Unmanned aerial vehicles (UAVs) have become a popular tool to monitor the equipment condition in industrial Internet of Things (IIoT) with insulator string defect detection as an important application scenario. However, the limitation of energy, which can be affected by UAVs' weight, flight time, data transmission efficiency and tasks computing efficiency, is still an issue hindering the performance of UAVs.
Recently, the edge computing research group at Shenyang Institute of Automation of the Chinese Academy of Sciences (SIACAS) proposed a cloud edge collaborative intelligent method for object detection, in order to reduce the computational load for intelligence computing of UAVs. The study is published in IEEE Internet of Things Journal.
In the process of transmission line inspection based on UAV, some targets such as insulator strings have large aspect ratio. The classical methods of target location and recognition based on deep learning obtain recognition results by classifying the annotation boxes with specific aspect ratio and scale. For the recognition of rotating targets with large aspect ratios, it is necessary to rotate the annotation boxes along the specific reference direction to complete the target recognition.
In order to improve the detection accuracy, the number of reference directions need to be increased, which will greatly increase the calculation of the model, making it unsuitable for UAV patrol and other resource constrained front-end devices. Therefore, how to determine the number of reference directions and how to reduce the high amount of computation caused by multi-directional detection become core issues.
To solve these issues, the researchers from SIACAS analyzed the impact of the extremely large aspect ratio of object on the detection accuracy and the computational load, and established the quantitative relationship between the number of reference directions and the aspect ratio of the target. Based on this, they presented a novel cloud edge collaborative intelligent method for defect recognition of insulator strings.
First, an ultra-lightweight direction estimation method is proposed based on the observation that the shape of targets with extremely large aspect ratios can be approximated by ellipse. Second, a lightweight and reliable insulator string defect recognition method is proposed, in which pixel level high-precision segmentation method is used to obtain the boundary of insulator string, and the defect is identified by the distribution of peak and valley points of the boundary. As the avoidance of target detection along all possible directions, the proposed algorithm can reduce the amount of computation by more than 90% without losing the recognition accuracy.
Experimental results showed that the proposed method can detect the defect of insulator strings with high accuracy, meanwhile has good generalization ability.
This study is the first work to analyze the impact of the extremely large aspect ratio of insulator string on the detection accuracy and the computational load.
52 Sanlihe Rd., Beijing,