Detect and Avoid (D&A) systems play a central role in the operational safety of Unmanned Aerial Vehicles (UAVs), enabling autonomous aircraft to identify and avoid obstacles in real time. However, the training of computer vision models for D&A faces significant limitations due to the scarcity of real-world data, particularly in high-risk scenarios such as near-collisions or adverse weather conditions. Acquiring such data in real environments is operationally hazardous, financially unfeasible, and often ethically impractical. In this work, we propose an approach based on synthetic data generation using high-fidelity graphics engines, specifically Unreal Engine and Blender, which enables the controlled and safe recreation of a wide range of critical scenarios. We detail the pipeline for virtual environment construction, physical flight simulation, dynamic obstacle parameterization, and large-scale generation of FPV videos for training computer vision models. Also, we demonstrate preliminary results in these datasets using Cnvolutional Neural Networks as detectors to locate other drones in the UAV trajectory. The expected outcomes include increased model robustness, improved performance in risk-prone situations, and enhanced sim-to-real generalization. This work lays the foundation for the creation of an open synthetic dataset for D&A in UAVs.