Deep-Learning-Based Insulator Detector for Edge Computing Platforms

Abstract

In the past years, the object detection applications have witnessed a rapid increase in usage of deep-learning (DL) based solutions, due to their accurate object detection, and robustness to illumination, scale, clutter, rotation changes, etc. Therefore, DL-based approaches started to be used in real-time applications. In an autonomous aerial inspection system, the robust detection and time requirement are critical aspects for the real-time perception and decision making. However, most of the DL models are not suitable for the edge computing platforms, due to their heavy sizes and poor reliability. This paper presents the experimental results obtained from a YOLOv4-Tiny model with a CSPDarknet-53 backbone on different single board computers. The study demonstrates that the performance of the adopted approach is highly dependent on the target platform; and the real-time object detection is reachable in specific cases.