it is not in my opust in work

I find local minima to detect the first and last change in the signal then I delete the signal before and after these changes. I replace zeros with nans and then interpolate it before I apply smoothing technique on the signal .

I get a good results. But I wonder if is this something good to do or I should use the moving average?

Which motion capture system are you using? If you are using Vicon, then they have a few methods of interpolation (e.g.spline filling) in the software (e.g. Nexus). There is a lot of manual work, visually going through frame by frame, and then cut/delete from when the marker goes wonky. Then you identify the same marker when it presents itself again so you have the trajectory of the same marker but with a gap in the middle. Use one of the spline filling/interpolation in Vicon to fill the missing gap. You may need to do this several times because the interpolation may not be to your standard yet. In essence the missing gap should get smaller with each attempt which makes the interpolation algorithms work better. After several run throughs, your marker will have the gap filled in a perceptually normal way.

This type of motion capture editing work is long, arduous and boring, but it will get you the cleanest data possible. 

I normally gap-fill with Qualisys/Vicon first and then with MATLAB. I suggest spline filling as it is more accurat than the linear method.

Whichever method you choose, I would simply remove the data if the gap is greater than XX frames (number of frames depends on sampling frequency and type of movement so there isn't a fixed answer for this).

Finally, I always plot my data with the gap-filled parts highlighed to visually check that the gap-filling is "sensible" as sometimes you get results that are clearly wrong, in which case you should just remove that portion of data from you analisys.

 Dear Rufaida Hussain

A correct way to solve the problem, see

https://www.researchgate.net/publication/262801364_ABOUT_TWO_THE_LEAST-SQUARES_METHOD_MODIFICATIONS_FOR_LOST_DATA_RECOVERY_IN_VIDEOANALYSIS_SYSTEM_BASED_ON_ACCELEROMETER_DATA

But it needs additional sensors

Regards

Pavel

Article ABOUT TWO THE LEAST-SQUARES METHOD MODIFICATIONS FOR LOST DA...

Hi,

I can see two problems here: gap filling and data smoothing.

If the regions of interest or some parts of the signal drastically changed when filtering the data (especially with a Butterworth filter) then it clearly means that you don't use the appropriate cut-off frequency for the filter.  The chosen cut-off frequency depends of the system noise used and the sampling rate of your data.  Have you heard of the residual error technique (page 41-43 of David Winter's book: "Biomechanics and Motor Control of Human Movement, 2nd Edition"). It is a technique that helps you to choose the best cut-off frequency for your data.

Concerning the gap filling issue, what software are you using to fill the gaps? Some functions are not designed to fill gaps when they appear at the beginning or at the end of the data set. This is called extrapolation.  If you are using Matlab, then I recommend you to have a look at the function "interp1".  https://uk.mathworks.com/help/matlab/ref/interp1.html . it gives you the possibility to choose different type interpolations ('nearest', 'next', 'previous', 'linear','spline','pchip', or 'cubic').  It also gives the possibility to extrapolate the data. 

As Eleonora said, it is very important to check the result of the interpolation.  It is not recommended to interpolate the data when the gap is too big. "Too big" is objective as it depends, among the other things, to the speed of the marker.  What is important to remember is that in biomechanics, the movement of interpolated data needs to be /look realistic.  If the gap is too long then it is like creating data / movement that didn't appear.  

The function "inter1gap" (https://uk.mathworks.com/matlabcentral/fileexchange/45842-interp1gap) gives you the possibility to set a threshold.  The function will only interpolate gaps that have a number of frame smaller then the threshold. 

Also, I recommend to filter the data first and then fill the gap. 

Cheers,

Bertrand Bru

Hi,

Actually there are plenty of methods in the state of the art that are much more sophisticated than interpolation. These method may use machine learning, body constraint and/or movement redundancy and low rank properties of matrices to recover data better than interpolation. Some methods are automatic (they use only the information contained in your mocap sequence), and some need training on other data, or additional information of the structure of your marker set.

I am currently working on an automatic method myself (that I hope to publish soon). For the time being, here is a quite recent, automatic and easy-to-use method where you can find the matlab code in open access. Moreover, it has specifically been tested on walking data (Gloersen et al. 2016) :

http://journals.plos.org/plosone/article?id=10.1371/journal.pone.0152616

In a few words, this method is based on PCA (i.e. using matrix low rank properties), and assigns bigger weight to trajectories of markers that are closer to the corrupted one for its reconstruction.

You can find a zip-file at the end of the article. Of course the method works better if you have long motion capture files and a lot of markers, as it gives more information to the PCA algorithm. (And of course my method will be better ;) )

As Bertrand suggests, it may be useful to filter data first and then fill gaps. The problem when doing so is to keep correct data at edges of the files and more importantly at the borders of the gaps in your data.

You can find my own filter implementation in attachment (only tested on my data with Matlab 2017a) if you find the use of it. It consists of a sliding symmetric window with size adaptation near borders to keep the window symmetric (otherwise there can be a delay in your filtered data, or undefined comportement, such as adding more nans or adding fake data (zeros) near borders to use the same window size).

I hope it can help,

Best regards,

Mickaël Tits

Thank you all.

I'll try them.

Hi again,

Here is a version of the filter that deals with a whole mocap track directly, the previous one was supposed to be used individually on each trajectory (or each 3 column).

Cheers,

Mickaël Tits

Dear Mickaël Tits

Thank you again.

But I'm sorry to tell you that I couldn't use your m file because I'm using matlab2014 which doesn't contain move median function. Sorry about that.

Have a nice day,

Rufaida

Dear Rufaida Hussain,

Sorry, I just saw your answer (I recoded the mean and median filter for compatibility, it seems to give the same result).

Try this version then (see attached), I hope it works better for you. I made the assumption that nested functions work in matlab 2014. If not, just put them in separated files.

Mickael

Dear Mickael Tits

Thank you it works well.

Hi,

My gap filling method is under submission for a journal article, but you can already find the code on github:

https://github.com/titsitits/MocapRecovery

Mickaël

Edit: here is the publication: Article Robust and automatic motion-capture data recovery using soft...

hi,

thank you for pointing to our paper in your earlier answer. I'll be looking out for your solution to the problem. :-)

Best regards, Peter

Hi Peter,

I should plan to publish a preprint here then! (either here or arXiv...)

Dear Rufaida, the signal pattern you have shown is quite similar to the pupil diameter when an eye blink occurs. Based on that, a cubic spline interpolation as an imputation technique might be a good approach!