A Systematic Accuracy–Efficiency Evaluation of YOLO11 for Fine-Grained Vehicle Detection in Urban CCTV Traffic

Abstract

Fine-grained vehicle detection in urban CCTV imagery requires balancing recognition quality and inference efficiency. This paper presents a systematic empirical evaluation of YOLO11 on the BMD-45 urban traffic dataset using the provided experimental results. The study evaluates model scale, input resolution, augmentation configuration, hardware platform, and class-wise performance. For YOLO11n at 640×640, the reported precision, recall, mAP@0.50, and mAP@0.50:0.95 are 91.40%, 87.20%, 93.10%, and 68.20%, respectively, with a mean latency of 18.7 ms and 53.48 FPS on the reported T4 configuration. Scaling from YOLO11n to YOLO11l increases mAP@0.50:0.95 by 5.40 percentage points but increases latency from 18.7 to 48.6 ms. Increasing resolution from 416×416 to 768×768 increases mAP@0.50:0.95 by 5.30 percentage points while latency rises from 11.8 to 27.9 ms. The supplied hardware results range from 77.52 FPS on an L4 configuration to 5.43 FPS on the reported CPU baseline. Under the supplied conditions, the analysis identifies YOLO11n at 640×640 as a practical low-latency operating point.