Comparing YOLO Models for Self-Driving Car Object Detection

For self-driving cars, precise object detection is needed to ensure safety. In this paper, three different YOLO models, YOLOv5, YOLOv8, and YOLOv11, are tested on the KITTI dataset to find out which one shows the best balance between accuracy, speed, and ease of use. YOLOv5 uses older methods and gets a mAP@0.5:0.95 of 0.49. It trains fast but sometimes overfits the data. YOLOv8 adds new features like dynamic labels and better loss functions. This improves its mAP@0.5:0.95 to 0.549, especially for small objects. YOLOv11 uses deeper networks and dynamic adjustments, reaching a mAP@0.5:0.95 of 0.546. The results show that YOLOv8 and YOLOv11 are better than YOLOv5 in accuracy, and YOLOv8 is slightly better than YOLOv11. Although YOLOv5 has lower accuracy, it has a fast training speed and is suitable for practical use. Although YOLOv11 has high accuracy, it requires more computational resources during training, which may limit its deployment on low-performance edge devices.