-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
AEB Won’t Prevent an Accident
- Thread starter Rusty1
- Start date
-
- Tags
- Driving Dynamics Model S
YauKwan
Member
Tam
Well-Known Member
Here Effectiveness of forward collision warning and autonomous emergency braking systems in reducing front-to-rear crash rates - ScienceDirect is a study showing impressive results for AEB, too. Forward collision warning helped, automatic emergency braking was better, and combined FCW & AEB was best. Here are the numbers:
Attachments
The reason is simple: radar clutter.
The car sees huge reflections coming back from stationary objects with a dopper shift equal to it's own velocity. It is very
difficult to be sure that the reflection is a car, and not the curvy-dish shaped end of a soda can reflecting a huge signal (for example).
But in normal operation the car is tracking OTHER CARS..they have a small velocity relative to your car, and so at that frequency
the car has no radar clutter to contend with.
Once you get close enough to a stationary car the radar CAN see it ...finally you have a honking huge reflected signal...and so the car can do stop-and-go following ok. Likewise the AEB can jam on the brakes at the last second with full confidence there really is a crash about to happen. (False alarm random AEB braking would be so BAD).
I have heard that AEB uses forward cameras trained on zillions of images of the rear ends of cars..... Not sure if true..but seems like
the future.
The radar is actually quite crude. I find this zero-velocity blindness deeply disturbing and think Tesla is irresponsible for not
telling EVERY DRIVER about it when they pick up the car. Yah obviously people should RTFM, etc. but they don't. Letting them
get hurt is still not right even if it is 'their fault'.....assuming the AP would stop is an extremely reasonable assumption..that just
happens to be WRONG>
There have been several crashes due to people thinking that the car will stop ..when it will NOT.
Lemme drop some science here. As a fighter pilot, we deal with radar and Doppler issues every day. Please read on if you’d like to know how this actually works.
It’s called a ‘Doppler notch’, and it’s physics. Simply, if you are driving at say 65 mph, the WORLD is rushing at the car at 65 mph. If Doppler was somehow ‘turned off’ the car would slam on the brakes immediately as it ‘sees’ the road, the tree, the light pole, that manhole cover, etc rushing towards it at 65 mph. To prevent this, engineers install a Doppler filter (same as in fighter jets) that basically ignores everything at the speed of the car. This is called a ‘notch’ in our parlance.
This is basic physics and radar theory. With the notch/filter in place, the car can now drive and essentially doesn’t see objects coming at it at the same speed as the car is traveling. Objects that a mph or two different than vehicle speed ARE seen due to Doppler effects, registered, and summarily avoided.
This is fact for all automakers with radar avoidance systems, as also explained in the article someone posted above.
Excellent summary of the technology.
It is for this very reason that "full" autonomy, level 4 and level 5, will require sensor redundancy. Radar is cheap and pretty good but obviously not perfect, especially in regard to the limitations described above or if the sensor is dirty etc. Camera tech is getting really really good at object recognition and can identify very useful things that radar can't, like stoplights and road signs. However, cameras also don't work in the dark so good, or fog, or heavy rain. FLIR can see in the dark 4-5 times farther than your headlights and accurately identify people, animals and other road hazards far beyond the capability of visual cameras. Similarly, LIDAR can see some objects more accurately than radar and doesn't depend on the doppler effect so it can give very accurate information about the speed and distance of objects from the car. You really have to put these technologies together with adequate processing power to get the levels of autonomy needed.
There is a vast amount of confusion regarding these technologies and they should be safe to use right out of the box. 90% of people buying cars don't know how to change a tire let alone how their AEB with FCW works. It's called autopilot and they assume they're good to go when they're not.
