87
\$\begingroup\$

Self driving cars rely on cameras, radar, and lidar to recognize the environment around them. Cameras of course don't interfere with each other, since they are passive sensors. Since a signal received directly from another transmitter is much stronger than a reflected signal from your own transmitter, what stops the transmitted signals from one radar/lidar interfering with the receiver of another?

Will radar/lidar still work when all cars are equipped with them? Assuming that they will, how will this be accomplished?

\$\endgroup\$
16
  • 2
    \$\begingroup\$ It's not necessarily a problem if you're picking up another radar emitter. That's a pretty strong signal that there's another car, and it is also "active" - even if it's currently not moving, its use of radar is a strong hint that it could start moving. \$\endgroup\$
    – MSalters
    Commented Jun 14, 2018 at 14:01
  • 1
    \$\begingroup\$ @MSalters - picking up another independent emitter doesn't give you range however, nor without good standards-referenced oscillators even doppler. There are many places where you can have traffic on a head on path towards you, which will not collide as the streets curve in between. \$\endgroup\$
    – Chris Stratton
    Commented Jun 14, 2018 at 15:17
  • 3
    \$\begingroup\$ What I personally hope is that pseudo-bluetooth car-to-car comms will come into being before we completely obliterate the radar bandwidths. \$\endgroup\$
    – Carl Witthoft
    Commented Jun 14, 2018 at 18:00
  • 4
    \$\begingroup\$ @CarlWitthoft If by IR cameras you mean thermal imaging cameras, then yes. And I know how cameras work, but if a camera system needs illumination to function, then it's not a passive system. \$\endgroup\$
    – gre_gor
    Commented Jun 14, 2018 at 21:14
  • 3
    \$\begingroup\$ This is a good question, the number of active Radar/Lidar systems at the intersection of congested highways with 5 lanes in each direction seems likely to present much greater problems than household WiFi which, even in dense urban areas, are static. \$\endgroup\$
    – RedGrittyBrick
    Commented Jun 16, 2018 at 9:38

6 Answers 6

Reset to default
82
\$\begingroup\$

You'd be surprised.

This is actually topic of ongoing research, and of several PhD dissertations.

The question which radar waveforms and algorithms can be used to mitigate interference is a long-fought over one; in essence, however, this breaks down to the same problem that any ad-hoc communication system has.

Different systems solve that differently; you can do coded radars, where you basically do the same as in CDMA systems and divide your spectrum by giving each car a collision-free code sequence. The trick is coordinating these codes, but an observation phase and collision detection might be sufficient here.

More likely to succeed is collision detection and avoidance in time: simply observe the spectrum for radar bursts of your neighbors, and (assuming some regularity), extrapolate when they won't be transmitting. Use that time.

Notice that wifi solves this problem inherently, much like described above, in a temporal fashion. In fact, you can double-use your Wifi packets as radar signals and do a radar estimation on their reflection. And since automotive radar (802.11p) is a thing, and the data you'd send is known to you and also unique, you could benefit from the orthogonal correlation properties of a coded radar and the higher spectral density and thus increased estimate quality of time-exclusive transmission.

There's a dissertation which IMHO aged well on that, and it's Martin Braun: OFDM Radar Algorithms in Mobile Communication Networks, 2014.

