This is not a "true" LiDAR in the classical sense (where you send out a laser beam and measure time passed until return). It's rather an indirect ToF sensor that uses modulated light bursts that flood the entire scene.
This approach typically works well for close-range convex surfaces (hand tracking, bin picking, face ID) but fails pretty miserably when longer ranges and concave surfaces are involved, due to quadratic signal dropoff and multipath errors.
As far as I understand, what the team has achieved is lowering the power requirements for the modulation part. It means they can spend the saved power on making the modulated light brighter, which should give them a bit more range. I haven't seen any other major improvements though and none of the other issues with iToF were addressed.
Not trying to downplay the achievement, just saying it is still affected by the usual tradeoffs and probably just occupies another niche in the high dimensional space of 3D cameras, rather than spanning many of today's niches.
This approach typically works well for close-range convex surfaces (hand tracking, bin picking, face ID) but fails pretty miserably when longer ranges and concave surfaces are involved, due to quadratic signal dropoff and multipath errors.
Exactly. This has been used before for short-range sensors, modulating the outgoing light electrically. Microsoft used it in the second generation Kinect. Mesa Imaging used it in 2013.[1] The prototype of that was shown in 2003.[2] I looked into this in my DARPA Grand Challenge days.
Since it's a continuous emission system, you need enough continuous light to overpower other light sources. Filters can narrow the bandwidth, so you only have to be brighter at a specific color. This works badly outdoors. Pulsed LIDARs outshine the sun at a specific color for a nanosecond, and thus can be used in bright sunlight. Also, they tend not to interfere with each other, because they're receiving for maybe 1000ns every 10ms, or 0.01% of the time. A little random timing jitter on transmit can prevent repeated interference from a similar unit.
So, short range use only. For more range, you have to use short pulses.
For short range, there's another cheap approach - project a pattern of random dots and triangulate. That was used in the first generation Kinect and is used in Apple's TrueDepth phone camera.
This is not directly my field, but in general I've found that many, many articles (and not just free-daily-rags, but also from reputable magazines and news sites) always seem to inflate every scientific achievement, where some minor discovery (still worth a published article) is presented as "new cure for cancer found!", and even worse if some well-known substance is used (eg. caffeine makes medicine delivery faster), the title would say "coffee cures cancer!".
And not just medicine... just look at the number of "new, 10.000x better battery discovered" articles posted here.
Nope. Structured light is projecting a geometric pattern into the scene and viewing it from from a camera that's offset by some amount vs. the projector. The original Kinect works this way.
Structured light is modulated in space rather than time. Accordingly, distance is inferred from offset in space (pixel, ie. angle) rather than time (phase).
I would have assumed they'd choose to work in a nice narrow absorption band like the 5.6u CO2 where solar illuminance is almost blocked. I know they commonly do that with structured light. It does add cost for filters and I don't know how well these ToF or standard CMOS imagers work there.
What happens in busy places when tens or hundreds of vehicles or robots illuminate the environment? How do several independent time-of-flight camera systems coexist? Are there existing solutions from other areas that are already using time-of-flight camera systems?
Cars (of 2035) ought to have multiple sensors and should be able to discard discrepant input - if (stereo) visual, radar, sonar and lidar disagree, the car should switch to a "degraded safe" self driving mode and signal it to cars around it via multiple channels (visual, radio, ultrasound, etc).
Cars of today already have this problem. If one sensor says there's a solid object right in front of you that just popped out, and another sensor says there probably isn't but is confused, what do you do? Slam on the brakes?
Teslas today have this "degraded mode" and the reaction is to sound an alarm and ask the driver to pay attention and grip the wheel, while slowing. This seems like a reasonable thing to do. Cars of 2035 that have no wheel better have perfect sensor integration and false-data elimination.
If you have two sensors, you have one. You can't resolve a disagreement of two sensors. Stopping the car is the only option in this case, but, say, if you have four sensors and one says there is 100% certainty of a solid object ahead of you, one says it's 50/50 and two others say it's a 100% no, then a mild braking action may be all that's needed, to give the car and its sensors more time to react and assess. Worst case scenario you still hit the solid object at an easily survivable speed. You also signal cars behind you to actuate their brakes accordingly (and, in turn, inform traffic behind them of their actions).
And it gets way more interesting when every care itself is a cluster, or node of sensors, and all of these sensors of the same type can form telemetry fabrics of all the families of sensor-types across the nodes.
