Benchmarking YOLO-based deep learning models for real-time object detection in hybrid ADAS and intelligent transportation systems
Mohammed Chaman, Anas El Maliki, Hamad Dahou, Abdelkader Hadjoudja
Abstract
• Systematic benchmarking of YOLOv8–YOLOv12 (2023–2025) for ADAS and ITS. • Unified 42,000-image dataset with 16 object classes for fair multi-class evaluation. • Comprehensive metrics: mAP@50, mAP@50–95, precision, recall, F1-score. • YOLOv12 shows highest accuracy and robustness; YOLOv10 provides lowest latency. • Results guide deployment for TSR, PDS, VDS, CAS, and ISA in real-time systems. This study presents a comprehensive benchmarking analysis of YOLOv8 through YOLOv12 to evaluate their performance in real-time object detection for Advanced Driver Assistance Systems (ADAS) and Intelligent Transportation Systems (ITS). The primary goal is to provide a fair and reproducible comparison of recent YOLO architectures (2023–2025) and determine their suitability for critical perception subsystems, including Traffic Sign Recognition (TSR), Pedestrian Detection Systems (PDS), Vehicle Detection Systems (VDS), Collision Avoidance Systems (CAS), and Intelligent Speed Assistance (ISA). A unified dataset of 42,000 annotated images covering sixteen traffic-related classes was constructed by integrating subsets of BDD100K, CCTSDB, and custom samples, standardized in YOLO format using the Roboflow platform. All models were trained under identical experimental conditions and evaluated using precision, recall, F1-score, mean Average Precision (mAP@50, mAP@50–95), confusion matrices, inference time, and frame rate (FPS). The results demonstrate clear generational progress, with YOLOv12 achieving the highest overall accuracy (precision = 97.3%, recall = 96.2%, mAP@50–95 = 82.2%) and an inference rate between 70–73 FPS, confirming its robustness and real-time capability. YOLO11 delivered comparable accuracy with slightly higher latency (62–65 FPS), while YOLOv10 achieved the fastest inference (68–70 FPS) with only a marginal drop in accuracy. YOLOv9 and YOLOv8 showed stable yet relatively lower recall and generalization performance. YOLOv12 provides the best trade-off between accuracy, inference speed, and computational complexity, making it the most balanced solution for embedded ADAS perception.