Blender Motion Capture - Walking, Jumping

Teleman

daylanKifky I'm not entirely sure, that's why I think it is a feature in matlab.

Here is a method from TimeSeriesDataBaseClass -- this is when the time array is created:

methods
        function obj = set.SampleRate(obj, sampleRate)
            obj.SampleRate = sampleRate;
            if(obj.SampleRate == 0)
                obj.Time = [];
                obj.TimeAxis = 'Sample';
            elseif(obj.NumPackets ~= 0)
                obj.Time = (0:obj.NumPackets-1)' * (1/obj.SampleRate) + obj.StartTime;
                obj.TimeAxis = 'Time (s)';
            end
        end
        function obj = set.StartTime(obj, startTime)
            obj.StartTime = startTime;
            obj.SampleRate = obj.SampleRate;
        end
        function samplePeriod = get.SamplePeriod(obj)
            if(obj.SampleRate == 0)
                samplePeriod = 0;
            else
                samplePeriod = 1 / obj.SampleRate;
            end
        end
    end

Here's from InertialAndMagneticDataBaseClass:

methods (Access = protected)
        function obj = Import(obj, fileNamePrefix)
            data = obj.ImportCSVnumeric(fileNamePrefix);
            obj.Gyroscope.X = data(:,2);
            obj.Gyroscope.Y = data(:,3);
            obj.Gyroscope.Z = data(:,4);
            obj.Accelerometer.X = data(:,5);
            obj.Accelerometer.Y = data(:,6);
            obj.Accelerometer.Z = data(:,7);
            obj.Magnetometer.X = data(:,8);
            obj.Magnetometer.Y = data(:,9);
            obj.Magnetometer.Z = data(:,10);
            obj.SampleRate = obj.SampleRate;    % call set method to create time vector
        end
    end

The data is imported, then obj.SampleRate creates the time vector -- but I do not see anywhere the calculating of the gyro/acc/mag arrays into the new time vector. (?)

Teleman

The arrays are resized (from 13420 to 11009) and interpolated right after the time array:

% -------------------------------------------------------------------------
% Manually frame data

% startTime = 0;
% stopTime = 10;

indexSel = find(sign(time-startTime)+1, 1) : find(sign(time-stopTime)+1, 1);
time = time(indexSel);
gyrX = gyrX(indexSel, :);
gyrY = gyrY(indexSel, :);
gyrZ = gyrZ(indexSel, :);
accX = accX(indexSel, :);
accY = accY(indexSel, :);
accZ = accZ(indexSel, :);

This is what appears to be a function of matlab.

daylanKifky

Teleman The arrays are resized (from 13420 to 11009) and interpolated right after the time array:

This seems to be discarding the part of the dataset where no action was occurring. So they are just discarding the rows that are outside of the [startTime : stopTime] range

daylanKifky

Breaking apart this line:

 obj.Time = (0:obj.NumPackets-1)' * (1/obj.SampleRate) + obj.StartTime;

obj.NumPackets : Number of rows on dataset
(0:obj.NumPackets-1) : Is using colon notation to create a vector with the values from 0 to the number of rows in dataset minus one
the ' operator transposes the vector.

So the first part of the expresion is just a column vector with the range [0:number of rows in dataset)

-(1/obj.SampleRate) and obj.StartTime are scalars

The complete expression is just creating the column vector and then performing an element-wise
multiplication and addition by scalars.

To create a similar expression in python numpy can be of great help.

on the other hand, these types of expressions:

obj.Gyroscope.X = data(:,2);

are only extracting the columns of the dataset. Indexes in matlab start at 1, so this particular line is extracting the second column of the dataset. The csv contains the gyro X axis at the third column, so it must be truncated somewhere before extracting it

daylanKifky

by the way, you can get a description of a native matlab/octave function with the help command. For example

help find

Teleman

daylanKifky I've pushed the latest changes in the Jupyter file to gitlab, DEV_teleman branch.

I wrote a little function to bisect a column vector -- there may be something in numpy that can do it better, but I am able to calculate a timed column vector based on the sample period from the raw data captured in the CSV file.

There is actually a bug in the matlab script that sets up the initial time column vector. The time range being 4 to 47 seconds, iterated by the sample period. But in the matlab script, the actual column vector cropped from the raw vector starts at 6 seconds.

For example with spiral stairs CSV, row number 675 is for packet 1025, which translates to 4 seconds, which is the first matching row of the time range. However, when the time array is recalculated, it starts with row 1025 (6.07421875 secs) instead of row 675. It doesn't really matter because the walking doesn't start until 10 secs in.

The Velocity plot now looks much closer to the Octave version:

I'll spend some more time analysing the differences in calculations.

Here is the Position plot:

I have not been able to get 3d plot out of the Octave version -- do you have a working version? I'd like to plot out the foot steps next. Cheers.

Teleman

Amazing what a little ' can do ...

I'll check into numpy. Thanks.

daylanKifky

Is looking great! 👏 👏 👏 👏 👏
will take a deeper look tomorrow

daylanKifky

Hi,
I had opportunity of reviewing the last commit. It is indeed looking much better, but there are still some not negligible deviations.
Taking the translational position plot as a reference again: it is now showing a much more plausible result, but there are still some incongruities. For example, in the spiralStairs dataset the subject climbed the stairs and then performed a couple of steps in the upper floor. As you can see, this is reflected in the matlab plot. The Z axis ramps up to ~ 5mts and stays there.

On the python result the Z axis ramps up to ~ 8 mts and then suddenly goes below 0. The other axis are also showing divergent behavior.

It seems like you are almost there, but there are still some minor tweaking to be done. A good sign is that this plot is clearly showing "steps" in most of its extension.

In order to see the last 3D plot: instead of fighting against octave it was easier for me to just export the final results as csv and then use them to animate an object in blender.
I pushed a commit which:

exports a csv from the matlab and python scripts, and omits the final plots.
includes an stationary only plot in the matlab script, and toggles to turn each plot on or off.
includes a script to apply the csv to an object in blender.

In order to use this last script select the object to which apply the animation (normally an Empty) and hit [Run script].
You can select which csv to apply by uncommenting the corresponding line at the beginning of the script. In order for it to find the csv you have to save the .blend file in root folder of the repo.

(I included a *.blend line in .gitignore so those files won't be committed)

Let me know how it goes. I haven't verified it, but I suspect that the detection of the stationary periods might be responsible for the deviations. Perhaps plotting a slice of it might help you compare the results better.

highmatergogo

Great work @Teleman and @daylanKifky (as always). This was a deal breaker for our project which is now on hold (also due to covid-19). We had a look at implementing translation ourselves but time and skillset avaded us.

I'm thrilled to see that this is being worked on.

italogsfernandes

Teleman Hey man, are you referring to my github repo? maybe can i add something up, there is a while since i last worked on it and the code is now a big mess haha

daylanKifky

Hi italogsfernandes, this thread was inactive for a long time. We were working on translating Madgwick's translation algorithm to python in order to make a first implementation for our motion capture system.
Not sure to which one of your repos @Teleman is refering to, but I see you work with inertial sensors. We are always open to contributions 😉

« Previous Page