I grazed a deer yesterday--no alert from car

[Jlwine]

I was driving about 45MPH in broad daylight yesterday in middle of the city, when two fawns jumped out right in front of me. There was oncoming traffic from the other direction--so I could not swerve into the adjacent lane. I braked hard and just barely grazed the second deer. I only know this because I heard the strike--fortunately and surprisingly there was no visible damage to the vehicle.
My question is, should I not have expected any alert from my Model S? I was not using auto-pilot at the time (I have AP 1 if that matters). However, I was under the belief--perhaps incorrectly, that it was designed to warn you or even brake if a "pedestrian" were to come into your immediate path. Is this not correct?

 

[jelloslug]

The car cannot detect something that small approaching from the side, especially when you are going that fast.

 

[alcibiades]

The car cannot detect something that small approaching from the side, especially when you are going that fast.

Why not?

 

[Jlwine]

The car cannot detect something that small approaching from the side, especially when you are going that fast.

I guess small is relative, yes it was smaller than full sized adult--but much larger than most kids 12 and under.

 

[Max*]

Again, a deer is not a person. A cat is not a person. A buffalo is not a person. A donkey is not a person.

If the car is trained to detect a person and classify everything else as other, then something like

IF(high confidence that what I'm seeing is a person)
THEN (apply emergency braking)
ELSE (do nothing/track target/add to database to decode later/play happy music/whatever)

Would still stop for a [small] person, but not for a huge cow.


I'm not saying this is how it works, it's an example. I'm just saying we need to stop assuming that if the car didn't stop for X, that it wont stop for a person.

 

[Max*]

The car cannot detect something that small approaching from the side, especially when you are going that fast.

And you have evidence of this...?

I'm sure the car traveling 45mph can DETECT a fawn with the camera. The real question might be does it have enough time to do anything about it? Does it even know what a fawn is? Is it programmed to classify a fawn? What about an adult deer? etc.

 

[Jlwine]

And you have evidence of this...?

I'm sure the car traveling 45mph can DETECT a fawn with the camera. The real question might be does it have enough time to do anything about it? Does it even know what a fawn is? Is it programmed to classify a fawn? What about an adult deer? etc.

Whether is is a fawn, a human, a polar bear or a concrete barrier--it is an object in its immediate path. I would like to have though that if I had enough time to see and react and all but avoid any impact, that the computer system would have had enough time to alert--at least after the first deer crossed directly in front.

 

[croman]

The car should've recognized a living breathing large mammal and at least alerted to the presence of an object that could cause injury or death to you.

Maybe @Max* is right and its programmed to only detect humans but that seems short sighted (pun intended!) with deer collisions being so common and dangerous.

 

[Max*]

Whether is is a fawn, a human, a polar bear or a concrete barrier--it is an object in its immediate path. I would like to have though that if I had enough time to see and react and all but avoid any impact, that the computer system would have had enough time to alert--at least after the first deer crossed directly in front.

I completely agree with you and @croman - it should do that.

But, and you know there's always a but.
-You're over simplifying the problem.

-AEB is not meant to stop for other things, so Tesla did the bare minimum

Automatic emergency braking (AEB) systems detect an impending forward crash with another vehicle in time to avoid or mitigate the crash.

-Your version of the code needs to be close to 100% perfect. And we all know nothing Tesla has right now is close to 100% perfect.
Having it accidentally apply full force breaking when it thinks something is going to hit your car (let's say a whicker chair flying out of a truck at 80mph) may be more dangerous than hitting the damn chair. Not detecting objects (misses) would be much safer than false positives (braking when nothing is there).
But then how do you advertise something that's not perfect? "We can stop 93.5% of the time for things in the immediate path of the car". No. You wait to get it to 99.xxxxx%, and then sell it.

-For moving objects this gets even harder. Going at 65mph, you're braking distance is about 200ft. You then need to make sure you only apply the brakes when a collision is imminent. But a deer could run away in those ~5 seconds that the car is braking.

-I'm not aware of any manufacturer* that does the "everything" detection. I know BMW/etc. have night vision that detects animals, I haven't read that it actually stops for you.


Again, all of this is possible to do. And I'm sure the technology will get there, and actually other manufacturers are very close.

*Volvo has a "large animal detection" system, not sure if they rolled it out in real vehicles yet nor how effective it is.
*I know other AEB tests have shown a brick wall and stopping.

