We are a dedicated team with deep knowledge and diverse experience in the automotive industry, autonomous driving, deep learning and safety.
Our team has built prototype autonomous vehicles from the very beginning of autonomous driving. For their homologation and certification for public roads, we have passed the assessment by TÜV Nord (the German Technical Inspection Association).
We have been researching AI safety for years. We began our journey long before AI safety was a publicly funded research topic. Today, we’re partners with OEM and various suppliers to the automotive industry. Based on our past accomplishments our research is now funded by the German Federal Ministry of Economics and technology as well as the European Union.
The Very Basics
Deepsafety is guided by the vision of true autonomous driving. Let’s see what this is all about from an engineering point of view: For an autonomous car to safely move around in the world, it needs to understand it’s surroundings. To build such a functional robotic model, the system needs to sense, plan and act.
Autonomous Cars use sensors to collect information about the environment. E.g: Cameras, Radar Systems and Lidar.
Analysing the Information
Collecting information is the simple part. Now you have to make sense of the information gathered. Therefore you need artificial intelligence. Until 2012 it was not possible to differentiate a child from a plastic bag. You cannot program a computer to do that. Only deep neural networks can learn to make these distinctions and identify objects correctly.
This is Us
To build a safe and functioning perception system, you have to combine sensors with an AI-based object detection system. This is where Deepsafety develops its technology.
The Situation 2022
Let’s have a look at the situation of self-driving cars today. Most people seem to think that self-driving cars already exist. That’s simply not true. Todays autonomous cars use some kind of driver assistance systems. In such systems the driver needs to be vigilant. We shall have a look at the various levels of autonomous cars, also called the SAE levels.
In the future, city traffic poses the biggest challenge for autonomous vehicles. In the city, autonomous cars will have to deal with a lot of difficult situations.
What’s the Problem?
So what’s the big deal to build a safe and effective perception system? To understand this, we have to look at a fundamental problem of AI. Let’s look at the following example: Say you want to train an AI to classify bears and zebras. Easy right? You just train the neural network with millions of images of bears and zebras.
And then …!
… then you give it a picture of a panda bear. The AI will classify the Panda either as bear or a zebra, because it only works with the classes it knows. It cannot identify the unknown.
…the problem is there will always be new stuff on our streets, which means there will always be unknown objects. Some of which can pose a danger in traffic.
How not to solve this …
Most attempts to solve this problem are based on workarounds. But it is futile to train for more classes, as there will always be new elements in the world.