Challenges in creating an experimental database for artificial intelligence: The EDF Store project

Within the framework of the EDF STORE project, a key objective is to collect and annotate images related to operational use cases, with the aim of evaluating and demonstrating the contribution of Artificial Intelligence (AI) in detecting threats on the battlefield. This initiative highlights significant challenges in creating an effective experimental database for training and testing AI algorithms. Two data acquisition campaigns were organized in France and Norway, two countries at different latitudes, offering a variety of environmental conditions (…), diverse backgrounds (urban, forest, lake, open field,…) to test AI robustness and varied weather conditions (sun, cloud, rain, snow, …), to simulate realistic conditions. The targeted objects were 6 military vehicles, to evaluate detection in various tactical contexts, 4 types of drones (given their increasing use as potential threats), and pedestrians. Up to 30 pieces of equipment were deployed to record scenes in visible spectral band, to capture visual details and in infrared spectral band, to detect
thermal signatures, useful especially at night or in low visibility conditions. Although videos are rich in information, they are insufficient alone for effectively training AI algorithms. The challenge for AI is to create metadata including: - Annotations: Precise labeling of objects and scenes. - Timestamps: Synchronized temporal information. - Localization: Geolocation data to contextualize scenes. - Weather Conditions: To understand the impact on detection. Creating an experimental database for the EDF STORE project highlights the complexities of collecting and annotating data, essential for effectively training and testing Artificial Intelligence algorithms in operational scenarios. Overcoming these challenges is crucial to demonstrate the added value of AI in detecting threats on the battlefield.