SoylentNews is people
SoylentNews
According to a [PDF] paper to be presented at the 2019 Conference on Computer Vision and Pattern Recognition, June 15-21 in Long beach, California, researchers have discovered a "simple, cost-effective, and accurate new method" of enabling self driving cars to recognize 3d objects in their path.
Currently bulky expensive lasers, scanners, and specialized GPS receivers are used in LIDAR (Light Detection And Ranging) sensors and mounted on top of the vehicle. This causes increased drag as well as being unsightly and adding another ~$10,000 to the price tag. Until now, this has been the only viable option.
Cornell researchers have discovered that a simpler method, using two inexpensive cameras on either side of the windshield, can detect objects with nearly LiDAR's accuracy and at a fraction of the cost. The researchers found that analyzing the captured images from a bird's eye view rather than the more traditional frontal view more than tripled their accuracy, making stereo cameras a viable and low-cost alternative to LiDAR.
According to the paper, which goes into this in considerable depth, it is not the quality of images and data which causes the difference in accuracy, but the representation of the data. Adjusting this brings the object detection results using far less expensive camera data for 3D image-analysis up to nearly the same effectiveness as much more expensive LiDAR.
Kilian Weinberger, associate professor of computer science and senior author of the paper, notes that
stereo cameras could potentially be used as the primary way of identifying objects in lower-cost cars, or as a backup method in higher-end cars that are also equipped with LiDAR.
The paper concludes that future work may improve image-based 3D object detection using the denser data feed from cameras further, fully closing the gap with LiDAR.
(Score: 4, Insightful) by ledow on Wednesday April 24 2019, @10:30AM (5 children)
Worse, this could be dazzled by a strong light at night.
Relying on optics the same way humans do is a really bad idea, and humans show you why - 65m years of evolution and we still can't judge distances or speeds accurately. That's why we invented cheap, easily-available tools that *can*.
Sure, we might be able to spear a gazelle from a standing start if we practiced for years, but in terms of judging anything in motion, we're useless. Which is why we have to have speedometers, braking-distance warnings, mnemonics, very powerful brakes, early-warning systems and sensors, and our complete attention to do so anywhere near reliably.
This is very much the old story of trying to anthropomorphise a computer-based approach to a problem. Rather than use computers, electronics and other hardware to do what they are actually good at, we throw away all their advantages to try to make something that replicates how we do it ourselves, which we do quite badly in the grand scheme of things anyway.
The solution is so damn obvious but nobody will say it... get these things away from human drivers and assign a lane / road to be automated driving only. Then you can mark fecking barcodes along the road to give away information like position, what the road ahead is shaped like, where the lane boundaries are, etc. and have the cars talk to each other and solve all these problems immediately, for very low cost.
(Score: 2, Interesting) by doke on Wednesday April 24 2019, @12:22PM (1 child)
Humans have trouble judging distance for many reasons that would have less effect on stereo cameras. We move our heads around a great deal. We move our eyes around. We even do micro-movements of our eyes, to shift the image edges around across cones. Cameras would be fixed, relative to the car, so an edge position can be easier compared between them. Our mental image processing is distracted by many other tests, ie is that object a gazelle that I can hunt and eat, or a bear that will hunt and eat me? The car would just want to know if there's something there to avoid. Red light cameras already do an excellent job of measuring speed, deceleration, distance, and location.
Assigning a dedicated lane for autonomous vehicles would only be viable in a few places. The main multilane roads around where I live (Northern Delaware USA) are highways, and they're overcrowded. Most of the roads are one lane each way, and there are quite a few single lane bridges.
(Score: 2) by Runaway1956 on Wednesday April 24 2019, @01:42PM
Single lane bridges? What part of the world do you live in? It's been a long time since I've used on on a paved road. I can still find some low water crossings, on gravel roads, and I know Pennsylvania still has some old wooden covered bridges that are one lane. It's been a long time since I've even crossed a "singing bridge", all steel with a steel grid road surface, and often times not wide enough for two trucks meet in the middle.
(Score: 3, Insightful) by c0lo on Wednesday April 24 2019, @12:38PM
Cheap tools like...?
You know binary vision is not human specific.
Say, how about we don't say we try to imitate our vision, but the kestrel's or sparrowhawk's vision (you know? those predatory birds hunting with great precision and at high speed) - they have two eyes as well. Very much like the computer stereo vision [wikipedia.org]
O'realy? You mean we are so stupid we can't imagine how to improve the things we are making?
Picking an example of using "two separated eyes", we never made a coincidence rangefinder [wikipedia.org] to help us throw a 10cm-caliber shell at 15km distance [wikipedia.org], those didn't happen, right?
https://www.youtube.com/watch?v=aoFiw2jMy-0 https://soylentnews.org/~MichaelDavidCrawford
(Score: 1, Interesting) by Anonymous Coward on Wednesday April 24 2019, @01:32PM
I'm willing to bet that the self-driving car will also have access to a speedometer.
Whatever sensors the braking-distance warning uses, will likely also be included in the self-driving car. The rest is already software anyway (and an optical/acoustic signalling device, which the self-driving car won't need).
It won't need mnemonics, as doing calculations and reliably storing and retrieving data is exactly what computers excel at.
I also don't see why self-driving cars would have less powerful brakes.
The self-driving cars also will likely have the same sensors as a normal car. The early-warning system? Well, apart from those sensors that already is just computer code. There's no reason to assume that the computer code in the self-driving car won't be able to do what the computer code of those much simpler systems already are able to do.
And guess what human-driven cars do not have? LIDAR.
No. It's another story, much older than electronic computers: Trying to bring down cost. Make a LIDAR system that's cheaper than two cameras, and they will happily throw out those cameras for LIDAR.
(Score: 1, Insightful) by Anonymous Coward on Wednesday April 24 2019, @04:59PM
It shouldn't be one or the other; a combination of techniques may be the best approach to both reinforce each other and to provide a backup if one fails.
And, too much LIDAR may have other problems, such as damaging the human eye and/or cameras, and confusing other LIDAR vehicles. Under "ideal" conditions they are safe, but since when is anything ideal? If most the cars on the street have lots of LIDAR, the side-effects will become more pronounced. Light LIDAR and optical recognition in combination may be the best approach.