\$\endgroup\$
11
  • 3
    \$\begingroup\$ @DavidK frequency allocation: nah. You need to be adaptive. You can't give every potential car its own band, or even find a method for every car in a given radius. Simply because radar is inherently a bandwidth-intense problem, and spectrum is friggin sparse these days. \$\endgroup\$
    – Marcus Müller
    Commented Jun 14, 2018 at 14:05
  • 4
    \$\begingroup\$ @st2000 why would you design a radar to use any smaller duty cycle than possible? the higher the duty cycle, the either more observations you get to build your model or the better your single estimate becomes. This is basically saying "if we design a radar to be very suboptimal, we avoid the collision problem". (note that there's legislation that says your duty cycle has to be limited, but "a fraction of a percent" really isn't going to give you a low probability of collision in a crowded scenario... which incidentally is the scenario where you'd want a working radar.) \$\endgroup\$
    – Marcus Müller
    Commented Jun 14, 2018 at 14:06
  • 6
    \$\begingroup\$ note that the probability of collision is not just the duty cycle. stochastically, you run more into something that looks like the birthday paradox for continuous birthday ranges. \$\endgroup\$
    – Marcus Müller
    Commented Jun 14, 2018 at 14:10
  • 2
    \$\begingroup\$ Marcus, you cannot compare the CDMA or Wifi technology with the requirements of the auto industry. Are you really going to bet your life on the assumption that all other cars or devices will be kind and stay on their slot? A self driving car needs this data in a timely fashion unlike a communication line that can stand a one second interruption. \$\endgroup\$
    – Dorian
    Commented Jun 14, 2018 at 14:43
  • 3
    \$\begingroup\$ Good question @LastStar007, I just assumed that a car's LIDAR or RADAR would only be interested in the immediate surroundings. Say, 10 x the 2 second rule with a minimum of 10 yards. So that, roughly would mean a 100mph car would look 0.25 miles or ~1320 feet ahead of it. It takes light 0.00000268666 seconds to go that far and reflect back. So you open the window for at least that long. Even if you take a sample 100 times a second your window would only be open for 0.0268666 percent of the time. It would be rare to have collisions. \$\endgroup\$
    – st2000
    Commented Jun 15, 2018 at 15:19
8
\$\begingroup\$

There are military sonar and radar systems that see the world around them using reflections of "the other guy's" radar / sonar. They existed back in the days of 286 Intel processors... so it can be done much more cheaply today when a $5 ARM SOC is as powerful as a 1983 Cray XMP-48 (the machine I managed then...)

So while it is useful to use all the time domain and code domain multiplexing, it is also possible to compute the location of the other emitter and then use his signal to see the world around you.

I know this existed in the 1980's as I knew the engineer who built it for the military and visited his shop. It was secret then, now not so much.

Basically, call multiple emitters a "feature" and move on.

\$\endgroup\$
2
7
\$\begingroup\$

This is a rather old problem in radar engineering, dating from the era of jet aircraft carrying guns and supersonic missiles. This Wikipedia article on Chirp compression gives some clues about how the problem might be equally addressed at automobile speeds.

\$\endgroup\$
1
  • 8
    \$\begingroup\$ Welcome to EE.SE. While it is considered a good thing to link long lasting resources like Wikipedia, it is not so good to write no information in the answer itself. Please give a short abstract of what is relevant from the wiki article to ease the reading and make a search on EE.SE successful. \$\endgroup\$
    – Ariser
    Commented Jun 15, 2018 at 8:11
4
\$\begingroup\$

On a primitive level of explanation, radars work by sending a defined sequence of pulses (signature) and then waiting for a similar sequence to be received. This provides high selectivity in the presence of significant interference or noise.

By making radar signatures sufficiently long and unique it's possible to allow multiple radars to coexist in the same environment, where every radar can distinguish its own signature even if other signatures are simultaneously present in the signal.

\$\endgroup\$
3
  • 1
    \$\begingroup\$ Won't two simultaneous sets of pulses on the same frequency range look like one unrecognizable mashup? \$\endgroup\$
    – cHao
    Commented Jun 15, 2018 at 16:22
  • \$\begingroup\$ Lidar pulses must be very short, though, since the ones in self-driving scenarios send millions of pulses per second. I wonder how much data you can actually encode into such a short pulse to make a unique identifier while still providing enough power to detect the reflected signal. \$\endgroup\$
    – Adrian McCarthy
    Commented Jun 15, 2018 at 20:53
  • 1
    \$\begingroup\$ @cHao no. Cross-correlation works wonders. Well, proper code design does. \$\endgroup\$
    – hobbs
    Commented Jun 16, 2018 at 18:20
2
\$\begingroup\$

I don't think that they will rely to much in the radar technology, I think this is just a patch until the cameras and AI will be advanced enough to take over entirely (that if ever the technology will be safe enough to be adopted on a large scale).

Just imagine that you could silently blind all cars around you. The hacking resources will be unlimited. What can you do to stop some somebody to do that ?

