Scenario-Based Synthetic Data Generation for an AI-based System Using a Flight Simulator

Abstract

In recent years, algorithms based on machine learning have significantly advanced many technical areas, including computer vision. Since the performance of machine learning applications is data-dependent, a sufficient amount of high-quality data must be available to achieve robust and stable performance. However, the collection of large amounts of real-world data that covers the operational parameters of the AI-based system is often a difficult task because of availability, cost, or even potential danger. Therefore, synthetic data generation is often used to supplement data sets with additional required data samples. In this paper, we propose a baseline for an automated toolchain to generate synthetic image data of aircraft for machine-learning computer vision applications using a flight simulator. Scenario-based approaches have shown applicability to systematically generate valid test cases for system safety evaluation. We leverage a similar approach to generate data for training of AI-based systems. Our approach requires the user to create scenario models using our modelling tool. These models define the operational ranges for a set of parameters that characterize executable scenarios. The scenarios defined by the models are used to automatically produce images from simulations carried out with the FlightGear open-source flight simulator. We distinguish between a static and a dynamic simulation approach. The static approach generates a sequence of independent scenes, while the dynamic approach creates situations that mimic a collision avoidance scenario. With our approach, we can automatically generate large amounts of raw image data covering the relevant parameter ranges based on the models created by the user.