SELF-DRIVING cars – AI AND VISUALIZATION
Training neural networks on synthetic data
Self-driving technology comes with many benefits for both customers and society, and it's no wonder that all the major players in the automotive industry today are conducting research on how to increase the level of autonomy on our roads. Berge has with its knowledge in computer graphics and machine learning been able to contribute in autonomous drive research by exploring new ways to teach artificial intelligence (AI) how to understand the world around us.
When facing the challenge of autonomous driving, it is no longer an exception to look to AI methods such as deep learning to understand the environment around the vehicle. Techniques that learn from data, rather than relying on rules defined in code by a human, come with powerful possibilities but also new challenges. The product you build will not be better than the data you used to train it – hence, data is key. Acquiring enough data with the right characteristics is an extensive task. Annotating it for AI training is then another big challenge.
These challenges are not exclusive to the automotive industry, and lately there has been a trend to use synthetic video data to reduce the need for real data when training different AI-based vision algorithms. Realistic computer images can be generated in real time from video games, and promising research has been done using the commercial game Grand Theft Auto V which is built around a large and visually realistic open-world environment, focused on the first-person perspective of cities and city traffic. The years spent developing the visual realism of the game may give it advantages when used for AI training compared to other solutions. However, commercial games, like GTA V, have drawbacks. Unless you get access to the source code, you are limited in what data and what annotation you can generate.
Berge is now helping a customer to understand the requirements of a virtual environment for use in creating algorithms for autonomous driving. With our experience in computer graphics, game development, deep neural networks, and autonomous driving, we have developed a tool tailored for training AI vision algorithms. Additionally, we have implemented a vision algorithm using deep learning and a method using that algorithm to evaluate different data sources.
The work is ongoing and experimental, but the benefits are clear. Availability of a sufficiently large and varied synthetic world can save a lot of time and money in the development of vision algorithms for the automotive industry. It could even be key to becoming the world leader within autonomous driving.