Of course you can blind a camera or even the driver to but he (or the passenger) will know that and take action.

A short failure of a CDMA or Wifi network can pass unnoticed. A car cannot wait one second until it gets data again, the requirements are much much higher.

Update Actually I don't see a future of the self driving. Many "smart" devices are making our life less secure and less private day by day.

I think that in the end there will be some centralized driving , rather smart roads that will guide each car in the traffic.

Apart from that, I see many disregarding the power of the image processing even I don't think the cars will ever drive using only cameras.

In the first place a camera has a huge redundancy. I don't see how an insect eye can better than a human eye.

  • "The radar can read the speed"

The camera (one or more) can also read the speed of an object. In 3 directions.

  • "The radar can read the speed with a great accuracy"

I'm not sure about this point. I really can't tell the speed without looking at the speedometer (which is not accurate at all) and I'm still driving well.

  • "Cameras can't see in the fog"

It's true. Then drive slow through the fog. A pedestrian or a dog also don't see in the fog your smart car approaching.

It's hilarious how CDMA or Wifi is given as example of sharing the same bandwidth for a car radar. Are you using Arduino to run your ABS?

Pushing ahead the smart driving cars is more a marketing move, It will not make the streets safer in the near future and the driving skills of the people will get lower and lower, I don't even know how this will work, you won't need a driver license, the car will get you everywhere?

I also see in the Marcus Brown's document linked by Marcus Muller how the smart cars will nicely and honestly cooperate into avoiding obstacles and also sharing the radar bandwidth. What a wonderful future!

That means that a Russian hacker car could make my car jump outside the road at 100 km/h as a better option than hitting a wall that only he sees?

\$\endgroup\$
19
  • 2
    \$\begingroup\$ @Galaxy Exactly as you see at night. I jut pointed that it should be something that a human supervisor can easily detect. \$\endgroup\$
    – Dorian
    Commented Jun 14, 2018 at 16:26
  • 3
    \$\begingroup\$ I was just pointing out that the use cases & reliability of camera's are very sensitive to environmental conditions be it lack of light (night) or fog etc. The accuracy of a camera system to calculate distance is extremely dependent on those conditions hence I don't see radar being completely replaced by cameras whether or not an AI is attached to them. \$\endgroup\$
    – Galaxy
    Commented Jun 14, 2018 at 16:39
  • 3
    \$\begingroup\$ @Galaxy Humans don't have radars, they have eyes. Our road system is built for visible light. There's no way to use radar to read lane markings and traffic signs, for example. Therefore, we install headlights, taillights and streetlights. \$\endgroup\$
    – user71659
    Commented Jun 14, 2018 at 17:47
  • 2
    \$\begingroup\$ @Dorian also, you ignore the fact that you need relative speed information much much exactly than exact positioning. Velocity is very easy to get with radar – Doppler is the physical effect and it's easy to estimate. Your camera system could, at best, perform as well as much easier, more reliable, and most importantly: cheaper automotive microwave radar. You think to highly of camera systems, to be honest. "A short failure can go unnoticed" nonsense. If your board computer senses the radar has failed, it will react, just like you react to being blinded. Only with 1/1000 of the latency. \$\endgroup\$
    – Marcus Müller
    Commented Jun 14, 2018 at 17:55
  • 5
    \$\begingroup\$ Your answer really tries hard to justify your preference for visual systems, ignoring the physical limitations of light. Also, the multiple billions worth of automotive radars built each year and the statistics that say those save lives be the hundreds each year already, though mandatory radar for trucks is a thing that's still relatively new in the EU. Did I mention radar is mandatory for trucks in the EU? Also, your attacker scenario ignores that the machine learning approaches of the last ten years have led to algorithms that can easily be led astray with, to the human eye... \$\endgroup\$
    – Marcus Müller
    Commented Jun 15, 2018 at 6:05
1
\$\begingroup\$

I would think the opportunities for them to cooperate when in close proximity far outweighs the difficulties they cause for each other. A simple handshake as they recognize each other and then some information sharing as well as resource allocations... Surely any build out at any density will have to include some protocol exchange. They'll cooperate, not compete.

\$\endgroup\$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.