Evaluating Fisheye Vehicle Detection in YOLO Using Multi-Strategy Data Augmentation

Fisheye cameras are increasingly used in traffic surveillance due to their ultra-wide field of view, yet their inherent geometric distortion poses challenges for conventional object detection models. This study investigates whether multi-strategy data augmentation can enhance the performance of YOLOv11 and YOLOv12 in detecting vehicles within fisheye images. A dataset of $\mathbf{1, 0 9 3}$ annotated fisheye images from Jakarta traffic scenes was used, incorporating 12 object classes with over $\mathbf{5 1, 0 0 0}$ labeled instances. The training pipeline applied a combination of geometric, photometric, and occlusion-based augmentations to improve model generalization. Experimental results show that models trained with augmentation significantly outperformed their non-augmented counterparts, with YOLOv12 achieving up to $85.1 \% \text{mAP} {@} 0.5$ and 80.9 % recall. The analysis confirms that augmentation enhances robustness and convergence stability without requiring modifications to the model architecture or preprocessing steps. These findings suggest that training-time augmentation is a practical and effective strategy for improving object detection in fisheye imagery and can be readily integrated into real-time detection pipelines without architectural modification.