Pin Bolt State Identification Using Cascaded Object Detection Networks

Abstract

Unmanned aerial vehicle-based transmission line inspections produce a large number of photos; significant manpower and time are required to inspect the abnormalities and faults in such photos. As such, there has been increasing interest in the use of computer vision algorithms to automate the detection of defects in these photos. One of the most challenging problems in this field is the identification of defects in small pin bolts. In this paper, we propose a pin state identification framework cascaded by two object detectors. First, the bolts are located in the transmission line photos by an initial object detector. These bolts are expanded in the original picture and cropped. These processed bolts are then passed to a second object detector that identifies three states of the pins: normal, pin missing, and pin falling off. The proposed framework can attain 54.3 mAP and 63.4 mAR in our test dataset.