The article presents the development of a system for simulating the behavior of groups of forest animals in the three-dimensional environment of Unreal Engine 5 in order to generate synthetic data for computer vision tasks. The simulation includes realistic animal behavior, group dynamics, and the use of a virtual drone with a camera to simulate aerial photography. Annotations are automatically generated for each object, suitable for subsequent training of neural networks. The experiments demonstrated the effectiveness of mixed datasets containing both real and synthetic images when training the YOLO model. The results obtained confirm that the generation of synthetic data can significantly improve the quality of learning with a limited amount of real data. The experiments demonstrated that adding 30% synthetic data to real datasets increased mAP50 from 0.1866 to 0.3605, while Precision improved from 0.7529 to 0.8411. In contrast, training solely on synthetic data reduced mAP50 to 0.0286, confirming the necessity of combining real and generated data. The results indicate that synthetic data can significantly enhance training effectiveness under conditions of limited real-world images.
Vorobyev et al. (Mon,) studied this question.