AI at work to predict car accidents

The new Autopilot 8 on Tesla cars harvest data from radar and proximity sensors around the car to create a dynamic model on what "will happen". Credit: Tesla

Taking decision when driving seems a seamless process to us. Actually, we are (almost) able to do that as we are engaging in conversation with our passengers apparently without really paying attention to the road.

Our brain performs plenty of "computations" in parallel and most of that is not perceived unless something triggers our attention. A soda can on the road is unlikely to trigger a reaction whilst a stray cat on the tarmac will surely engage us in some diversion actions. Actually a can or a cat, from a purely safety consideration, won't make that big difference, however hitting a can is not a big deal, hitting a cat is just bad! Our "unconscious" brain takes into consideration all these factors before coming up with a "alarm".

These factors will need to be considered by a car autopilot (or collision alert) system. And this is what the Tesla Autopilot V8 is doing. 

Taking a look at the article on Slashgear describing it is very interesting, at least it was to me, because it points out the "hidden" issues we are usually disregarding when thinking of an autopilot.

Also interesting is the way Tesla is trying to make sense of the surrounding by "cars-sourcing". They harvest data coming from all their cars as they move around and based on that they create a georeferenced data base that will be used by car in a specific location to work out the meaning of data received by the on-board radar. A concave object, as an example, can appear much bigger than it actually is based on radar data (since its concavity spreads the radar beams simulating a much larger reflecting area). 

The extraction of "meaning" from sensor data is crucial for the autopilot that has to be sure of what is going on. Erring on the safe side is ok, but it shouldn't be overdone, otherwise your car will keep braking and you'll get annoyed pretty soon.  

To decrease the false "positive" (no need for braking) and to make sure there are no "false" negative (when you need to brake!) Tesla is getting better at understanding the behaviour of other cars in the surrounding. A good example of its capabilities is shown in the video where the Tesla systems is able to predict an accident one second before it actually happens involving two cars preceding it.

Apparently, autopilots is on the way of getting smarter then we are. Besides, it does not have to engage in conversation with passengers, the same way we do!

Author - Roberto Saracco

© 2010-2018 EIT Digital IVZW. All rights reserved. Legal notice. Privacy Policy.