David Lindell: A camera that can see around corners

Recorded atNovember 23, 2019
Duration (min:sec)07:21
Video TypeTEDx Talk
Words per minute187.99 fast
Readability (FK)47.71 difficult
SpeakerDavid Lindell

Official TED page for this talk


To work safely, self-driving cars must avoid obstacles -- including those just out of sight. And for this to happen, we need technology that sees better than humans can, says electrical engineer David Lindell. Buckle up for a quick, groundbreaking tech demo as Lindell explains the significant and versatile potential of a high-speed camera that can detect objects hidden around corners.

Text Highlight (experimental)
100:12 In the future, self-driving cars will be safer and more reliable than humans.
200:18 But for this to happen, we need technologies that allow cars to respond faster than humans, we need algorithms that can drive better than humans and we need cameras that can see more than humans can see.
300:32 For example, imagine a self-driving car is about to make a blind turn, and there's an oncoming car or perhaps there's a child about to run into the street.
400:41 Fortunately, our future car will have this superpower, a camera that can see around corners to detect these potential hazards.
500:49 For the past few years as a PhD student in the Stanford Computational Imaging Lab, I've been working on a camera that can do just this -- a camera that can image objects hidden around corners or blocked from direct line of sight.
601:03 So let me give you an example of what our camera can see.
701:06 This is an outdoor experiment we conducted where our camera system is scanning the side of this building with a laser, and the scene that we want to capture is hidden around the corner behind this curtain.
801:18 So our camera system can't actually see it directly.
901:21 And yet, somehow, our camera can still capture the 3D geometry of this scene.
1001:27 So how do we do this?
1101:29 The magic happens here in this camera system.
1201:32 You can think of this as a type of high-speed camera.
1301:35 Not one that operates at 1,000 frames per second, or even a million frames per second, but a trillion frames per second.
1401:45 So fast that it can actually capture the movement of light itself.
1501:50 And to give you an example of just how fast light travels, let's compare it to the speed of a fast-running comic book superhero who can move at up to three times the speed of sound.
1602:02 It takes a pulse of light about 3.3 billionths of a second, or 3.3 nanoseconds, to travel the distance of a meter.
1702:10 Well, in that same time, our superhero has moved less than the width of a human hair.
1802:16 That's pretty fast.
1902:18 But actually, we need to image much faster if we want to capture light moving at subcentimeter scales.
2002:24 So our camera system can capture photons at time frames of just 50 trillionths of a second, or 50 picoseconds.
2102:33 So we take this ultra-high-speed camera and we pair it with a laser that sends out short pulses of light.
2202:40 Each pulse travels to this visible wall and some light scatters back to our camera, but we also use the wall to scatter light around the corner to the hidden object and back.
2302:51 We repeat this measurement many times to capture the arrival times of many photons from different locations on the wall.
2402:58 And after we capture these measurements, we can create a trillion-frame-per-second video of the wall.
2503:04 While this wall may look ordinary to our own eyes, at a trillion frames per second, we can see something truly incredible.
2603:12 We can actually see waves of light scattered back from the hidden scene and splashing against the wall.
2703:19 And each of these waves carries information about the hidden object that sent it.
2803:24 So we can take these measurements and pass them into a reconstruction algorithm to then recover the 3D geometry of this hidden scene.
2903:33 Now I want to show you one more example of an indoor scene that we captured, this time with a variety of different hidden objects.
3003:40 And these objects have different appearances, so they reflect light differently.
3103:44 For example, this glossy dragon statue reflects light differently than the mirror disco ball or the white discus thrower statue.
3203:52 And we can actually see the differences in the reflected light by visualizing it as this 3D volume, where we've just taken the video frames and stacked them together.
3304:02 And time here is represented as the depth dimension of this cube.
3404:07 These bright dots that you see are reflections of light from each of the mirrored facets of the disco ball, scattering against the wall over time.
3504:16 The bright streaks of light that you see arriving soonest in time are from the glossy dragon statue that's closest to the wall, and the other streaks of light come from reflections of light from the bookcase and from the statue.
3604:29 Now, we can also visualize these measurements frame by frame, as a video, to directly see the scattered light.
3704:37 And again, here we see, first, reflections of light from the dragon, closest to the wall, followed by bright dots from the disco ball and other reflections from the bookcase.
3804:48 And finally, we see the reflected waves of light from the statue.
3904:53 These waves of light illuminating the wall are like fireworks that last for just trillionths of a second.
4005:05 And even though these objects reflect light differently, we can still reconstruct their shapes.
4105:11 And this is what you can see from around the corner.
4205:15 Now, I want to show you one more example that's slightly different.
4305:19 In this video, you see me dressed in this reflective suit and our camera system is scanning the wall at a rate of four times every second.
4405:27 The suit is reflective, so we can actually capture enough photons that we can see where I am and what I'm doing, without the camera actually directly imaging me.
4505:37 By capturing photons that scatter from the wall to my tracksuit, back to the wall and back to the camera, we can capture this indirect video in real time.
4605:48 And we think that this type of practical non-line-of-sight imaging could be useful for applications including for self-driving cars, but also for biomedical imaging, where we need to see into the tiny structures of the body.
4706:01 And perhaps we could also put similar camera systems on the robots that we send to explore other planets.
4806:08 Now you may have heard about seeing around corners before, but what I showed you today would have been impossible just two years ago.
4906:15 For example, we can now image large, room-sized hidden scenes outdoors and at real-time rates, and we've made significant advancements towards making this a practical technology that you could actually see on a car someday.
5006:28 But of course, there's still challenges remaining.
5106:30 For example, can we image hidden scenes at long distances where we're collecting very, very few photons, with lasers that are low-power and that are eye-safe.
5206:41 Or can we create images from photons that have scattered around many more times than just a single bounce around the corner?
5306:48 Can we take our prototype system that's, well, currently large and bulky, and miniaturize it into something that could be useful for biomedical imaging or perhaps a sort of improved home-security system, or can we take this new imaging modality and use it for other applications?
5407:05 I think it's an exciting new technology and there could be other things that we haven't thought of yet to use it for.
5507:11 And so, well, a future with self-driving cars may seem distant to us now -- we're already developing the technologies that could make cars safer and more intelligent.
5607:21 And with the rapid pace of scientific discovery and innovation, you never know what new and exciting capabilities could be just around the corner.
5707:30 (Applause)