You are wrong. The physics for the radars you describe and those for automotive radars are the same, but the implementations for automotive use are not as you imagine. For proof, see the following illustration from here http://cdn.euroncap.com/media/1384/...13-0-a837f165-3a9c-4a94-aefb-1171fbf21213.pdf . These vehicles all responded to a stopped object ahead, including those with radar. Also, my own experience since the 1990s with automotive radar is that no automotive radar system ignores stationary objects in the expected path. To reduce false alarms, many of them wait after first detection for more confidence the object is not overhead or outside the lane, but even those react before collision. That's what defines today's "collision imminent braking" which very often uses radar. Also, here in the U.S., the Insurance Institute for Highway Safety tests vehicles with automatic emergency braking, including for braking for a stopped vehicle ahead, and many of those systems are radar-based -- and their tests show they work.
stopcrazypp
Well-Known Member
Yeah, the myth that automotive radar can't detect stationary objects has existed too long and is no longer applicable to modern radar.You are wrong. The physics for the radars you describe and those for automotive radars are the same, but the implementations for automotive use are not as you imagine. For proof, see the following illustration from here http://cdn.euroncap.com/media/1384/...13-0-a837f165-3a9c-4a94-aefb-1171fbf21213.pdf . These vehicles all responded to a stopped object ahead, including those with radar. Also, my own experience since the 1990s with automotive radar is that no automotive radar system ignores stationary objects in the expected path. To reduce false alarms, many of them wait after first detection for more confidence the object is not overhead or outside the lane, but even those react before collision. That's what defines today's "collision imminent braking" which very often uses radar. Also, here in the U.S., the Insurance Institute for Highway Safety tests vehicles with automatic emergency braking, including for braking for a stopped vehicle ahead, and many of those systems are radar-based -- and their tests show they work.
In fact they advertise exactly that they can detect stationary objects:
"Independent mode for height measurement using an elevation antenna, enabling the system to reliably classify obstacles and brake safely, even when the object is stationary"
Long-range radar sensor
The most recent 79GHz radar with higher bandwidth can even detect the shape of a stationary car, a pedestrian near a car, whether doors are open or not:
https://pdfs.semanticscholar.org/152d/3c733cccd528c26b6afd91a43738acbff688.pdf
Did you read the study? ALL of the cars tested were using LIDAR, with one car using Radar only, and one car using a fusion system of multiple sensors.
Radar came out with better performance than LIDAR.
ALL cars were UNABLE to stop and avoid a collision above 25 or 30 kph. With the exception of the fusion stereo system car. Read that again.
Here’s a summary from the actual study. “One RADAR-only system has been tested and was found to offer higher speed avoidance (up to 30km/h) than any of the LIDAR systems but this had no mitigation effect at higher speeds. Two multiple sensor systems were tested and both offered greater performance than either LIDAR or RADAR alone. The stereo camera system was most effective, with full avoidance from test speeds of up to 50km/h.”
You are wrong. The physics for the radars you describe and those for automotive radars are the same, but the implementations for automotive use are not as you imagine. For proof, see the following illustration from here http://cdn.euroncap.com/media/1384/...13-0-a837f165-3a9c-4a94-aefb-1171fbf21213.pdf . These vehicles all responded to a stopped object ahead, including those with radar. Also, my own experience since the 1990s with automotive radar is that no automotive radar system ignores stationary objects in the expected path. To reduce false alarms, many of them wait after first detection for more confidence the object is not overhead or outside the lane, but even those react before collision. That's what defines today's "collision imminent braking" which very often uses radar. Also, here in the U.S., the Insurance Institute for Highway Safety tests vehicles with automatic emergency braking, including for braking for a stopped vehicle ahead, and many of those systems are radar-based -- and their tests show they work.
Then why didn't it work in Florida or with the fire truck or for me about 2 weeks ago when I had to intervene?
Pure science, Doppler notch.
Even the study concluded, if you are above 25-30 kph...no system will detect and stop. The engineers HAVE to factor in Doppler notch and force the radar to not see objects coming at the car at the same speed the car is traveling...or we’d have thousands of false stops.
Even the study concluded, if you are above 25-30 kph...no system will detect and stop. The engineers HAVE to factor in Doppler notch and force the radar to not see objects coming at the car at the same speed the car is traveling...or we’d have thousands of false stops.
