The paper proposes a system for modeling the collective behavior of flocks of birds to generate synthetic data in Unreal Engine 5. The system provides flexible adjustment of flock dynamics parameters, including species, speed and flight radius, as well as simulation of complex interactions in a virtual environment. Automated data markup is performed in YOLO (You Only Look Once) formats and with the preservation of spatial metadata of objects. A comparative analysis of the influence of the proportion of synthetic data in training samples on the quality of detection of computer vision models using the example of YOLOv11 is carried out. The results showed that the inclusion of up to 60 to 80 percent of synthetically generated data in the training dataset allows achieving an optimal balance between the accuracy of the model and the cost of collecting real data. The experiments showed that combining real and synthetic data significantly improves detection quality. For example, with 20% real and 80% synthetic data the YOLO v11 model achieved mAP@0.5 = 0.7157, Precision = 0.8157, and Recall = 0.6396, which substantially exceeds the results obtained when training only on real images. The developed system opens up prospects for application in the tasks of airspace monitoring, environmental research and virtual reality.
Sharin et al. (Mon,) studied this question.