By using multiple antenna elements, they can create an electronically steered phase array.
More simply: with two sets of fixed elements, they can have wide and narrow beams.
-
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
By using multiple antenna elements, they can create an electronically steered phase array.
More simply: with two sets of fixed elements, they can have wide and narrow beams.
Still doesn't help with roads that dip, or curved off ramps that have barriers near lane of travel...
The vision system has great horizontal and vertical resolution, and could predict travel path...
Ok fine, I’ll write what I was going to write then.
It would be nice to have true sensor fusion between radar and vision, a neural net each combined by another net...
Oh, and drive like humans drive - by remembering the visual cues for each stretch of road so that you are more proficient driving a road after you’ve driven it 50 times.
Tesla is so far away from that, it is depressing.
I want a system that drives a route perfectly the second time (out of all the cars)
Unless we are talking a race course... Then it can incrementally improve.
My feel is the system needs the NN for object detection, objects then get weights for avoidance, apply rules of the road, then plot a minimum cost path though them. But I also think a lot can be done by applying a good prefilter to the vision data (a combination change detection/ motion estimation and perspective warp), to reveals what is flat, what is solid, and what is moving toward you.
We can't say where Tesla is (internally) since they can download a totally different software set at any time. Just like we went from a world without the Roadster 2020 to a world that has one (or three or whatever).
Seems to me that Professor jimmy is saying software is the road to the promise land. I'm pretty sure jimmy has solved full self driving but he wants us to feel like we played a part.
I could be mistaken in current utilization. However, beam steered radars are out there.
This article describes a Dephi radar that outputs angle to target: http://www.mdpi.com/1424-8220/18/3/837/pdf
Dephi's ESR 2.5 is Electronically Scanned Radar.
Delphi Electronically Scanning RADAR | RADAR | AutonomouStuff, LLC
Performance data reporting bearing resolution on <1 degree
http://www.araa.asn.au/acra/acra2015/papers/pap167.pdf
I've also seen 2-D images showing range (Y) and intensity (shading) over time (X), and range vs speed over time.
My personal experience is with synthetic aperture radars which use a long set of pulse data from a fixed beam moving radar post processed to create the image.
Yes, both the AP 2 and 2.5 radars use phased array techniques and beam forming. When the wavelength is 4 mm, a lot of radiating elements can be packed into an antenna that is a few inches across.
Good question. I've seen research papers talking about 16 elements. More elements gives a wider scan direction and better resolution. Another paper discusses 23 and 33 element spherical scanning systems.
There is a 24GHz system with 44.
Would love to find a technical data sheet for current versions.
It does seem to me that beyond better sensors, the real problem with moving towards self driving or even enhancing auto pilot is software. For example, I think it's likely essential for the system to build a model of what's going on around the car with some characteristics of all the different detected objects, such as speed, past movements, any signaling, etc. If a car is obscured by a hill or a truck or whatever, it doesn't cease to exist, it just isn't visible. If it takes an exit, it's gone from the model and so on. On the other hand, if a truck in an oncoming lane turns in front of the vehicle, just because it's shape and color changes, it didn't cease to exist. If the model can't figure out what happened to it, there should be an alarm.
Given a dynamic model like than, the car then can do something like a chess playing program and assess what's possible and likely to happen and drive accordingly.
While I think all this is quite possible, the software effort to build it is quite non trivial and I'm not convinced that Tesla has the team and, more importantly, the corporate culture to do it.
Given a dynamic model like than, the car then can do something like a chess playing program and assess what's possible and likely to happen and drive accordingly.
While I think all this is quite possible, the software effort to build it is quite non trivial and I'm not convinced that Tesla has the team and, more importantly, the corporate culture to do it.
Here's a tech sheet for the Continental ARS 404 radar, which is the one in AP 2.5 systems. There is a lot of good tech data here.
If you go to Figure 9, it appears that this radar uses an 11 x 11 element array.
Thanks for the spec sheet. It would be interesting to see a programming manual for it showing how to operate it. I am still shocked at how low speed the CAN bus is. 500 kbps? Yikes. There are automotive 1Gbps networks available now...
Awesome, thank you!
I may also be able to track down info from the FCC database now that I have a part number.
The 404 has 2 TX and 2 (near) +1 (far) RX antennas that use digital beam forming likely aligned that figure you mentioned. It may only be a 11x1 array with no vertical adjustment.
CAN is great for fault tolerance and determinism. Also, effective bandwidth depends on number of nodes. If radar has it's own bus, that is plenty fast for the 250 untracked or 100 tracked objects at the 60 mS cycle rate.
jimmy_d
Deep Learning Dork
The part currently used by Tesla is the ARS410 (or so it says on the case of the radar). Any source for data on that part?
Start about half-way down this page:
HW2.5 capabilities
Fundamentally, the Bosch radar in AP 2.0 is an FMCW radar, and the Continental AP 2.5 radar is a more sophisticated pulse-mdulated, frequency-chirped radar with more sophisticated signal processing. The Conti radar can provide a lot more data to AP than the Bosch unit (like vehicle ID by 1-D downrange imaging), though it isn’t obvious that Tesla is using the additional features.
Similar threads
