Vehicle Detection from Overhead Satellite Images in Saudi Arabia with YOLOv12

Abstract

One of the primary challenges in urban areas of the Kingdom of Saudi Arabia is managing traffic flow and implementing effective urban planning strategies in densely populated cities. Utilizing satellite imagery for vehicle counting presents an effective solution to enhance urban planning and economic monitoring. Based on a review of the literature, the lack of vehicle counting datasets using satellite imagery is a significant gap. To address this, we introduce the Saudi Vehicles Dataset, which comprises 410 images, each 500x500 pixels, with 5,609 manually annotated vehicles. To ensure the dataset's quality and generalization, we benchmark it against two other satellite datasets: Car Synthetic and COWC. Benchmarking was performed using the latest version of the YOLO object detection algorithm, YOLOv12. The results indicate that our dataset yields comparable performance metrics to the other datasets. Specifically, when YOLOv12 was trained on the Saudi Vehicles Dataset, it achieved the high scores for mAP50, recall, precision, and F1-score, with values of 0.977, 0.939, 0.945, and 0.942, respectively. While the best performance was observed on the synthetic dataset, achieving a perfect precision score of 1, the Saudi Vehicles Dataset showed superior performance in terms of recall, precision, and F1-score. The dataset is publicly available for use and can be accessed via this link: https://github.com/Syukrondzeko/saudi-vehicle-dataset