 

[Jlwine]

I completely agree with you and @croman - it should do that.

But, and you know there's always a but.
-You're over simplifying the problem.

-AEB is not meant to stop for other things, so Tesla did the bare minimum


-Your version of the code needs to be close to 100% perfect. And we all know nothing Tesla has right now is close to 100% perfect.
Having it accidentally apply full force breaking when it thinks something is going to hit your car (let's say a whicker chair flying out of a truck at 80mph) may be more dangerous than hitting the damn chair. Not detecting objects (misses) would be much safer than false positives (braking when nothing is there).
But then how do you advertise something that's not perfect? "We can stop 93.5% of the time for things in the immediate path of the car". No. You wait to get it to 99.xxxxx%, and then sell it.

-For moving objects this gets even harder. Going at 65mph, you're braking distance is about 200ft. You then need to make sure you only apply the brakes when a collision is imminent. But a deer could run away in those ~5 seconds that the car is braking.

-I'm not aware of any manufacturer* that does the "everything" detection. I know BMW/etc. have night vision that detects animals, I haven't read that it actually stops for you.


Again, all of this is possible to do. And I'm sure the technology will get there, and actually other manufacturers are very close.

*Volvo has a "large animal detection" system, not sure if they rolled it out in real vehicles yet nor how effective it is.
*I know other AEB tests have shown a brick wall and stopping.

I agree that it needs near 100% certainty of a crash before breaking--its the fact that there was no alert that surprised me.

I did call Tesla just now and they said they would pull the logs to see if they see anything indicating any detection and let me know. I hope they do follow-up--for my curiosity if nothing else.

 

[oktane]

Why? It's because AEB and FCW on Tesla is a joke. It is useless. The feature works much better on my former luxury brand car from 2002 than Tesla. Tesla's solution is a half-baked rush job meant to insulate from lawsuit. It should not be trusted. Just like side collision, the features don't work as advertised, and not to the level of $30,000 cars even.

 

[Max*]

Why? It's because AEB and FCW on Tesla is a joke. It is useless. The feature works much better on my former luxury brand car from 2002 than Tesla. Tesla's solution is a half-baked rush job meant to insulate from lawsuit. It should not be trusted. Just like side collision, the features don't work as advertised, and not to the level of $30,000 cars even.

Two questions:
1. Have you ever had AEB activate?
2. Have you ever crashed into a car in front of you?

 

[oktane]

Two questions:
1. Have you ever had AEB activate?
2. Have you ever crashed into a car in front of you?

I had FCW activate once erroneously. Never had AEB (fortunately). On most of my other cars, the car will properly and consistently warn me when there I am closing fast on slowing traffic. I never get any alert on the Tesla, even though I should. It's a nice feature for there to be a beep if potentially getting into a dangerous situation (like a car deciding to stop and turn left all of a sudden).

I've had my other car save me while passing at freeway speed. Another car didn't see me and jumped into my lane. I had my foot on the accelerator but not only did the car cut off my accelerator but applied full brake, which is the only thing that prevented contact. I was amazed at how well it worked. I do not have any confidence in the Tesla product, which works only at low speeds, given I get no warnings when there should be audible warnings, and seeing how sluggish and poorly the blind-spot feature functions.

 

[Max*]

I had FCW activate once erroneously. Never had AEB (fortunately). On most of my other cars, the car will properly and consistently warn me when there I am closing fast on slowing traffic. I never get any alert on the Tesla, even though I should. It's a nice feature for there to be a beep if potentially getting into a dangerous situation (like a car deciding to stop and turn left all of a sudden).

I've had my other car save me while passing at freeway speed. Another car didn't see me and jumped into my lane. I had my foot on the accelerator but not only did the car cut off my accelerator but applied full brake, which is the only thing that prevented contact. I was amazed at how well it worked. I do not have any confidence in the Tesla product, which works only at low speeds, given I get no warnings when there should be audible warnings, and seeing how sluggish and poorly the blind-spot feature functions.

I have pretty much the opposite experience, but I think this might be an AP1 vs. AP2 discussion in that case.

I have my FCW set to early, it has a lot of false positives. But each time it mattered it has beeped at me consistently (similar to your previous cars). I could change it to normal (or whatever the setting is), and have less false positives, but I'm more than happy with the way it is.

