MRTMD : a multi-resolution dataset for evaluating object detection in traffic monitoring systems

Abstract

Traffic monitoring reduces congestion, improves safety, and supports environmental sustainability. Real-time flow tracking, anomaly detection, and efficient management are key. Convolutional Neural Networks (CNNs) have become integral due to their compact size and easy deployment. However, their effectiveness depends heavily on the quality of the input data, especially image resolution. With highresolution cameras, especially 4K, balancing image quality, detection accuracy, and system efficiency is critical. We propose the Multi-Resolution Traffic Monitoring Dataset (MRTMD), which captures transport scenes at resolutions ranging from 2160p to 360p. This dataset serves as a benchmark for standard object detection models, enabling the development of more efficient and cost-effective traffic monitoring solutions. MRTMD will be freely available on GitHub, offering a valuable resource for researchers and practitioners. We evaluate leading object detection models—YOLOv9, YOLOv8, YOLOv7, Faster R-CNN, FCOS, SSD, and RT-DETR—across varied resolutions. Our analysis focuses on mean Average Precision (mAP), recall, and processing time. We also assess the accuracy of Number Plate Recognition (NPR) for tasks that require fine-grained detail extraction. Our findings show that detection performance typically varies within ±0.01 to ±0.03 in mAP and recall across resolutions, suggesting higher resolutions are not always advantageous. However, they remain crucial for tasks like NPR. The multi-resolution dataset enables a comprehensive evaluation of the trade-off between image quality and task performance. Ultimately, our analysis highlights the importance of resolution selection in large-scale deployments, informing system designers and policymakers. This dataset is a vital tool for balancing performance, cost, and practical constraints in real-world traffic monitoring.