Fast and accurate cable detection using CNN

摘要

In recent years, unmanned aerial vehicle (UAV) based vision inspections have been widely applied in electricity systems for both efficiency improvement and labor cost saving. Cable detection is essential for both navigation and flight safety of aerial vehicles. However, power cable detection is widely regarded as a challenging task since the targets are very weak and very easy to be confused with cluttered backgrounds. Traditional line and edge detectors are lack of robustness to scene variations. Recent deep learning based methods also can not support fast and stable power cable detection well for onboard applications . In this paper, a new convolutional neural network (CNN) based cable detection method is proposed. First of all, we encode cables by groups of evenly distributed key points, which reduce the complexities of detection tasks. By this approach, the proposed model detect grouped key points of cables from aerial images directly and the detailed pixels of cables can be restore with the curve equations which are implicitly behind those grouped key points. Subsequently, new methods of data labeling and augmentation, sample matching, post clustering, and performance evaluation for cable key points detection are presented. Finally, comprehensive experimental results demonstrate the efficiency and accuracy of our proposed cable detection method.