I've had AEB activate once (Emergency Braking in Progress), I can tell you for a fact that on an AP1 car it's not a gimmick and it's not something that was put in place to just get a check mark. It works.

I've also had the side collision avoidance activate twice (Side Collision Avoidance - disappointing?), both times were false activations, and I honestly expected more from it. I feel they have a lot of room for improvement here, with AP2 (I have no hopes for my AP1 car).

I believe when you mentioned the side collision, you were talking about the blind spot monitoring (in both AP1 and AP2), I agree - that's a gimmick at best.


But to put this thread back on topic. I think stopping for "everything" will happen, but I'm not surprised it's not there yet.

I think first they should put the "stop for everything" code into the FCW, that'll decouple it from any dangerous false braking events that AEB can trigger, and would satisfy a lot of people. But again, I don't think this is a trivial thing to do.

 

[whttiger25]

I was driving about 45MPH in broad daylight yesterday in middle of the city, when two fawns jumped out right in front of me. There was oncoming traffic from the other direction--so I could not swerve into the adjacent lane. I braked hard and just barely grazed the second deer. I only know this because I heard the strike--fortunately and surprisingly there was no visible damage to the vehicle.
My question is, should I not have expected any alert from my Model S? I was not using auto-pilot at the time (I have AP 1 if that matters). However, I was under the belief--perhaps incorrectly, that it was designed to warn you or even brake if a "pedestrian" were to come into your immediate path. Is this not correct?

Was the deer ok?

 

[MrAustraliaTax]

It didn't activate because the deer were not not crossing over a bridge you were driving under at the time.

EDIT: Never mind. It was AP1.

 

[scottm]

The deer had mistakenly believed all the hype it had been hearing about Tesla cars, and truly thought it would a safe bet walking in front of the car. However, this type of "testing" is not advised, says so in the manual. But the problem is deer can't read and hooves don't work so well on the touch screen.

There's always this too: Donate Wild Game – Hoosiers Feeding the Hungry
but fawns are a little lite..

 

[cgiGuy]

If it detected everything now, we wouldn't have a reason to upgrade our S/X's when the new cars come out in 6 months with FLIR sensors.

 

[cgiGuy]

I had FCW activate once erroneously.

With it set to "early" I get warnings quite frequently. I usually say, "Really? That concerned you, car?" And my wife usually says, "It concerned me, too."

 

[wwu123]

The car should've recognized a living breathing large mammal and at least alerted to the presence of an object that could cause injury or death to you.

Maybe @Max* is right and its programmed to only detect humans but that seems short sighted (pun intended!) with deer collisions being so common and dangerous.

I'm not in the AI business, but all this talk about machine learning and faster processors I think makes people ambitious in what can be practically computed today. Tesla has no LIDAR or IR, so all it has to process vision and identify objects from the background is bunch of pixels, maybe a smear of blurry pixels if the item is literally "jumping out" in front of the car (think of how clear your TV looks when you pause an action sequence).

45 mph is a closing rate of nearly 70 ft per second. At 2 seconds from impact, that's 140 ft away, and a 4ft tall deer has an angular size of less than 2 degrees. Is that small spot of brown blurry pixels a deer 140 feet away, or just a big leaf blowing across the road 6 ft in front of the car? Those have the same angular size, and there's no sense of distance. How does it know it's a large living breathing mammal? Just programming it to recognize the shape of the front of a deer, the back, side, antlers, jumping, standing, running, might be a huge amount of memory and computation in itself, maybe equal to humans; then double it to add moose, elk, etc.

Sure, add in motion, and with 20-30 frames of data it starts to get easier. But now at one second, with a lot more pixels to distinguish a deer from a leaf blowing across the windshield, there's a lot less time to judge and then brake (since I'm sure auto-swerving is just not something computers can or should do) to avoid hitting the obstacle, if even possible at that distance.

I'd prefer they start just trying to recognize and react to trucks, buses, bicycles, garbage cans, orange cones, and other urban hazards that the Autopilot is more likely to hit, before moving on to deer and other large mammals.

EDIT: After writing this I read that AP 2.0 does have radar, which I'd assumed it doesn't after many articles debating LIDAR. Radar would certainly help with distance of an object and ultimately identification of the object, in addition to decision to brake. Nonetheless still may be a challenge, at least today vs a few more years from now.