Hi Andrea, we had a similar problem when we wanted to mount the Kinect to track workers from above. To our knowledge both the Kinect SDK and the OpenNI / NITE middleware can only track horicontally. However in your case I suppose you only need to rotate the image by 90 degrees before it reaches the middleware. OpenCV can do that.

Skeleton tracking occurs when the app using Kinect defines the joints of the skeleton. Therefore, it can easily detects the sitting and standing positions. When it comes to lying on the floor, kinect cannot detect the joints and that is why it fails to detect the skeleton unless you do relocate your kinect in a suitable orientation to the human body. In a conclusion, you should explore the best location of kinect to be to detect the 3 positions of human body.

Try to explore Omek Beckon software for full human body skeleton detection.

I explored Omek Beckon software before and published a video which you can find it below

https://www.researchgate.net/post/Omek_Beckon_Software

For my point of view, if you don´t have enough distance between the floor and the body, the kinect won´t be able to detects the skeleton. What you could do, as the kinect work with infrared, it is to create a gradient of temperature. I already notice, if the surface is warm (up to 45C), you just get a black gap on the control screen where you have the hot spot. May be, if the floor is warm enough, only the body will be visible by the kinect, making the skeleton easier to detect. This is just an idea, I didn´t test that myself.

We wrote software to do custom tracking of hands and arms from above with KinectArms. While that doesn't directly help you, the fundamentals of getting it working were simple enough. You could capture your stream with a second layer of your application and write custom code that kicks in when the API's skeletal tracking fails.

Our code is open-source on GitHub (search KinectArms or contact me if you can't find it). It might be a reasonable starting point for bootstrapping the process.

Good luck!

As Geoffrey Marchal said, if you don´t have enough distance between the floor and the body, the Kinect won´t be able to properly track the skeleton, you can try tricking the sensor I haven't tested with warm surfaces but mirrors or certain black cloths can help, you can test how much reflection a material gives by accessing the infrared stream or the depth stream in the Kinect Explorer, although the best solution would be to write your own code for that specific situation even in this case "adding" space behind the user improves detection and simplifies code said code.

If you view some Microsoft Research videos on the Kinect, you'll find that the skeleton tracking works off of a specific set of poses that MS trained it on. They took a model of a human (cg) with each part (arm, leg) labeled, and then programmatically manipulated the "puppet" into all the different poses expected to be encountered, and then fed the render data, along with programmatic depth data, and labeled parts into a Decision Tree network that learned to classify the pixels as belonging to the leg, arm, head, etc. So in the final pass, what is happening is that your image and depth data is entering the pre-trained classifier, and the classifier is outputting one of the predefined poses (which explains where the info for "hidden" body parts comes from), and that pose is matched to your image on the screen, and then the joint locations are "projected" from the puppet onto your image+depth and final coordinates given.

In this case, their pose library did not include lying down, so their is no chance of the classifier giving skeleton data because their "puppet" was never in that pose.

So you can either put a kinect above, and if it detects a "standing" person, then the person is lying down. Or you can implement the machine learning algorithm yourself or use an off the shelf one and train it with lying down poses.

take a look here: http://research.microsoft.com/en-us/projects/vrkinect/

Kinect is designed to read human body movement w.r.t gaming, hence it expects that the person in front of the camera is standing.

If you really want to detect a lying person then you need to get into the programming level of the Kinect. There are SDk and all the required helper software available to develop Kinect based application. a good place to start with will be:

http://www.microsoft.com/en-us/kinectforwindowsdev/start.aspx

the SDK is available here:

http://www.microsoft.com/en-in/download/details.aspx?id=40278

Most of than standard API will not consider the lying position, so u need to write one extension of your own. Need to get into the vary details of the API's given in the SDK.

Best of luck.

Hi,

May I suggest first placing kinect above then detecting the person, if this is correct, try to detect standing with the configuration above. It may work.

Cheers,

Pedro.

Detecting body parts from single depth image may help your problem. If you train a dataset over lying people with different scales and orientations regarding the camera, you may get a good result.

A good paper regarding that matter can be found here: http://dl.acm.org/citation.cfm?id=2398381