I am working on a Kinect project using the infrared view and depth view. In the infrared view, using CVBlob library, I am able to extract some 2D points of interest. I want to find the depth of these 2D points. So I thought that I can use the depth view directly, something like this:
coordinates3D[0] = coordinates2D[0];
coordinates3D[1] = coordinates2D[1];
coordinates3D[2] = (USHORT*)(LockedRect.pBits)
[(int)coordinates2D[1] * Width + (int)coordinates2D[0]] >> 3;
I don't think this is the right formula to get the depth.
I am able to visualize the 2D points of interest in the depth view. If I get a point (x, y) in infrared view, then I draw it as a red point in the depth view at (x, y)
I noticed that the red points are not where I expect them to be (on an object). There is a systematic error in their locations.
I was of the opinion that the depth view and infrared views have one-to-one correspondence unlike the correspondence between the color view and depth view.
Is this indeed true or is there an offset between the IR and depth views? If there is an offset, can I somehow get the right depth value?
The depth sensor contains a monochrome CMOS sensor and infrared projector that help create the 3D imagery throughout the room. It also measures the distance of each point of the player's body by transmitting invisible near-infrared light and measuring its "time of flight" after it reflects off the objects.
The device features an "RGB camera, depth sensor and microphone array running proprietary software", which provide full-body 3D motion capture, facial recognition and voice recognition capabilities.
To get the depth data from the kinect, simply change the types of the framesource, framereader, and frame. Note also that the resolution of the depth camera is different from that of the color camera: 512*424 instead of 1920*1080.
An average depth accuracy error of under 2 mm was observed in the central region of the cone for the ranges from 0.5 to 3.0 m, 2 to 4 mm for ranges from 3.0 to 3.5 m, and over 4 mm for distances beyond 3.5 m. Accuracy evaluations of the Azure Kinect have been limited to date.
Depth and Color streams are not taken from the same point so they do not correspond to each other perfectly. Also they FOV (field of view) is different.
cameras
my corrections for 640x480 resolution for both streams
if (valid depth)
{
ax=(((x+10-xs2)*241)>>8)+xs2;
ay=(((y+30-ys2)*240)>>8)+ys2;
}
x,y
are in coordinates in depth imageax,ay
are out coordinates in color imagexs,ys = 640,480
xs2,ys2 = 320,240
as you can see my kinect has also y-offset which is weird (even bigger then x-offset). My conversion works well on ranges up to 2 m
I did not measure it further but it should work even then
do not forget to correct space coordinates from depth and depth image coordinates
pz=0.8+(float(rawdepth-6576)*0.00012115165336374002280501710376283);
px=-sin(58.5*deg*float(x-xs2)/float(xs))*pz;
py=+sin(45.6*deg*float(y-ys2)/float(ys))*pz;
pz=-pz;
px,py,pz
is point coordinate in [m] in space relative to kinect
I use coordinate system for camera with opposite Z direction therefore the negation of sign
PS. I have old model 1414 so newer models have probably different calibration parameters
There is no offset between the "IR View" and "Depth View". Primarily because they are the same thing.
The Kinect has 2 cameras. A RGB color camera and a depth camera, which uses an IR blaster to generate a field light field that is used when processing the data. These give you a color video stream and a depth data stream; there is no "IR view" separate from the depth data.
UPDATE:
They are actually the same thing. What you are referring to as a "depth view" is simply a colorized version of of the "IR view"; the black-and-white image is the "raw" data, while the color image is a processed version of the same data.
In the Kinect for Windows Toolkit, have a look in the KinectWpfViewers
project (if you installed the KinectExplorer-WPF
example, it should be there). In there is the KinectDepthViewer
and the DepthColorizer
classes. They will demonstrate how the colorized "depth view" is created.
UPDATE 2:
Per comments below what I've said above is almost entirely junk. I'll likely go edit it out or just delete my answer in full in the near future, until then it shall stand as a testament to my once invalid beliefs on what was coming from where.
Anyways... Have a look at the CoordinateMapper class as another possible solution. The link will take you to the managed code docs (which is what I'm familiar with), I'm looking around the C++ docs to see if I can find the equivalent.
I've used this to map the standard color and depth views. It may also map the IR view just as well (I wouldn't see why not), but I'm not 100% sure of that.
If you love us? You can donate to us via Paypal or buy me a coffee so we can maintain and grow! Thank you!
Donate Us With