Evaluation of Real-Time Train Overhead Line Component Detection on Edge Device

To ensure the safety of train operations, regular inspections and early detection of damage are essential. Currently, inspections are conducted visually by humans during nighttime hours, which incurs significant costs and time. As a solution, the establishment of automated inspection technology using drones is being explored. In this study, as part of this initiative, the focus was on detecting railway overhead equipment using edge devices and improving detection speed. Building upon prior research involving YOLO and ONNX, a dataset was created for training the AI, which was then used to detect railway overhead equipment on a Raspberry Pi. The trained models achieved mAP and F1-scores of 0.959 and 0.96 for Yolov7, and 0.956 and 0.94 for Yolov7-tiny, respectively. Detection times on the Raspberry Pi were 16.43 seconds for Yolov7 and 2.74 seconds for Yolov7-tiny.