IEEE Access (Jan 2024)
Efficient Oil Tank Detection Using Deep Learning: A Novel Dataset and Deployment on Edge Devices
Abstract
This paper investigates the usage of advanced deep learning (DL) approaches in detecting oil storage tanks, which are widely used in the energy sector. It focuses on the evaluation of the latest emerging YOLOv7 and YOLOv8 object detection models and assesses their performance when deployed on embedded edge devices, marking the first deployment of its kind in this field. A comprehensive dataset of 12,948 images and 171,809 annotations is created, incorporating high-resolution Google Earth captures and images from public datasets, including Satellite pour l’Observation de la Terre (SPOT) images. This represents the most extensive collection for this application. The analysis covers YOLOv8 variants from nano to extra-large, and YOLOv7 large, standard, and tiny models. Findings reveal that the YOLOv7 standard model surpasses the YOLOv8 extra-large variant, achieving a precision of 92.41% and average precision (AP) of 86.63% at an intersection over union (IoU) of 50%, indicating a balanced precision-recall trade-off ideal for detecting oil tanks. The models consistently deliver accurate detections even in the presence of significant variations in environmental conditions and object presentation, reinforcing their reliability for comprehensive and adaptable oil tank detection across varied operational scenarios. Additionally, the study examines the models’ effectiveness on high-end GPUs and embedded devices, such as the NVIDIA Jetson Nano and Xavier NX. The latter demonstrates adaptability with up to 16.55 frames per second (FPS) using YOLOv8 nano, indicating efficient real-time monitoring capabilities. This comparison of YOLOv7 and YOLOv8 models advances object detection technology in the energy sector, providing valuable insights for enhancing infrastructure monitoring and risk assessment with a focus on accuracy, efficiency, and flexibility.
Keywords
