Hacker News Reader

Comment by northben | Hacker News Reader

Fellow HN readers: We're a smart bunch of hackers. Why don't we have a startup working on a technology that could recognize and alert lifeguards when swimmers are in danger? I bet a few cameras and some (admittedly advanced) machine learning could spot swimmers in need of assistance.

View on HN

This already exists. There are two categories of systems:

* with wristband (http://www.sealswimsafe.com/#!how-it-works/cok0)

* with underwater camera (http://www.poseidonsaveslives.com/ http://www.poolview.co.uk/swimeye/ http://www.angeleye.it/)

I was actually spending quite some thought about this. At least for water parks, I think this should be doable. My idea was to have a wide angled camera above the pool (distortion is not a problem here), with an algorithm that would recognize any structure that is not moving for a given time (lets say 30 seconds). This would create many false positives, probably, but the system could just activate a buzzer and zoom to the spot on a display that would be permanently monitored by some lifeguard. The number of false positives could probably be dramatically reduced, if there was some way to only react to non-moving-structures underneath the surface, maybe by some cameras from the side?

Problem is, when I was doing my lifeguard training they showed us / told us how the surface of the water distorted swimmers and who / what is below the surface. A towel or swim toy can look like a person once fully submerged. (it is required to wear polarized sunglasses in chair) You would need cameras under the water, focused on the bottom of the pool to be of any use I imagine. As a lifeguard, our main objective is scanning for signs of struggle yes, but also to be checking for bodies below the surface. With unresponsive victims, every second counts.

I like the idea, but the stakes are really high in downing. At some point you want it to be helpful, but not to be so good as to make the human lifegaurd not pay full attention. Machine learning might not be the way to go here..

It definitely in the realm of possibilities. They're starting to do the camera assist in the checkout product theft prevention.

http://www.stoplift.com/how-it-works/

Anthropic fallacy. It seems like you are assuming a human is or will always be better than another solution.

There may be solutions that are vastly superior to a human lifeguard.

They're pretty clearly not saying that no solution could be better than a human. They're saying that a solution which isn't as good as a human, combined with a human, could be worse than a human without the inadequate solution, which is true.

We have some pretty good machine learning tech out there, but I'd reckon we're quite a ways off from the proposed startup focusing on replacing lifeguards with computers even at spotting functionality, let alone the other functions necessary (retrieval, resuscitation iff actually necessary, etc.).

Sure, someday we'll have robots do everything for us. We're not quite there yet.

A rectangular pool with a flat bottom (probably accounts for the majority of pool volume world-wide) would be pretty easy to have full camera coverage over with minimal image distortion. Each object could be easily tracked for time underwater, depth, etc. (and expanded to track fine motion as well) Objects that stay near the bottom are suspect. The system could alert lifeguards to the exact location of a victim.

You could even have an auto-drain feature if you desired.

Right, but my point is that we're not quite at the point where a computer can reasonably process that much visual input, determine whether or not something is exhibiting the signs of drowning, and alert a lifeguard with a minimum of errors. False positives will lead to personnel confusion, and false negatives will lead to dead swimmers.

This isn't to mention that automation does encourage complacency. One of the primary reasons for the recent Malaysia Air jet disappearances according to investigations is an excessive reliance upon automated systems to fly planes, and insufficient knowledge of flying without such tools. I reckon a similar situation could be a real danger here as well (if not amplified considerably; most lifeguards (at least where I've lived) were usually highschool students doing it as an extracurricular/volunteer activity, summer job, etc., and teenagers aren't exactly known for having an above-average attention span).

> False positives will lead to personnel confusion

False positives can lead to death too. This article talks about alert fatigue that contributed to a fatal medication error at UCSF:

https://medium.com/backchannel/beware-of-the-robot-pharmacis...

I'd be wary of giving a lifeguard who might be straining to pay attention to n things an n+1th thing to pay attention to.

That's what I meant by "personnel confusion", yes. A.k.a. the "cried 'wolf'" effect; how can a lifeguard be expected to treat this alarm seriously when all the other ones have been false positives?

As a non-lifeguard, I would rather have a machine than trust myself. So ideally, you would have both a lifeguard and another independent monitoring system for non-lifeguards. They could be fail-safes for each other.

Most of the startup scene is about making a new JavaScript framework, not actually solving the hard problems.

Startups often follow the money. I would be hesitant to blame startups for what is really a far deeper and far more pervasive cultural problem.

The first challenge is training data. 10-100 videos of people beginning to drown would be a good start. It also might be highly pool dependent. The different shapes, styles and attractions at the pool will all make a difference. Wave pool vs slope vs dropoff etc

Go for it.

Reply on news.ycombinator.com