Monitoring Infrastructure Faults with YOLOv5, Assisting Safety Inspectors

Abstract

Recent advances in AI technology paved a wave for real-time computation applications like remote infrastructure inspection. In this paper, we propose preliminary test results of a YOLOv5n-based algorithm diverse enough to visually identify faults in electrical insulators, roads, pavements and concrete, an approach for a single solution for different infrastructural fault inspection in smart cities. This is achieved using a modified algorithm using YOLOv5n, wherein there is a master detection algorithm (MDA) that drives a slave detection algorithm (SDA). The MDA broadly classifies the visual data into different broad categories and transfers the data to the SDAs - fine-tuned algorithms for detecting faults in a single category. Furthermore, using the YOLOv5n made the algorithm lightweight in order to be implemented on an AI device - NVIDIA Jetson Nano. The implemented algorithm resulted in detection time of 90 ms with an overall accuracy of 93 %. © 2023 IEEE.