Analyzing the Fusion of Synthetic and Real Datasets for Dynamic Object Detection in Traffic

The rapid advancement in autonomous vehicle technology and traffic monitoring systems demands robust object detection models capable of performing accurately in diverse traffic scenarios. This paper presents an extensive study on the efficacy of integrating synthetic data, generated from the Carla simulator, with real-world data to train object detection models. We specifically focus on the YOLOv7 algorithm, recognized for its speed and accuracy in real-time detection tasks. Our methodology involved creating datasets with varying ratios of synthetic to real data, evaluating the performance of each using metrics such as precision, recall, mAP@.5, and mAP@.5:.95. The study reveals that models trained on a balanced mix of real and synthetic data, particularly with a 10:90 synthetic to real ratio, outperform those trained solely on real or synthetic data. Additionally, we discover that in scenarios with limited real data, increasing the proportion of synthetic data compensates effectively, enhancing model robustness and generalization. These findings underscore the value of synthetic data in training more effective object detection models for traffic applications.