Real-Time YOLOv5-Based Object Detection for Autonomous Vehicles on Nigerian Roads

Abstract

Autonomous vehicles (AVs) offer substantial improvements in safety and efficiency; however, global perception benchmarks are primarily based on structured traffic environments in developed countries and do not capture the informal and complex road conditions common in Nigeria and other African nations. To address this gap, the present study developed a custom dataset comprising 8,093 real-world images from Nigerian roads, annotated with 22 contextually relevant classes, including tricycles (Keke NAPEP), Okada motorcycles, street vendors, potholes, animals, and degraded traffic signs. Three lightweight You Only Look Once (YOLO) models (v5, v8, and v11) were trained and evaluated using extensive data augmentation. YOLOv5 demonstrated the highest performance, achieving an mAP@0.5 of 96.8%, a peak F1-score of 0.94, and real-time inference at 82 FPS on an NVIDIA RTX 2060, outperforming YOLOv8 and YOLOv11 while requiring the fewest parameters. Notably, strong results were observed for safety-critical classes (pedestrians: 98.3%; tricycles: 92.6%), while degraded signage and small or occluded objects remained the primary limitations. These findings indicate that high-accuracy, vision-only perception for autonomous vehicles is feasible in resource-constrained, unstructured African traffic environments using low-cost cameras and consumer-grade hardware, provided that a locally collected dataset and robust augmentation strategies are employed. This work establishes a scalable, cost-effective benchmark and provides publicly available resources to support AV development in Nigeria and comparable developing regions.