Region-Specific Road Object Detection in African Environments: Dataset Curation, Benchmarking, and Edge Deployment

Road safety in Africa faces substantial challenges that current public and custom datasets for autonomous driving do not adequately address. Public datasets frequently underrepresent object classes crucial to road safety in the African region, which features unique obstacles such as wild and domestic animals, informal public vehicles, and infrastructure hazards. Without these classes, detection models often miss road obstacles, thereby reducing safety on African roads. This study introduces a custom dataset of 3,236 original images with annotations for 11 object classes. It includes wild animals (kudus, elephants), domestic animals (cattle, goats), informal public vehicles (boda bodas, tuk-tuks, minibus taxis), and infrastructure hazards (potholes, unmarked speed bumps). Three of the most recent You Only Look Once (YOLO) object detection models (versions 8, 9, and 10) were trained on our custom dataset and evaluated on an NVIDIA RTX 3050 GPU. Of all the models, YOLOv8 achieved the highest mean Average Precision (mAP)@0.5 of 94.2%, followed closely by YOLOv10 and YOLOv9 at 92.5% and 90.6%, respectively, demonstrating strong detection performance on challenging, geographically relevant data. To assess the models in real-world deployment, they were optimized with TensorRT and deployed on an NVIDIA Jetson Nano embedded platform. This optimization achieved a 77% reduction in inference time with a very small accuracy drop of 0.64%, proving that the models are capable of fast, accurate, real-time execution on inexpensive edge devices. This study addresses geographic bias in autonomous-driving datasets in Africa and offers deployment-ready solutions for inexpensive edge hardware.