The best thing to do is to set the parameters to get the data as raw as possible and do your own processing, preferably save the .xed with the Kinect studio, this will be helpful in case the next version of the SDK does a better job tracking the skeleton you will be able to re run your system with the new SDK also allowing you to re run the system with different Transform Smooth Parameters so you can test the best settings for your scenario, although the .xed files are very large it has proven to be very useful for my team to always have the .xed of the tests that we run.

Dear Armando,

Thank you for the hint. I'm getting ready the system of up to 4 Kinects tracking the patient simultaneously, hoping the correlated data will be more valuable. Will publish the results (if any ;-) ). I like the idea with XED files, anyway will try this.

BTW - do you know if there is a programmatic method to get the raw data from the Kinect sensors (I mean using my proprietary code in C# instead of the Kinect Studio)? I wish I had the synchronized streams of the XED and my database of the joints when starting and stopping the experiment (the exact goal is to start recording and stop recording of the XED files simultaneously to the grabbing of the skeleton data thus have them correlated). Moreover - one more question. I cannot find any MS docs on .XED file format. How do you access those data?

I read in one of the many Kinect book a solution for this (it had most of the implementation in the book), but I haven´t done it yet and I can´t recall which book it was but the author proposed to make a copy of all the classes involved (depth frame, skeleton frame etc.) also this is not the same because you cant feed your saved depth frames to the Skeleton Tracking Engine so it beats in my opinion the main purpose for which I save the .xed which is being able to take advantage of further improvements on the Skeleton Tracking Engine of future SDKs. The .xed files are opened with the Kinect studio I haven´t tried to read them directly.

I think you ask this because it is difficult to synchronize various .xed with Kinect studio assuming it is even possible to emulate more than one Kinect Sensor with the Kinect Studio, I think its not.

A possible solution (I haven´t given it much thought) could be to set the Kinect Studio on different PC´s one for each angle and stream the data to one pc that would process all the views that is assuming you don't need to evaluate real time analysis of your system, in which case to synchronize all this you just need to have buffers for each stream and match the time frames.

There are sample codes of how to stream Kinect data, I haven´t tested any of them.

By the way we are working on similar projects maybe we could do some collaboration.

I have been browsing the web for a while for the .XED API and found one guy from Microsoft that was saying on their MSDN support that they will provide API "soon". The post was from the 2012 so I guess they have given up the idea. At the moment I'm collecting all of the joints in the relational database using skeleton tracking engine, organized by the frames and experiments (test sets) I believe for my experiments it would be enough at the moment to fine tune smoothness of the joints so this is the reason I'm seeking for the perfect set of the parameters. The goal is to constitute gait analysis for stroked patients. Surely we can collaborate, with pleasure.

I think that smoothing the data would result in loss of information, these parameters are intended for UI purposes and games in which you want the experience of using your app to seem as smooth as possible but for analyzing data, this would be like cheating the system as it would attenuate your patients deficiencies during gait, that's why we set them as raw as possible that would be:

Smoothing: 0

Correction: 1

Prediction: 0

MaxDeviationRadius: 0

JitterRadius: 0.3

Following my recommendations JitterRadius should be a big number, as the idea is to leave the raw data as is, but you know that tracking can have glitches set this accordingly depending on haw fast your patients are moving, best practice i.m.o is to set 0.3 which is 30 cm and minimize noise, avoid using dark or shiny clothes, avoid sunlight, if you are using harnesses for your patients avoid occlusion of lower members at all times, if you are using a treadmill use one that has reverse record patients turned around walking forward with the treadmill in reverse to avoid occlusion with the dashboard and use the Kinect in optimal distances 1 to 3 meters.

As a side note the majority of my patients are Parkinson patients to capture jitters is very important maybe in your patients it is not and you could smooth the data in any case since you don't do real time analysis you really don't need it, save your data raw and do your own post processing and filtering for whatever specific noise you may encounter.

Dear Armando

I've tried your settings today and it seems like this is working RAW indeed. I do had the prob lems with foot and ankle identification but this may be perhaps of the dressing (black trousers) of the model-person. I also found that with those settings cause the huge delay when starting the skeleton tracking (I mean tracking by setting TrackingState to Traced) byt once it starts, then traces quite well. I need to move to another lab with more space to remove any reflections of the walls, desks, etc. btw.

One more question if you mind - did you experiment with the Kinect height and relation to the patient position? Any hints? Should be on the chest level, head level?

Regards,

P.

Regarding height it depends on which part is the most important for you, in my case we are only looking at the feet, we are not paying attention to the arm swing or torso, we have it at the pelvis looking a bit down therefor torso and arms get a bit distorted but we have a good view of the legs and feet this is to minimize errors in the parts that we are focused.

It also depends if you are using the floor plane provided by the SDK, if you want to use it, its better to have it as low as possible even so the floor plane provided by the SDK is VERY unstable at least in my experience.

To minimize loosing track of the ankles and feet put the Kinect as close as you can, start tracking at a distance that you see the hole person and then bring it as close to the user as possible without loosing track of the parts that you are interested in.

As I said before avoid dark colors specially black, to get a good understanding of what's going on do a run of the experiment looking what the infrared camera sees, it will allow you to see all the sources of noise

that is: sunlight, shiny or reflective surfaces and very dark materials that absorb the Kinects infrared pattern you want to see this pattern as clearly as possible all the time.

Dear Armando - regarding your previous post - are you able to track the lower part of the body only (i.e. without shoulder center or spine joints)?

BTW - I successfully connected four Kinects simultaneously today, so all I need is a treadmill. Seems promising.

Regards,

P.