Object detection of Train Line Equipment using SSD and YOLOv7

In modern society, the early detection of damage to train line equipment and regular inspections were essential for ensuring the safe operation of trains. Traditionally, inspections of train line equipment were conducted manually at night. However, this method posed issues such as the need for significant labor and limited mobility. To address these problems, efforts were made to improve mobility by installing cameras on the roofs of test vehicles. However, the analysis of the images captured by these cameras was still conducted manually. In recent years, the use of drones, high-resolution cameras, and AI for inspections had been explored. This paper reported on a preliminary study concerning damage detection of train line equipment using deep learning and its application to autonomous drones. In this study, two models were used for object detection of train line equipment. The target equipment in this study included three components: hanger, connector, and insulator. The dataset contained a total of 2,400 images, with 800 images each for the hanger, connector, and insulator. To evaluate the training progress of the models, the mean Average Precision (mAP) was used for both the training and test datasets. The training was conducted on a Windows PC with CUDA support. The object detection models used were the Single Shot multibox Detector (SSD) and You Only Look Once (YOLOv7). After simulations, the following results were obtained. For the training data, the SSD model achieved an mAP of 73.51%, while the YOLOv7 model achieved an mAP of 97.5%. For the test data, the SSD model achieved an mAP of 53.74%, while the YOLOv7 model achieved an mAP of 76.30%.