Two big computer vision papers boost prospect of safer self-driving vehicles Like nuclear fusion and jet-packs, the self-driving car is a long-promised technology that has stalled for years - yet armed with research, boffins think they have created potential improvements. Citizens of Phoenix, San Francisco, and Los Angeles are able to take one of Waymo's self-driving taxis, first introduced to the …
Yeah I have yet to hear a report of a self driving car having an accident where it reacted too late. They crash because they did the wrong thing - like the Teslas that hit trucks crossing the road or parked emergency vehicles. They're going full speed when those accidents occur, they aren't trying to brake at the last second but lacked sufficient time/distance to stop.
Having an infinitely fast reaction time doesn't help if you aren't reading the situation correctly in the first place.
Having a faster reaction time gives the system more time to read the situation correctly.
Of course a lack of time isn't the root cause of all errors in reading the situation, but a system that has more time can be built to use more and varied methods to make a decision.
It's not the answer but its plausibly part of the answer.
So how many accidents do we see with self-driving vehicles, and how many accidents do we see with drunk taxi drivers? That would only be a joke if I had seen any reports of taxi drivers getting prosecuted for any kind of driving accidents, I haven't seen any reports but my concern is that every time I drive through Wales I'm happy with the single lane roads because if I see anyone driving towards me in Wales, both vehicles normally slow down or stop and pull over to let the other car gently drive past.
How would a self driving vehicle react on a Welsh single lane road?
An increasing number of humans are having difficulty driving round the single lane country roads round here. They don't know how to reverse, they don't know where the passing places are, they don't want to risk the meerest hedge scratch on their shiny hire purchase car.
"So how many accidents do we see with self-driving vehicles, and how many accidents do we see with drunk taxi drivers?"
Loads of people are lousy at playing the piano, so let's abandon all those Broadwood grands and bulld nothing but paper roll player pianos instead.
If playing the piano badly caused people to die, I'd seriously have to consider whether it was worth it. There are many things that have been automated because they have safety implications. In some cases, the older machines are so dangerous that they are no longer considered legal to operate in the way that they once were.
This is far from saying that cars have reached that point or that the replacements are sufficient, because so far the replacements are not good enough. Comparing deaths caused by human drivers to bad pianists is a faulty comparison, and I think you already know it.
Just remove all humans. Remove motorcycles, bicycles, rickshaws, unicycles, pedestrians, anything that's controlled by a human. Computers can't handle deviations from protocol, and humans just can't reliably follow protocols, so just eliminate the uncontrolled variable.
I was in the charming city of Naples last year, and in Texas this year (Austin, Fredericksburg, and Stonewall for the eclipse). I can see self-driving cars work in the generally quite disciplined traffic in the USA, but see some serious challenges in the, let's say rather more colourful driving style in Naples. Only transmitting changes in the image could potentially saturate bandwidth in the kind of pandemonium I observed there. Assessing the situation and planning for the seemingly erratic behaviour of the myriad cars, motorbikes, cyclists and pedestrians is very difficult. Somehow, humans seem to adapt, as I did not see any accidents during my stay there. However, an AI trained on US traffic data would probably fail in Naples. Other places with "interesting" driving behaviour I have visited include Kampala, Uganda, and Samara, Russia. Once AI can handle those situations safely, we can truly speak of AI. I am not saying it is impossible, just that it is a HUGE challenge.
Similarly on motorways. Where I live in the UK you tend to allow faster approaching vehicles in the out side lane to over take before pulling out. When travelling through Belgium the process seems to pull out regardless and the faster approaching vehicles slow down behind you. Driving 'UK' style in Belgium will result in being permanently boxed in, and driving 'Belgian' style in the part of the UK I live in could result in accidents if the approaching driver does not expect the other driver to pull out at the last minute (i.e reading the road). Even across the UK there are major different driving styles such as the M25 versus the M6 in Cumbria. Are self driving cars going to have different rules programmed in for different locations?
Agree about Kampala Uganda.
Was there a few years back with a charity and using the same local taxi driver for repeat visits. On the first visit it was found that his eyesight was way below legal for UK driving. Spectacles were sufficient to bring him up to legal. On the repeat visit it was amazing how he usually won the scrums with other taxi drivers by driving closer to them than they would to him. I wondered if the other drivers had uncorrected vision and so were less confident of the proximity of their vehicle to others.
Fortunately they were only low-speed scrums in Kampala.
One key point that neither of these papers addresses is that the human visual system is not a camera. The actual image on the retina is a very small part of visual perception, which primarily depends on interpretation (some elements of which are still not fully understood). The weakest link in all current automotive 'perception' systems seems to be a lack of that capacity to interpret and infer from that interpretations possible outcomes -- i.e. to think.
> One key point that neither of these papers addresses is that the human visual system is not a camera.
At the risk of stating the obvious, that's because this research didn't set out to recreate the human visual system. This research has solved one specific problem: achieving low-latency object detection at reduced bandwidth. *Other researchers* can now build on this.
The other thing about human vision is the ability to place the identified objects in 3-dimensional space and compute their trajectories, primarily relative to oneself. I suppose we have to thank a long evolutionary history of hunting, being hunted and swinging through treetops for helping us to drive and, of course, play cricket.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2025
