At the shareholder meeting Tuesday, Elon said that maybe a 4mm radar would work better than lidar for autonomous driving. Bosch already makes an automotive radar in the 4mm wavelength (Mid-range radar sensor (MRR)). So, did Elon mis speak? Does he not realize it is already available? I’m confused...
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
4mm Radar
- Thread starter Cosmacelf
- Start date
That just makes me even more confused. It means that AP2 already uses a 4 mm radar. I just read a bad transcript of the exchange, and maybe Elon was just saying that 4mm radar is the way to go as opposed to LIDAR. If so, it would be nice if they figured out how to detect stationary objects in the cars’s path using it.
My car does a great job under 50mph at detecting untracked static cars at stop lights. Seems dicey but stops without panic stop like it did before 2018.10.4. definitely can be improved at higher speeds.
jimmy_d
Deep Learning Dork
Here's what Elon is saying:
If you're going to use an active photon emitter 4mm is a good wavelength. 4mm is small enough that diffraction limits don't impact your ability to delineate features of interest but it's immune to rain and fog. At 4mm building a phased array transceiver doesn't require exotic technology and you can generate and process the signals with existing ICs - no need for gallium arsenide or indium phosphide active optics (which all solid state laser systems rely on). It has excellent range, wonderful scattering properties and zero natural background noise. You don't need "lenses" or moving parts on any scale and the underlying tech is very low power, requires only a few square centimeters of antenna, integrates easily into existing vehicle bodies, isn't sensitive to temperature variations (unlike lasers), and doesn't care if your car is dirty (also unlike lasers). Plus transmissions can be digitally coded to eliminate interference from competing sensors on nearby vehicles.
It's a great wavelength - much better than 700nm, or 1500nm. And best of all it's complementary to cameras in a good way. Any self driving system *must* have excellent vision. No amount of ancillary sensors will compensate for crappy vision. And if you need good vision then your backup sensor should be good in exactly the places where vision is weak. It makes no sense to have your backup sensor operate at the same wavelengths (optical) where your primary sensor (vision) operates. You'd much rather have your backup sensor be reliable in the most common environmental situation where your primary sensor is having problems - heavy precipitation. When precipitation is bad enough to make vision fail it's also going to make lidar fail. But mm wavelength radar is nearly impervious to precipitation.
My addendum: it doesn't have to be 4mm - anything from 1mm up to 10mm is pretty workable, but 4mm happens to be a good point today when you consider the FCC and the current state of IC tech. Getting into the THz range you run into absorption issues that limit range and the IC tech isn't there yet - anything bigger than 10mm and you don't get good resolution on small objects like cats and tree branches.
If you're going to use an active photon emitter 4mm is a good wavelength. 4mm is small enough that diffraction limits don't impact your ability to delineate features of interest but it's immune to rain and fog. At 4mm building a phased array transceiver doesn't require exotic technology and you can generate and process the signals with existing ICs - no need for gallium arsenide or indium phosphide active optics (which all solid state laser systems rely on). It has excellent range, wonderful scattering properties and zero natural background noise. You don't need "lenses" or moving parts on any scale and the underlying tech is very low power, requires only a few square centimeters of antenna, integrates easily into existing vehicle bodies, isn't sensitive to temperature variations (unlike lasers), and doesn't care if your car is dirty (also unlike lasers). Plus transmissions can be digitally coded to eliminate interference from competing sensors on nearby vehicles.
It's a great wavelength - much better than 700nm, or 1500nm. And best of all it's complementary to cameras in a good way. Any self driving system *must* have excellent vision. No amount of ancillary sensors will compensate for crappy vision. And if you need good vision then your backup sensor should be good in exactly the places where vision is weak. It makes no sense to have your backup sensor operate at the same wavelengths (optical) where your primary sensor (vision) operates. You'd much rather have your backup sensor be reliable in the most common environmental situation where your primary sensor is having problems - heavy precipitation. When precipitation is bad enough to make vision fail it's also going to make lidar fail. But mm wavelength radar is nearly impervious to precipitation.
My addendum: it doesn't have to be 4mm - anything from 1mm up to 10mm is pretty workable, but 4mm happens to be a good point today when you consider the FCC and the current state of IC tech. Getting into the THz range you run into absorption issues that limit range and the IC tech isn't there yet - anything bigger than 10mm and you don't get good resolution on small objects like cats and tree branches.
NeverFollow
Active Member
This type of radar is used by various car for "smart cruse control" to keep a safe and constant with a car that you are following.
One issue is that in a curving road, when the car in front of you is not visible or detected, the cruse control starts to accelerate.
/mittelbereichsradadrsensor_functioning.jpg)
There is no such limitation, that's a Tesla forum legend. If you don't believe it, google FMCW automotive radar. They way they work is really clever.
