Seite - 101 - in Proceedings of the OAGM&ARW Joint Workshop - Vision, Automation and Robotics

Reading of an Analog Liquid Level Gauge on an Oil Platform with a Mobile Robot using 2-D Images Peter Heno¨ckl1 Abstract—An approach to automatically read oil platforms’ liquid level gauges, originally designed to be read by human operators is presented in this paper. Grayscale image data is acquired from different heights to enhance reliability and minimize deviations due to outdoor influences like reflections and translucence. The position of the level gauge in the scene image is determined, the liquid column is extracted and the level of the liquid is returned using image processing methods. I. INTRODUCTION Measuring the level of a liquid in a container is seen as a solved task. However, as the gauge may not be altered in any way, conventional methods of detection using ultrasonic, magnetic, mechanical, pneumatic, conductive, microwave or capacitive sensors cannot be applied. New optical methods as reciprocally placed photo-LEDs and transistors described in [1] look promising, but as there is no possibility to reliably getbehind thegauge, the sensorchosenhere is a2-Dcamera. To acquire the liquid level of the level gauge a mobile robot (fig. 1) approaches the level and captures the gauge taking images. Although camera based level detection is already greatly described e.g. in [2], not having a closed environment with a correctly positioned bottle on a conveyor belt, brings a big increase in complexity similarly found in [3], [4] and [5]. Coping with different lighting, reflections, backgrounds, objects shining through or alternating weather conditions and locating the level gauge in the scene image brings new additional challenges. The acquisition of the level is done in three consecutive steps. First the position of the level gauge in the image is determined. To improve the reliability of the following level reading and reduce influences as reflections or translucence of objects in the background, that can be seen in fig. 2, this step is performed repeatedly using different images. Compared to tests with polarized filters and usage of a flash combined with a very short aperture time of the camera’s shutter, using multiple scene images makes the biggest enhancement in readability and reproducibility of the same quality. Images of the same level gauge are taken from different angles. As the level is a horizontal feature, horizontal disturbances have a much higher influence than vertical ones and are to be compensated. To achieve this the height of the camera is altered. Secondly warped images of the liquid column are created giving optimal conditions for level detection. Based on these images an estimation of the 1Peter Heno¨ckl is with Faculty of Electrical Engineering, Au- tomation and Control Institute, TU Vienna, 1040 Vienna, Austria peter.henoeckl@gmx.at Fig. 1. Mobile robot approaching the liquid level gauge and positioning its arm to take a picture for level detection using image processing. Fig. 2. Reflections (left) and translucence of a pipe behind the liquid level gauge (right) are examples for challenges for a correct reading. level follows accepting multiple level hypotheses for each column. II. METHODS Fig. 3 shows the structure of the proposed method. Due to preprocessing steps in other parts of the overall code, grayscale images are the basis for the detection. Changing colors of the liquid in the gauge make RGB images of minor importance. The position of the level gauge in the 3D space is known and the pose of the mobile robot and its arm can be 101