaches three to five, it becomes obvious, thataccording todifferent sizedboxes in the images, the template screws have to be adapted, thus resized to perfectly fit the screws in the scene image. To cover not just frontal images of the level gauge but also those taken from slightly above and slightly below the height of the box center, also screw templates have to be chosen accordingly. The set of templates has to contain screws photographed from different angles. When running the different algorithms for box detection, it showed that every single one of them has its strengths and weaknesses. Different approaches perform best depending on illumination, weather condition, distance of the camera to the 104