It seems to me that the real limitation on what I believe is the current Tesla system is vertical resolution. These systems use electronic scanning but the resolution depends on the number of antennas in the X and Y planes and as I understand it, there are many fewer vertically. To get a real map of what's in front of the unit, the scan needs to be full resolution in both X and Y.
The main advantage LIDAR has is very high resolution in both planes.
jimmy_d
Deep Learning Dork
As others have mentioned, the idea that there's an absolute limitation on detection of stationary objects is a misunderstanding. Almost all automotive radars use doppler frequency discrimination to determine the relative velocity of targets that bounce back the transmitted signal. This has a fringe benefit, which is that if you can discriminate frequencies finely enough to measure relative velocity you can use the same discrimination filter to separate out the signals from things moving at different speeds, which gives you better signal-to-noise ratio for each individual return target. That improves your accuracy a lot. However this trick doesn't work for things which are moving at the same speed as the overall environment. An object which is stationary with respect to the surrounding trees, road, buildings, and parked vehicles can't be separated out, in a doppler sense, from background reflections. That means that discrimination is relatively poorer for stationary objects than for moving objects. How much this matters depends on what techniques other than frequency discrimination are being employed. Since you get frequency discrimination for free and since adding other discriminators increases the cost of the radar unit the very cheapest and simplest radars are going to have relatively more problem with stationary objects than more sophisticated radars will.
As the radar tech employed in cars gets better through improvements in IC technology and the cost improvements from mass production the stationary object disadvantage is getting ever smaller. The Conti radar spec calls out stationary object performance as one of it's advantages and even the original Bosch could detect stationary objects, though not at the same range that it could detect the same object when in motion.
As for vertical versus horizontal resolution of radar: one of the basic elements of the design of this generation of radars is how many antenna elements are employed in the radar's antenna grid. To a first approximation the cost of the electronics scales with the number of elements, so inexpensive radars try to keep the number of elements down and to optimize how they use those elements to match the application. Generally antenna elements for automotive radar are laid out in a rectangular grid that is wider than it is tall so they spend more resources on having fine horizontal resolution than they do on having fine vertical resolution. That's generally a good plan for cars, but when you have just a half dozen elements to work with it can mean that your vertical resolution get's reduced to just "low" and "high", which was kind of the case with the original Bosch unit, but seems to be less so with the Conti unit.
It's worth noting that car makers can add additional processing to the output that they get from the radar, in some situations, and effectively improve the radar performance with better software. Tesla seems to do that to some extent.
jimmy_d
Deep Learning Dork
I hear that $100 is the upper bound for stuff that goes into volume production.
deonb
Active Member
How does it work when you have 150 cars all beaming their radar at an overhead sign? When the signals come back, how do they differentiate which signals belong to which cars?
Signals are signed (coded) so that each car knows which is its own signal. The same way multiple people can use cell phones simultaneously in the same area.
deonb
Active Member
So you can use the Radar signal as a carrier wave? Frequency or Amplitude modulated?
All the automotive radars I'm aware of use FMCW: Frequency Modulation Continuous Wave (the transmitter never shuts off but the frequency varies) to send out a series of "chirps", that is, a signal the rises in frequency then drops suddenly back down. The shape when plotting frequency vs time is a sawtooth and there are various shapes.
The important point is that the peak of the shape can be recognized in the returning echo which gives two pieces of information. The first is the time of flight of the pulse from the delay between sending out the peak of the sawtooth and getting it back. That also allows the receiver to determine the frequency shift of the signal which gives the relative velocity of the sender and target because it knows the frequency at the peak during transmission and it can measure the reflected frequency. It behaves like both a pulsed and doppler radar system. (Note: the way this is actually done is not trivial)
Scanning the beam(s) allows the system to come up with a 3D plus velocity profile of what's in front of the radar. The reason I think vertical resolution is important is to notice things like 1m high crash barriers in front of the car. With low vertical resolution, distinguishing between the crash barrier and the road itself isn't possible until the car is very close indeed to the barrier.
My understanding is that in quantity radar units are a few hundred dollars. However as was pointed out earlier, more vertical resolution means more hardware and more $.
Principle of FMCW radar
The important point is that the peak of the shape can be recognized in the returning echo which gives two pieces of information. The first is the time of flight of the pulse from the delay between sending out the peak of the sawtooth and getting it back. That also allows the receiver to determine the frequency shift of the signal which gives the relative velocity of the sender and target because it knows the frequency at the peak during transmission and it can measure the reflected frequency. It behaves like both a pulsed and doppler radar system. (Note: the way this is actually done is not trivial)
Scanning the beam(s) allows the system to come up with a 3D plus velocity profile of what's in front of the radar. The reason I think vertical resolution is important is to notice things like 1m high crash barriers in front of the car. With low vertical resolution, distinguishing between the crash barrier and the road itself isn't possible until the car is very close indeed to the barrier.
My understanding is that in quantity radar units are a few hundred dollars. However as was pointed out earlier, more vertical resolution means more hardware and more $.
Principle of FMCW radar
Similar threads
