Thumbs up while walking

So, Turing had a birthday and Galileo recanted, the recent nearby asteroid is actually 1km in size and will cause some "issues" in 750 years and Monsters was a pretty good alien film.

Using the some ideas from http://www.andol.info/hci/1661.htm as a base of doing defect detection in opencv, I combined some recent developments together. Discarded for initial speed issues was a SURF system looking at a hand image. Sure, it can find the parts that match but I don't really need to know that I'm looking at a hand anymore, I've pretty much assured that.

Walking the +-x and +y directions got me my original (too) tight bounding box, so I decided to start walking the depth pixels +-y as well to get a decent coverage of what I should be seeing. Throwing in a +- depth check to get pixels probably belonging to what is a hand, making a rectangle and padding that by +30 pixels in all directions gave me a really good box at most distances from the Kinect.

This suffers more from a walking pixel than the previous method. Slapping your hand on your chest will pick up most of your body. I've an alternate idea that was provided to me by a good friend using a circle system if this does prove troublesome.

Anyway. I have a region of interest now that I can do some things with, contours and defects. Taking a straight bmp from the coding4fun tobitmap output really wasn't finding what I wanted it to. Turn on drawing of all contours and at certain distances curve your fingers towards the Kinect, they'll switch from the bright white through to black/dark and then the contour will show.

Pop in a ThresholdBinary and we can work on the biggest polygon contour, nab the convex hull and its defects. With defects and their start, depth and end points, we have our fingers.

Since I know which hand I'm working on as it's based from a skeleton point, my next step is to look at the new rotational information the 1.5 SDK is supposed to have. Get orientation and I can start looking at thumb distances and crafting a virtual hand. I think the fingers and numbering of could be tricky, but measuring distance gaps (taking into account depth) and providing a "Hold your hand up for calibration" and storing that info per recognised person (why yes, yes I will investigate voice and facial identification) should make it low maintenance.

I know I'm missing more of what I've been up to, but most of what I've been up to is crawling through fictional towns to avoid zombies while hunting for beans and ammo in DayZ.