Fitting Landmark Detection on High-Voltage Cables

Fittings play an important role in ensuring the safety and stability of transmission lines. Effective detection of these fittings is a prerequisite to upholding their reliability and security. Classic detection techniques only focus on the region and type of fittings, which needs to be improved to find potential problems such as misalignment. We propose a novel method to detect the exact location of the fittings on high-voltage cables. First, a Single Shot MultiBox Detector (SSD) is used to detect the type and region of the fitting. To improve the detection accuracy of small objects, we replace the VGG16 with ResNet as the backbone network and employ Feature Pyramid Network (FPN) for feature extraction. Next, A Stacked Hourglass network is utilized to detect the heatmap of a given fitting image, thereby identifying the landmarks of the fitting and achieving precise localization. The mean average precision(mAP) value increased by 5.2% and the mAP50 value by 4.1% compared with those of SSD classification. In terms of keypoint detection, we achieved an average percentage of correct keypoints(PCK) value of 90.35 across 34 fittings.