Comparative Evaluation of YOLO Models on an African Road Obstacles Dataset for Real-Time Obstacle Detection

Public datasets are used to train road obstacle detection models, but they lack diverse and rare object classes found on African roads, negatively impacting the performance of models trained on them. Although attempts have been made to create custom datasets to train road obstacle detection models, they lack the unique challenges posed by African wildlife and livestock commonly encountered on African roads. This leads to poor performance of road obstacle detection systems in the African context. This study presents a custom dataset with rare African object classes and compares the performance of three YOLO models on it using mean Average Precision (mAP). The images were collected from multiple sources to ensure a wide range of scenarios. Offline data augmentation was applied to increase dataset diversity and simulate real-world road scenarios. The models were trained and evaluated, with YOLOv5 demonstrating superiority over the other two models, with an object detection accuracy of 94.68% mAP at an Intersection over Union (IoU) threshold of 0.5 with data augmentation. Offline data augmentation significantly improved all models' object detection accuracy, especially for YOLOv3. The results reveal the effectiveness of the custom dataset and highlight the importance of data augmentation in improving object detection.