A Vision-Based Analog Meter Reading Method for Inspection Robots

Abstract

Computer vision technology has been widely applied in reading recognition of analog meters. However, it is still a challenge to quickly and accurately read various types of analog meters under different environmental conditions. We propose a fast-reading method for analog meters based on keypoint detection, which is applied to inspection robots. First, we use the YOLOv5s network to locate the analog meter. Second, the HRNet network is used to detect the keypoints of the pointer and scale on the dial. Third, an objective image quality assessment method that includes multiple indicators is established to select the optimal image for reading recognition. Finally, we calculate the reading of the analog meter based on the deflection angle of the pointer. The experiment shows that our method can accurately read the readings of analog meters, with an average reading error of 3.81%. It can be effectively applied to inspection robots to read analog meter readings.