Heh, my first job as a software/firmware engineer was literally to fact-check suppliers in the context of the auto industry. I would read things with enticing marketing claims just like what you're doing here, order the part, and attempt to build a prototype demonstrating the ability in order to help this unnamed system integrator supplier select which parts to use.
I have to say, the success rate is pretty much the same as weight loss infomercial products. I would strongly caution against reading into specific marketing claims and confusing that with a productized system. The Bosch radar sensor used in AP1.0 and AP2.0 has the same claim, that there's a tall radar beam and wide radar beam for eliminating overpasses and hence enabling stationary radar braking. But judging by how much AP1 and AP2 slammed on the brakes back when radar-only-braking was more active, I'm guessing it didn't work out of the box super well.
I do agree that radar technology could theoretically deal with stationary objects through advanced signal processing, but I honestly still doubt anything on the market actually works like that.
All cars did not have lidar. As you said yourself, one car had radar only. You said that radar can't detect stationary objects. That car with radar only DID detect and respond to the stationary object. That was my point.
Here in the U.S., the Insurance Institute for Highway Safety ("IIHS") tests vehicles with automatic emergency braking ("AEB") systems and publishes results. Many of the vehicles they test use radar for that function, and one if their tests is for the case of a stopped vehicle. Here is their description for that test: Front crash prevention tests . Using that test, from this review 2017 Honda Civic the results above is one example of a vehicle (among many) that uses radar.
I don't disagree that these systems should respond earlier than any of those tested did. I'll explain more in responses to some of the other comments after yours on this thread that had additional points and questions.
Last edited:
Someone claiming to know the driver posted on Reddit that the Tesla was following another vehicle, and the driver couldn't see the fire truck until the lead vehicle swerved into another lane. The driver reportedly said he didn't have time to avoid the collision, and wasn't sure whether his Tesla did any braking. If that report is largely correct, then when the details are finally reported I bet they will show that the Tesla's systems responded as quickly as the radar's view of the fire truck was not blocked, including automatic braking. Many people seeing the crash photographs have commented that a crash from the reported original speed of 65 mph should have been more severe. My suspicion is that automatic braking -- perhaps in addition to the driver's braking -- reduced speed before the crash.
If, when the details become public, we learn there was an unobstructed view of the fire truck, then I'm also puzzled why the Tesla would not have responded.
Last edited:
I worked for a company in the 1990s that designed, manufactured, and sold radar collision warning systems for heavy vehicles (mainly class 8 trucks). I assure you that our systems responded to stationary objects, and when they had enough confidence that the detected stationary object in the travel path was not an overhead or roadside structure, they provided the system's highest level of collision warning to the driver. The radar engineers who designed those systems had military radar experience and were very, very familiar with the issues of Doppler based systems -- but they designed clever ways of dealing with them for our use cases.
I have carefully watched progress of collision warning/avoidance systems since then, and am involved with some now. The unaided radar systems still have more trouble distinguishing stopped vehicles from normally-present road structures near the travel lane, like bridges, overhead signs, and sometimes roadside objects in curves ahead. To gain more confidence that the stationary object detected is not just such "road furniture", the typical approach is to delay action after first detection until the host vehicle is closer and can increase the chances that the detected object is really blocking the lane ahead. I suspect, but don't know, that is the case for all types of sensors -- but I know it is true for radar.
That is for unaided radar, and various approaches are used to improve those results. Typical ones are to use a vision sensor and compare results from both. Another is what Tesla introduced with software version 8.0, with changes to their approach for detecting stationary objects blocking the travel path that they explained here Upgrading Autopilot: Seeing the World in Radar . It is an approach that aids the radar with information previously gathered by other Tesla vehicles' radar and GPS, to evaluate whether the stationary object seeming to be in the travel lane appears to only be the same that is normally detected in the same location. It is not comparing the current detection to a generalized version of all possible road furniture. Instead, it is comparing to what has been detected at the same location by radar sensors on Tesla vehicles that previously drove the same path. That approach should give additional confidence whether a stationary object detected by a Tesla vehicle using that approach is actually a stationary object blocking the path (whatever it is -- stopped vehicle or "a landed flying saucer" as the Tesla post jokingly mentions) or simply road furniture that is normally in the same location as the currently-detected stationary object.
