An accident is always a disgrace, but until now we attributed them to human errors and they passed almost without transcendence.
When something becomes recurrent, like entering a roundabout like Carlos Sainz in a rally, we just do an awareness campaign, we complain about each other and… we keep falling on the same rock.
But that’s it. We are what we are, human individuals and our brains are not connected, so everyone learns on their own according to the experiences they have had.
And the worst doesn’t stop there, some even from their experience influence others on how to react in certain situations and these are not always applicable so again, they translate into an accident. How many times have common sense or bad information made us do something wrong with the car?
All this now has a solution, or rather, a potential for improvement that today is feasible thanks to technology. Yes, Tesla, like many others, are using this technology to take advantage of all these experiences and build a system that knows which is the best reaction to an event and thus avoid an accident.
Yes, we are talking about Autopilot, but what is it based on? How can a computer take advantage of the experience of all drivers and improve? Well, this technique is called Deep Learning and I will try to explain in a very simple way how it works.
We could say that it all starts with image recognition. We need something that interprets what the cameras see and translates it into objects. A car, a truck, a traffic light, road signs, a cyclist,… all this, in all the shapes, colours and typologies that exist in each and every one of the countries.
On the other hand, it is necessary that, this same system together with some as it can be a radar, gives us more information as the speed to which they move, in which direction, if they are accelerating, braking, their possible trajectory, …
This is, as you can imagine, a complex system, which must also be installed in the car and I tell you that you need a VERY powerful computer to identify everything in a matter of 1 second or less. It needs to be as fast as possible so as not to hinder the rest of the system.
With all this information, the system is silly. It doesn’t know what to do, so you have to use Deep Learning to create the necessary “reaction algorithms” for each of the data.
This system sends all that information of what it sees, identifies and measures the “cloud”. A supercomputer stores all that information, but also how each driver has reacted to each of those situations.
With all this, Deep Learning looks for that pattern or mathematical function to know what to do in each situation. That is to say, and to simplify it a lot, imagine that we have 2 points in space (identified by a coordinate x and y) and we make a linear regression. That supposed line that passes through the 2 points.
If instead of 2 points, we have 3 or more, perhaps a line is not the most adequate and we would have to look for a curve or another type of figure to try to go through all the points. The difference between the point and the figure we are looking for would give us the precision of the system.
Each value of X would be each of our driving situations. At what speed are you going, what vehicles are around, if there are traffic lights, … The value of Y would be what each driver does in that situation.
At this point, you can imagine that X is not a number and in fact, are a lot of different variables that define each of the possible situations that we are given in driving.
With all these values, both situations and reactions of each driver, the system builds that function. This is what is called “training” and takes quite some time taking into account the amount of information.
This function is the one that in the updates is sent to the cars so that they don’t have to calculate it every time. In this way, each time you activate the Autopilot, the system already has that mathematical function that once interpreted what the cameras see, knows what to do. If X is given, do Y.
This is why the more information the system has, the better it interprets and evaluates, and the more different drivers have been exposed to different situations, the better it will know how to react and the more accurate it will become.
It even allows the “luxury” of anticipating the layout of the curves without having to see them. It even anticipates some accidents without being noticed.
It’s only a matter of time before the system gets the best out of each driver in different situations and knows how to drive better than any of us.
Article translated using DeepLearning thanks to DeepL.com 😉