A robust pointer meter reading method for inspection robots in real industrial scenarios

In real industrial scenarios, inspection robots can replace humans to automatically read pointer meters, which greatly improves productivity and safety. However, existing automatic reading methods perform poorly in complex robot operating environments. To this end, we propose an automatic reading method for pointer meters, which can be better applied to inspection robot working conditions. Firstly, we propose a meter detection network, Yolo_Meter, which combines an attention mechanism and an adaptive feature fusion module. This network can accurately locate the meter from the perspective of robots and crop out images that are suitable for automatic meter readings. Secondly, we propose an oriented pointer detection network (OPDNet) to fit the tip position of the pointer precisely. Thirdly, we design a deep neural network OCR_Meter to obtain the scale and unit information of the meter by text detection and a filtering algorithm, which is adaptable to multiple types of meters. Finally, we propose a polar pixel method for locating the main scale lines and design the local angle method to calculate the readings of the pointer meters. Adequate experiments demonstrate the high accuracy and robustness of our method in real-world scenarios, with an average global error of only 0.73%.