That's probably more than you wanted to know, but this is my world and I try to share pieces of it when I think it might help.
Last edited:
Please see my answer in post #56 in this thread to a similar point by chillaban.
Also, here are comments related to your point about the tested vehicles not stopping in time when initial speeds were above 25 - 30 kph. I suspect the sensors in those vehicles detected the stationary object at much greater distance than when they started braking. As I explained in my reply to chillaban, it is usually a concern about detections of overhead &/or roadside structures that prevent systems from acting immediately upon detection of stationary objects in the expected travel path. Any approach, that gives greater confidence that the detected object is truly blocking the travel path ahead, should allow braking to start earlier. With sufficient confidence, appropriate braking could start immediately after first detection. Automotive radar typically offers 160 meter (like Tesla's Bosch sensors) to 250 meter (like some truck collision mitigation systems) range. With long detection ranges like those, plus greater confidence that detected stationary objects are actually blocking the travel lane, warnings to the driver &/or non-emergency vehicle braking could be started much earlier than today's systems provide. (Tesla's may already be offering such improvements, based on greater confidence with their "fleet learning for radar" approach.)
Last edited:
Sorry -- I only addressed one of your three questions. After the Florida accident, Musk tweeted that the radar saw the crossing tractor-trailer but mistook it for an overhead object. Tesla addressed that issue with their "fleet learning for radar" introduced with software version 8.0.
What were details of your situation "about 2 weeks ago when I had to intervene", and what version of Tesla software did your vehicle have?
I have to ask... Let's say they generate a GPS map of all the overhead road signs to ignore but what if a large vehicle is disabled on the travel lane right beneath the overhead sign that is marked as a false positive. Will the vision then kick in to activate emergency braking overriding the radar false positive?
I don't know how Tesla is implementing their approach, beyond what they have reported and what I've seen reported by people in forums like this one. So I don't know whether vision is currently being used, after Tesla changed their approach in software version 8.0 as explained here Upgrading Autopilot: Seeing the World in Radar . However, they may not need vision in addition to radar to address the issue you posed. Their "fleet learning for radar" described in that Tesla article may be enough.
Even from their "ADAS map" data discussed on this thread Tesla Autopilot maps , particularly looking at the image below from page 6, I can't tell whether Tesla limited their approach to only locations of previously-detected roadway objects that were judged not dangerous. There potentially are additional characteristics, besides those needed to convert into location, of objects reported by radar that could be captured, sent to their servers, processed, and then included in the ADAS map tiles. A simple one is radar signal return strength, or radar cross section, which is derived from signal return strength. With either of those (or other characteristics some radar could capture such as size, shape, etc.) the same characteristics of the stationary object detected in real time could be compared to those contained for an object in the ADAS map tile for an object in a matching location. In the situation you posed, the stopped vehicle's radar return, radar cross section, size, shape, and/or vertical location might distinguish the stopped vehicle under the structure from the structure alone. If so, that would allow fast identification of that situation, versus presence of just the previously-detected, not-dangerous object (the overhead sign, in your example). The descriptions in the image below were the best guesses people made about the data. If the value guess-labeled as "Height" is correct, then the presence of a stopped vehicle underneath that object might cause the reported "Height" to differ. In that case, the mismatch could allow the stopped vehicle under the overhead object to be identified -- not as two separate objects, but as one "blended" one that does not match what was detected earlier. If instead the value is related to radar signal return strength and not height, I made the same point above about using that characteristic in addition to location, to identify something more than the previously-detected object is likely present.
Similar threads