Each a plane, and then ultimately being able to see the relationships of patterns of sensor groups. I wonder how that information will become useful?
First use could be to model traffic patterns - as undoubtedly all online mapping apps already do. As the meshes get denser, the cars can become aware of traffic patterns around them and make decisions in a coordinated way, forming "trains" to reduce drag in highways, for instance. Eventually enough data is gathered that we have predictive models for traffic jams and accidents and can act to prevent them.
This seems fanciful - "if it'd work so well, why don't we already do it" is a real point against it. I think people imagining this are forgetting problems with communications overhead, trusting each other, and how much it'd actually be useful.
What we do have are realtime traffic alerts and Waze, and presumably a robot car could use those as well as we do.
True, but the usual action is to hand the decision to the human, who can look out, figure out which (if any) sensor is right, and take appropriate action (737 Max feelings).
Humans are better at using our big two sensors because we can interrogate the world with them - i.e. move our head or refocus our eyes on what we want to look at, or just put the sun visor down if there's too much glare.
I assume the people working on cars also know they could build this, but most people don't seem to notice it's a failing of passive sensors like most camera setups.
You can't and won't get perfect sensing. It's simply not possible.
A safe vehicle design avoids the situation of having to decide whether to slam on the brakes constantly by realizing that's a fundamentally unsafe mode to be operating in. You should already be slowing/able to safely brake/pulling over/stopped before that point excepting rare "everything explodes simultaneously" situations you can try to engineer away with safety ratings and MTBF numbers.
And, since you aren't driving a fully automated vehicle, it driving below speed limits becomes much less annoying (and, if all cars can coordinate their speeds, prevent traffic jams altogether).
Sonar for bats and dolphins etcetera. I think illumination from different sources can help resolve the location of objects (using reflections on objects from other sources, or tracking the source itself such as a car with LIDAR).
While this is cool, I think we should seriously take a look at how insects navigate 3d spaces. They typically have compound eyes, where each eye-let incredibly simplified (in fact, early-stage evolution) photoreceptor that has independent connections in the brain. They have crappy resolution but incredibly good response time and spatial awareness. And they are super low power.
But cars don't navigate 3d spaces, it seems like that would make results not so transferable. How does a fly avoid in-air collision with another fly? Answer: it doesn't need to.
> As lidars slim down and scale up, the days of pure two-dimensional image sensing seem numbered. You probably saw your first lidar system on a so-called self-driving car. You might see your next one on your phone.
The futuristic crime dramas and games where they replay events in a simulated 3D space is getting closer. Imagine public spaces constantly scanned and recorded right down to the dropped cigarette.
It's not real time, but Apple's map scans are much more 3D. Like Google Maps you can't make small movement adjustments in between designated viewpoints, but the animation of moving from one point to another is pretty slick.
Now, imagine that, using car sensors, you could time travel and see the same point in space, but at any point in time (after the introduction of cars that contribute their sensor data to a consolidated repository of "4D" data.
I was thinking more along the lines of being able to perfectly map your environment so that it could be incorporated into whatever application you're using. That along with more comfortable and higher resolution headsets will mean you can spend more time with the headset on. This is important for productivity or social VR/AR since you don't want those sessions to be limited by discomfort or minor interruptions like getting a glass of water. It would also unlock the ability to change spaces or positions seamlessly e.g. from chair to standing to couch to floor etc. This is one of my favourite things about VR vs pancake computing, you aren't locked to your desk or forced into bad posture by hunching over a laptop. You can simply move naturally around your space when the mood takes you. I think we're going to look back on the 20th and early 21st as eras of tortuous treatment of the human body where we became sedentary and crippled by our devices.
This approach typically works well for close-range convex surfaces (hand tracking, bin picking, face ID) but fails pretty miserably when longer ranges and concave surfaces are involved, due to quadratic signal dropoff and multipath errors.
As far as I understand, what the team has achieved is lowering the power requirements for the modulation part. It means they can spend the saved power on making the modulated light brighter, which should give them a bit more range. I haven't seen any other major improvements though and none of the other issues with iToF were addressed.
Not trying to downplay the achievement, just saying it is still affected by the usual tradeoffs and probably just occupies another niche in the high dimensional space of 3D cameras, rather than spanning many of today's niches.