Performance Benchmarking of YOLOv11 Variants for Real-Time Delivery Vehicle Detection: A Study on Accuracy, Speed, and Computational Trade-offs
Rabinandan Kishor
*
Department of Commerce & Management, Sunrise University, Alwar, India and P.G., Artificial Intelligence and Machine Learning, Purdue University, West Lafayette, USA.
*Author to whom correspondence should be addressed.
Abstract
The YOLOv series represents state-of-the-art technology for single-stage object detection, excelling in speed and accuracy. In many scenarios, it outperforms traditional two-stage detection frameworks, making it ideal for real-time applications. This study evaluates YOLOv11 model variants (n, s, m, i, x) on a custom dataset of 2,285 labelled images representing four delivery vehicle classes: FedEx, Other-Vehicles, UPS, and USPS-Truck. The dataset is meticulously curated to capture diverse delivery vehicle scenarios and split into training, validation, and test sets. Each variant was fine-tuned using uniform settings: 20 epochs, an input resolution of 640×640 pixels, and a batch size of 16.
Performance was assessed using metrics such as mean Average Precision (mAP, a standard metric measuring detection accuracy) across Intersection over Union (IoU) thresholds from 50% to 95% (a range defining the overlap between predicted and ground-truth bounding boxes), precision, recall, and inference speed on GPU and CPU. The results highlight trade-offs between model complexity and performance: smaller variants like YOLOv11-n achieved faster inference speeds (170.74 FPS on GPU and 5.86 ms on GPU), while larger models like YOLOv11-x excelled in detection accuracy and recall but at the cost of slower speeds (240.03 FPS on GPU and 4.17 ms on GPU). YOLOv11-s, for example, offered a balance with the highest FPS (1120.46 GPU FPS) but with moderate accuracy and recall. These findings demonstrate the adaptability of YOLOv11 variants to varying application requirements, from high-speed real-time systems to scenarios prioritizing detection accuracy.
This research advances object detection by providing a detailed performance benchmark for YOLOv11 variants. It offers practical insights for deploying YOLOv11 in diverse fields, including logistics, delivery tracking, and other domains requiring efficient and accurate object detection.
Keywords: YOLO, YoloV11, object detection models, deep learning computer vision, neural Networks, dataset evaluation, model performance metrics, image processing real-time applications