Industrial Meter Reading For Patrol Robot Of Electric Power Substation

Power patrol robots are designed to detect and read various electric meters.How-ever,traditional methods always fail in face of complicated on-site conditions.In this paper,we propose an efficient method for meter detection and display reading.This method is robust to different pose,illumination,scale and partial occlusion.This paper has three contributions as below:1.We propose a two-level cascaded parameter regressor for aligning any target me-ter image to the same pose and scale.This algorithm performs much better than traditional methods which are based on key points matching in alignment rate,processing speed and location precision.2.We propose a coarse-to-fine approach for meter detection.It consists of a coarse detector,a two-level cascaded regressor and a post verifier.Our method achieves true positive rate at 95.5%,while keeping the false positive rate at 1.35%.The realtime performance is also excellent with 15.5 frames/s.3.We propose different methods for meter digits recognition and pointer display reading.The digits recognition method makes use of template matching,skip-ping the trouble of extracting digits area by binarizing images.With geometry constraints,the accuracy of digits recognition is improved from 94.3%to 95.50%.The pointer display reading method transforms the sector region to a smooth rect-angle image by bilinear interpolation algorithm,and then detects the pointer with HOG feature by sliding windows.This method achieves pointer detection rate at 95.6%with keeping the relative error of display reading below 6.6%.