Industrial Manometer Detection and Reading for Autonomous Inspection Robots

Abstract

Autonomous mobile robots for industrial inspection can reduce cost for digitalization of existing plants by performing autonomous routine inspections. A frequent task is reading of analog gauges to monitor the health of the facility. Automating this process involves capturing image data with a camera sensor and processing the data to read the value. Detection algorithms deployed on a mobile robot have to deal with increased uncertainty regarding localization and environmental influences. This imposes increased requirements regarding robustness to viewing angle, lighting and scale variation on detection and reading. Current approaches based on conventional computer vision require high quality images or prior knowledge. We address these limitations by leveraging the advances of neural networks in the task of object detection and instance segmentation in a two-stage pipeline. Our method robustly detects and reads manometers without prior knowledge of object location or exact object type. In our evaluation we show that our approach can detect and read manometers from a distance of up to 3m and a viewing angle of up to 60° in different lighting conditions with needle angle estimation errors of ±2.2°. We publish the validation split of our training dataset for manometer and needle detection at https://tudatalib.ulb.tu-darmstadt.de/handle/tudatalib/2881.