Hacker News Reader

Comment by konker | Hacker News Reader

"So the final product is a multi-camera, rear-facing bike-light, which uses object detection and depth mapping to track the cars behind you - and know their x, y, and z position relative to you (and the x, y, and z of their edges, which is important).

What does it do with this information?

It's normally blinking/patterning like a normal bike light, and if it detects danger, it does one of two actions, and sometimes both:

1. Starts flashing an ultra-bright strobe - to try to get the driver's attention

2. Initiates haptic feedback - to let the person riding the bike that they're in danger

3. Honks a car horn. This is if all else fails, and only happens if you are in certain danger"

Source: https://discuss.luxonis.com/d/8-it-works-working-prototype-o...

View on HN

I see in Germany older folks often using a mirror on the left handlebar. Once I borrowed such bicycle, it was somehow weird, but I got used to that. Such device identifies cars and dangerous situation very well. Zero false positives, works pretty well in the night too. No batteries needed and costs between 5 and 15€.

Yes, mirrors are a fantastic solution. And for getting the attention of the distracted driver (as sometimes the dynamics of the bike preclude maneuvering out of the way in time), LOUD bicycle is a great solution: https://www.kickstarter.com/projects/lansey/loud-bicycle-car...

This sounds like a great idea in theory but in practice my guess is that its too little (this won't help cars pulling out of junctions into bikes and many other dangerous scenarios) and too late (how much time will the driver have to see the strobe/hear the horn and take evasive action).

The Radar powered Garmin Varia (https://buy.garmin.com/en-GB/GB/p/601468) is probably simpler and cheaper and I'd guess more likely to work reliably in bad lighting conditions.

But a good set of continuously flashing ultra bright strobe lights, high visibility/reflective clothing, and bike lanes or driver education is by far superior.

Thanks for the comments! So the radar solution is great for touring riders, where one doesn't see a car for half an hour or a couple hours. It mentally wakes them up to the presence of a car.

It's not intended for use in commuting though... as in commuting you don't care about the 500 cars that will pass you safely on your ride (and you don't want 500 alerts), you care about the 1 that is on a trajectory to kill you.

So with computer vision you can tell that trajectory. You can also with radar, but only with super-expensive radars ($100k+), wheras you can do this w/ computer vision for $50 or so.

So it's like the radar-based Garmin RTL5xx (née Backtracker) but with the added benefit of "hit or pass" lane information and all the warning escalation you can do with that. I'm generally extremely sceptical of technical solutions to driver inventiveness (they will eventually out-risk-compensate everything you throw at them, it's a one-sided arms race) but they might be on to something here. The Garmin is already surprisingly popular and proper use of lateral information would greatly add to the appeal. A contender for the annual Garmin Fitness acquisition 2021 perhaps?

edit: Still, I'd be afraid that falls from startled reactions by the cyclists themselves might in the end cause more bloodshed than the extremely rare rear-ender it's supposed to prevent. A car preparing a well timed overtake (close in fast inside the lane to sweep left just after oncoming traffic passes) will be a false positive you can't avoid without directly guessing intention and skill of the driver. It will surely be a net-positive in some communities with particularly low cyclist awareness amongst drivers but not necessarily in others.

Thanks for the comments usrusr! And great points. So one of the things we did early-on is to prototype the whole flow to see if you could indeed get incredibly low false-positives. The key ended up being the combination of disparity depth (which gives over 1 million depth points) and neural inference so that the system can know incredibly granularly (i.e. within inches) the edges of things and their predictive trajectories.

So at the outset of this effort (back in 2017) such a computer-vision based device that was inexpensive and embeddable was impossible. The Movidius Myriad X came out, which has all the requisite CV/AI processing to make such a thing solvable... but there was no hardware/firmware/software/AI ecosystem which allowed producing such embedded spatial AI problems.

So we had to build that first to solve this safety problem and we're releasing it in conjunction with OpenCV as the OpenCV AI Kit (OAK). https://opencv.org/opencv-spatial-ai-competition/

So the `why` of this is it allows others to solve problems on embedded systems which were previously intractable. And there are a ton of them... the number of applications we've seen has blown our mind.

We'll have a KickStarter around this OpenCV AI Kit which will be going live on July 14th: https://www.kickstarter.com/projects/opencv/opencv-ai-kit

Thoughts?

Thanks, Brandon

Reply on news.ycombinator.com