Snow sets off vehicle detection

[Winstonflame]

Every time it snows I have to turn off vehicle detection because I guess snow flakes look like cars. Any ideas on how to improve detection? 

[firmwaredev]

I'm not aware of any way to improve it.   I'm sure Google is working behind the scenes to better train the AI in general, so it may just get better as firmware updates are rolled out.   I see the same issue, with both snow and rain.   For me, this only happens at night.   If you watch snow or rain live at night on the camera, the relatively low frame rate and the use of IR makes it looks like streaks of light, which probably is why it's triggering as a vehicle.  Headlights crossing into my detection zones (with vehicle detection set), trigger similarly.  You can turn down the overall detection sensitivity in the app.  I think I did try that, and it didn't help for rain/snow - plus I don't want the sensitivity turned down, as it starts missing things I want to capture then.  Better AI training is likely the only way to really improve this, and that's up to Google.

[jbrooxie]

It isn't an AI training issue, it is a neural network implementation issue. They are not taking temporal data into account. Everytime the image of your car gets obscured by snow, rain, etc, their algorithm says hey "new car found". Their implementation needs to take into account past images to know that a recent car is not a new car. This is a design limitation/flaw of Google's software and the only way to get it fixed is to complain to Google and hope they take it seriously. It would be a major overhaul of their software, but it would put them ahead of their competition. The Netatmo Smart Outdoor Camera has the same issue.

[firmwaredev]

I can't speak for the OP, but the issue I'm talking about is not one of mistakenly seeing a car again that was obscured (and then unobscured) by snow (or something else).   Yes, I have seen that behavior before ( in my case a flag waving in front of car), and yes I would expect that.   What I'm saying is that it will misidentify snow or rain as a car - with no cars in the scene at all (or in the recent past).    That sounds like a basic AI training issue to me, but I'm no expert in this field, so maybe the neural network implementation might play a part in that as well.  In any case - I'm talking about a different problem than the one you describe.   Same answer to the OP of course - it's a Google problem.

[jbrooxie]

I agree, miss identifying snow for cars, I would expect to be a training issue. I haven't personally seen that issue. I was all bundled up in winter snow gear while shoveling and it was having trouble identifying me as a person.

[firmwaredev]

I've never seen it during the day - only at night on IR when such events look like big streaks of light going across the frame.  It's rather annoying, actually, generally requiring me to turn off notifications when it starts happening, which defeats the point of being notified at night for things I want to be notified of.  It's going to be a while I suspect before this gets to the point where it's free of all such false positives, if that's even ever possible.

[Brad]

Hey folks,

Thanks for visiting the Community. Since this thread hasn't had activity in a while, we're going to close it to keep content fresh.

We hope you were able to get the help you need, but if you're still having trouble, please feel free to submit another post, and provide as many details as possible so that others can lend a hand. Hope this helps!

Best regards,

Brad.