Hacker News new | ask | show | jobs
by umvi 2082 days ago
I think detecting when a driver is falling asleep could be an amazing safety feature. You would need privacy guarantees about where camera data is going, but I think it would be worth the tradeoff.
2 comments

> I think detecting when a driver is falling asleep could be an amazing safety feature.

Welcome to 14 years ago, which is when Toyota first introduced their DMS/DAM: https://en.wikipedia.org/wiki/Driver_Monitoring_System

More generally, https://en.wikipedia.org/wiki/Driver_drowsiness_detection

Not even close in terms of complexity. I used many of the "Driver Drowsiness" detection from Toyota, BMW and MBenz on my previous cars. They always had false positives, inconsistent and pretty rudimentary. What you're implying is akin to saying Nokia/Ericsson/Palm had smartphones before the iPhone 1 came along. Therefore the iPhone is a non-event. ️

The "hacker" mentioned in this article discovered these detection labels on the Model 3/Y "selfie" camera:

BLINDED

DARK

EYES_CLOSED

EYES_DOWN

EYES_NOMINAL

EYES_UP

HEAD_DOWN

HEAD_TRUNC

LOOKING_LEFT

LOOKING_RIGHT

PHONE_USE

SUNGLASSES_EYES_LIKELY_NOMINAL

SUNGLASSES_LIKELY_EYES_DOWN

Maybe I'm missing something. But it seems like an apples to oranges in comparison to me. Correct me if I am wrong, but Comma.ai's driving monitoring feature is currently the only one in production that is comparable.

Hmm, "HEAD_TRUNC" presumably means out of the frame, but given we're talking about hurtling pieces of metal, it conjures up nastier images for me somehow.
False positives are a minor annoyance. False negatives are potentially catastrophic. So for this kinds of systems it makes sense to skew towards false positives. It is always a tradeoff.
They were definitely annoying alright. Even the collision detection on our previous MBenz SUV would go off at certain roads that were perfectly clear of obstacles.

Since OTA is not a thing on most cars. This becomes a permanent problem that a dealership cannot fix. Even if they wanted too.

With a Tesla (only one in the industry AFAIK), you can submit a bug report at the exact location via voice and a future software update might address it.

Open Pilot's driver monitoring could be improved. After the last update I got a ton of false positives when driving at night. Nighttime would make driver monitoring more difficult, but it's also when people are more likely to be distracted.
I agree. The great thing about software is you can always iteratively improve it over time. As long as you're not reaching the hardware limits.
Super Cruise also has a DM system which from what I hear is pretty good
Yes. Super Cruise is pretty good. The only caveat is that it only works on limited-access freeways.

https://www.gm.com/our-stories/technology/cadillac-to-increa...

Correction: atleast both seeing machines and smart eye has system in cars today that are more advanced. (GM, BMW etc)
I have actually tried BMW Smart Eye and GM super cruise. Comma.ai’s Open Pilot driver monitoring is far superior to both of those.

What OP was implying, is that the Tesla system mentioned on the article is on par with offerings from other OEMs from 14 years ago.

How did you try them? I thought they were hidden deep down away behind the OEM layers?

I have tried their the demo systems at several occasions, (they have been developong them in 20y) but not comma.ai's.

For example they calculate the opening speed of your eyelid with high precision, the gaze direction output is within a few degrees of error. You can almost cover your complete face and it will still track your head pose with extreme precision. The list of outputs are long. Pupils, glasses, faceid, blinks (velocity, duration), drowsiness /, perclose, the 3d geometry of the head, facemasks, gender, and a lotbof aggregatet values.

But comma.ai maybr has this too?

I don't believe you.

Open Pilot can't use your blindspot cameras or any other additional sensors in your car. If I try to merge into another lane, and someone is in my blind spot, Open Pilot will try (and will crash my car).

Super Cruise and Smart Eye don't have this problem.

And yes, I have an Open Pilot for my Lexus RX 450h.

I thought the topic of this thread was driver monitoring and driver drowsiness detection?

What you're describing is the entire semi-automated driving system + driver monitoring feature. Another Apple to Oranges comparison...

It's obvious to anyone familiar with Open Pilot that it can't access OEM blindspot cameras since it only uses the Comma Two smartphone cameras and hooks into CANBUS. So I am not really sure what point you're trying to make here.

Seeing machines and Smart Eye are DMS (driver monitoring systems). They have only acces to 1 camera in these cars and outputs details of the driver (gaze direction, head pose, eyelid etc.). What and how the outputs are used for are up to the OEM.
> Not even close in terms of complexity.

You are asserting with no idea whatsoever, or relevance really given you're apparently ignoring that I was answering a very specific comment.

> What you're implying is akin to saying Nokia/Ericsson/Palm had smartphones before the iPhone 1 came along. Therefore the iPhone is a non-event. ️

One, it really is not, there's nothing impressive-looking so far. Two, tesla's record of "game changing" is mostly "game changing marketing", and while you should absolutely feel free to give them all the benefit of every doubt, I really don't feel so inclined.

> The "hacker" mentioned in this article discovered these detection labels on the Model 3/Y "selfie" camera:

These are just events it might be able to generate, it tells you nothing about how well the events are detected and how the consumer of those events integrates them.

What do you think exactly, that other manufacturers just get a magical yes/no blob?

> You are asserting with no idea whatsoever, or relevance really given you're apparently ignoring that I was answering a very specific comment.

I am directly addressing your comment about "DMS/DAM". How is that not relevant?

> One, it really is not, there's nothing impressive-looking so far.

What is impressive to you? That's what the critics said when the iPhone 1 was announced. That's what many also said about electric cars. Funny how the metrics (sales/safety) played out on that one.

> These are just events it might be able to generate, it tells you nothing about how well the events are detected and how the consumer of those events integrates them.

The fact that it can detect those events using a specialized redundant NPU hardware is a far cry from "Driver drowsiness detection" and "DMS/DAM" from "14 years ago". Hence, an apple to oranges comparison.

> What do you think exactly, that other manufacturers just get a magical yes/no blob?

It's interesting you ask that, because they mostly are binary logic if you look at the ECU firmware. Some "Driver drowsiness detection" (MBenz) systems will also just beep at random intervals after driving non-stop for 1 hour. But they still market them as being able to "detect" drowsiness.

George Hotz talked about this many times. They also don't have any firmware code signing whatsoever (esp Toyota) which presents a security risk.

Out of curiosity, which of the DMS/DAM systems do you have experience with?

If history is a teacher, Musk is going to invent it in few years.
An actual few years, or a "fully self driving cars are a few years away" few years?
There are a lot of features that are, in and of themselves, really great, but I would not use until the privacy and data sharing issues are resolved. This is unfortunately a growing problem that many device manufacturers and software makers are making worse.