How do you develop high-quality AI without compromising on privacy? Take a look inside the Project Hafnia engine room to see how Milestone puts responsible AI into practice.
Developing high-quality AI and protecting people's privacy should go hand in hand. That is the principle at the heart of Project Hafnia.
Led by Milestone Systems, Project Hafnia develops computer vision AI for real-world challenges. Project Hafnia uses real-world data of high quality to develop AI solutions that are specialized in specific settings such as understanding traffic conditions or identifying events in video. These data are carefully selected and prepared, with data-protection safeguards such as anonymization and annotation built into the process from the start and guided by the GDPR and EU AI Act. The ambition is to enable developers to create AI that is useful, reliable and capable of understanding the complexity of real environments.
So, what does this look like in practice?
How Project Hafnia prepares data for AI
- Selecting the right data from trusted sources Project Hafnia uses data supplied by data partners alongside datasets, under the terms that apply to each source. The steps below describe how partner footage is prepared for AI development.
- Use only the data that is neededNot all video is relevant to every AI task. Project Hafnia selects data based on its purpose and its value for developing and evaluating AI. Real-world conditions matter: road layouts, lighting, weather, traffic patterns and camera angles all influence how well an AI model performs.
- 3. Apply anonymization Before partner-supplied footage is used for AI training, advanced anonymization technology from Milestone’s brighter AI is applied. Identifying features, such as faces and license plates, are anonymized while preserving the real-world characteristics of the scenes that are relevant for AI development.
- Remove the original recordingsOnce the processing is complete, the original, raw recordings are deleted from Milestone's processing environment.
- Label the data so AI can learn from itVideo becomes useful for AI training when the relevant objects and events are annotated. This means adding information that helps a model learn what it is seeing, whether that involves vehicles, pedestrians, movement or traffic-related events. Annotation is done by trained personnel. Processing takes place in controlled environments, access is restricted to authorized personnel with a legitimate need to use the data, and organizational and technical safeguards help protect it.
- Use the prepared data to develop better AIThe resulting annotated data can be used to train, test and improve computer vision models. By working with data that reflects real-world conditions, engineers can build AI that is better equipped to handle the complexity and variation of the environments in which it will be used.
- Train the model without handing over the data
This allows developers to benefit from high-quality real-world data without receiving access to the underlying datasets.
Turning data into insights
The purpose of this process is not simply to prepare data. It is to help develop AI that can address real challenges in the world around us.
From traffic management and road-safety analysis to urban planning and clearer situational summaries for operators, Project Hafnia aims to make high-quality data useful for solving practical problems.
