xx We can use your help to test Keith's Video Mapper

[JonB256]

no tripod for phone. Certainly not rock-steady but probably representative of the videos that people will try to use.

[Phrog30]

Just sit the phone on a box or something.

[sean]

Keith wanted me to remind everyone that this capability is currently a work in process. I said we could see this in the next release, to be safe it will be done as time allows for the work. Keith has his work and family to balance and I do not want to put undue stress by setting an arbitrary dead line. We are still in April.  Keith and I are having discussions on the algorithms , Keith is doing all the work. He really could not proceed without the numerous videos so thanks for giving input to the process, ill keep you posted on the progress. The challenge is how large the data frames are from video.

Even a low resolution frame of 800x600 is 480,000 pixels. The video is 30 frames/sec .
So 14.4 million bits of info per second. For each frame of 480,000 pixels we need to look at their value and determine that there is a light there.
The lights create halos so instead of a pixel of 255 at say 300,400 we could get 300 pixels all in a circle, all 255. We need to process all those pixels to determine the center x,y location. Do this 30 times per second times 150 nodes.

You get the idea. It is not a trivial task.

xLights has always been very open on the future tasks/effects we are working on. We could have said it was a secret until july/aug or something. I like to let you know what we are working on early. Not everyone agrees with this philosophy.

So watch for periodic updates on the automatic custom models from Video project over the coming weeks/months.

sean

[Phrog30]

While you can't please everyone, I like that you guys are open and honest. There are other groups that keep things secret but also make hints and it just draws things out. I vote you keep doing what you have been doing. For me, I'm enjoying the ride. 

James

Sent from my SM-G900V using Tapatalk

[Gilrock]

When I first heard the idea it reminded me of when I used to do astronomy and asteroid discovery at my old house where I had an observatory in my backyard.  I had software that could do plate solving by detecting all the stars in the images and figuring out their alignment assuming you knew your approximate RA/DEC.  I also had software I worked on where we would detect camera blooming in images and automatically remove it.  This project sounds interesting although I usually am never able to actually power up my display elements before I sequence.

[arw01]

Perhaps reaching out to the student or amateur astronomy programmers could bring in some code to assist with this, or perhaps even use their star mapping software and then do the math on the results, certainly interesting capability of Kieth can pull it off.

I may be able to get a couple more James items videos made as well in the next few days.

Alan

[sean]

Gil; i was thinking of you last night as i searched numerous open sourced software for locating stars. Astronomy has it easier, normally looking at a few plates of info. Asteroids is more like this, look at thousands of stars on multiple frames and detect movement. Our problem is a little different since our "stars" look so large on frame it is like finding supernovas every second. Certainly an interesting problem


Sent from my iPhone using Tapatalk

[Gilrock]

I used to stack multiple frames to get a better signal to noise ratio.  If you look at this webpage the left images are 3 single frames and the images on the right are a stack of 6 images for each frame.  You can see how it reduces the apparent noise and you can now see the asteroid moving in the 3 frames.
http://www.threebuttes.com/TrackAndStack.htm

[flyinverted]

I wonder if Keith could incorporate the UI to allow the user to configure areas to ignore, just as many home surveillance camera software products do.

The next level would be to then make the green boxes smaller as shown in the photo. The user would then choose the boxes that have pixels in them and this would push the accuracy into the upper 90% range.

-Peace,

-Steve

[JonB256]

Steve, if you are going to do an overlay like that, then perhaps you could just do a "follow the dots and click on them as the light" with your mouse. Might need to slow it down a bit, though, especially if someone is using a trackpad. A touchscreen on an iPad or large tablet would be pretty cool, like using the touch sensitive whiteboards in so many classrooms now.

[keithsw1111]

I don't think this is necessary. Security systems are different beasts. The only differences I expect to see in frames is lights turning on and off. My biggest problems are flare and reflections. Both make identifying bulbs hard. But I am getting there.

[keithsw1112]

A bit of an update.

It isnt perfect yet ... but the results look pretty cool. Attached are 2 xmodel examples generated.

The first is a static light model of a policeman from an image I found on one of the lighting forums. A really challenging model to map due to diffuse incandescent lights ... and even worse blue ... the worst colour for contrast. Because it is static I can work off just a photo ... no video required (or supported).

The second is the model generated from the first video I was sent. Its not perfect. I am missing 2 nodes because of the settings I was playing with to try and reduce false bulb identification from reflections but still not too bad. These reflections are also the cause of the model being so large.

It will only get better from here.

[Gilrock]

Not sure if it would be useful but here's an algorithm I've seen used when I used to run software that did auto-focusing.  It calculates the half flux diameter of a star but it would work on a light.  The smaller the HFD the more in focus the light is.  You could use if you ever get into automatically adjusting brightness and/or focus.

http://www.lost-infinity.com/night-sky-image-processing-part-6-measuring-the-half-flux-diameter-hfd-of-a-star-a-simple-c-implementation/