24

I am a newbie in R, love it, but I am surprised by a complete lack of solid package to analyse motion capture data.

The simplest motion capture file is just a massive table with 'XYZ' coordinates for each point attached to a recorded subject, and for every frame captured. I know that I can find individual methods and functions in R to perform complex operations (like principal component analysis) or I can plot time series for all the points. But when I am looking for examples that could also educate me statistically about analysing human movement, and provide with nice toolbox for visual representation of data, R turns out to be a cold desert. On the other hand, MATLAB has Motion capture toolbox and MoCap Toolbox and especially the latter has quite good options for plotting and analysing the captures. But let's be honest - MATLAB has quite ugly visualisation engine comparing to R.

Some specific requests for R motion capture package would include:

  • reading, editing, visualizing and transforming mocap data
  • kinetic and kinematic analysis
  • time-series and principal component analysis
  • animating data

Am I missing something here (in my Googling) or is there really no mocap packages out there for R? Have anyone tried playing with motion capture data in R? Can you give me some directions?

Geek On Acid
  • 6,330
  • 4
  • 44
  • 64
  • 2
    You may not be missing anything. My favo(u)rite solution, `library(sos); findFn("{motion capture}")`, didn't come up with anything useful. There are cultural issues: it may be possible to do cool stuff with R, but if all the cool kids who work on motion capture are using MATLAB or Python, then that's where things will get done. I would definitely take a look and see what's been done in Python, and at interfacing Python with R for any statistical heavy lifting that's not already implemented in R ... – Ben Bolker Nov 16 '11 at 21:22
  • 1
    You can use the packages "forecast" and "ftsa" for timeseries and principal component analysis. – power Nov 25 '11 at 02:45

3 Answers3

4

UPDATE, December 2019: It seems like the mocapr package by Steen Harsted is a much more powerful tool than the one I built. Enjoy.

Have a look at my package, the mocap package: https://github.com/gsimchoni/mocap

It is far from perfect but it's a start, currently tested only on CMU Graphics Lab Motion Capture Database ASF/AMC files.

And here is a blog post with some more details.

Giora Simchoni
  • 3,487
  • 3
  • 34
  • 72
2

I used the package rgl to create an animation from a motion gesture dataset. Although it's not a package made specifically for gesture data, you can work with it.

In the example below, we have gesture data for 8 points on the upper body: spine, shoulder center, head, left shoulder, left wrist, right shoulder, and right wrist. The subject has his hands down and his right arm is making an upward movement.

I restricted the dataset to 6 time observations (seconds, if you will), because otherwise it would get to big to post here.

Each line from the original dataset corresponds to a time observation, and the coordinates of each body point are defined in sets of 4 (every four columns is one body point). So at each line, we have "x", "y", "z", "br" for the spine, then "x", "y", "z", "br" for the shoulder center, and so on. The "br" is always 1, in order to separate the three coordinates (x,y,z) of each body part.

Here is the original (restricted) dataset:

DATA.time.obs<-rbind(c(-0.06431,0.101546,2.990067,1,-0.091378,0.165703,3.029513,1,-0.090019,0.518603,3.022399,1,-0.042211,0.687271,2.987086,1,-0.231384,0.419869,2.953286,1,-0.299824,0.173991,2.882627,1,0.063367,0.399478,3.136306,1,0.134907,0.176191,3.159998,1),
                 c(-0.067185,0.102249,2.990185,1,-0.095083,0.166589,3.028688,1,-0.093098,0.519146,3.019775,1,-0.043808,0.687041,2.987671,1,-0.234622,0.417481,2.94581,1,-0.300324,0.169313,2.869782,1,0.056816,0.398384,3.135578,1,0.134536,0.180875,3.162843,1),
                 c(-0.069282,0.102964,2.989943,1,-0.098594,0.167465,3.027638,1,-0.097184,0.52169,3.019556,1,-0.046626,0.695406,2.989244,1,-0.23478,0.417057,2.943475,1,-0.300101,0.168628,2.860515,1,0.053793,0.395444,3.143226,1,0.134175,0.182816,3.172053,1),
                 c(-0.070924,0.102948,2.989369,1,-0.101156,0.167554,3.026474,1,-0.100244,0.522901,3.018919,1,-0.049834,0.696996,2.987933,1,-0.235301,0.416329,2.939331,1,-0.301339,0.170203,2.85497,1,0.04762,0.390872,3.142792,1,0.14041,0.186844,3.182172,1),
                 c(-0.071973,0.103372,2.988788,1,-0.103215,0.16776,3.025409,1,-0.102334,0.52281,3.019341,1,-0.051298,0.697003,2.991192,1,-0.235497,0.414859,2.935161,1,-0.297678,0.15788,2.833734,1,0.045973,0.386249,3.147609,1,0.14408,0.1916,3.204443,1),
                 c(-0.073223,0.104598,2.988132,1,-0.106597,0.168971,3.022554,1,-0.106778,0.522688,3.015138,1,-0.051867,0.697781,2.990767,1,-0.236137,0.414773,2.931317,1,-0.297552,0.153462,2.827027,1,0.039316,0.39146,3.166831,1,0.175061,0.214336,3.207459,1))

For each time point, we can create a matrix where each row will be a body point, and the columns will be the coordinates:

# Single time point for analysis
time.point<-1
# Number of coordinates
coordinates<-4
# Number of body points
body.points<-dim(DATA.time.obs)[2]/coordinates

# Total time of gesture
total.time<-dim(DATA.time.obs)[1]

# Transform data for a single time. observation into a matrix
DATA.matrix<-matrix(DATA.time.obs[1,],c(body.points,coordinates),byrow = TRUE)
colnames(DATA.matrix)<-c("x","y","z","br")
rownames(DATA.matrix)<-c("hip_center","spine","shoulder_center","head",
                         "left_shoulder","left_wrist","right_shoulder",
                         "right_wrist")

So, we have, at each point of time, a matrix like this:

                        x        y        z br
hip_center      -0.064310 0.101546 2.990067  1
spine           -0.091378 0.165703 3.029513  1
shoulder_center -0.090019 0.518603 3.022399  1
head            -0.042211 0.687271 2.987086  1
left_shoulder   -0.231384 0.419869 2.953286  1
left_wrist      -0.299824 0.173991 2.882627  1
right_shoulder   0.063367 0.399478 3.136306  1
right_wrist      0.134907 0.176191 3.159998  1

And now we use rgl to plot the data from this matrix:

#install.packages("rgl")
library(rgl)

# INITIAL PLOT

x<-unlist(DATA.matrix[,1])
y<-unlist(DATA.matrix[,2])
z<-unlist(DATA.matrix[,3])

# OPEN A BLANK 3D PLOT AND SET INITIAL NEUTRAL VIEWPOINT
open3d()
rgl.viewpoint(userMatrix=rotationMatrix(0,0,0,0))

# SET FIGURE POSITION
# This is variable. It will depend on your dataset
# I've found that for this specific dataset a rotation
# of -0.7*pi on the Y axis works

# You can also plot and select the best view with
# your mouse. This selected view will be passed on
# to the animation.
U <- par3d("userMatrix")
par3d(userMatrix = rotate3d(U, -0.7*pi, 0,1,0))

# PLOT POINTS
points3d(x=x,y=y,z=z,size=6,col="blue")
text3d(x=x,y=y,z=z,texts=1:8,adj=c(-0.1,1.5),cex=0.8)

# You can also plot each body point name.
# This might be helpful when you don't know the
# initial orientation of your plot

# text3d(x=x,y=y,z=z,texts=rownames(DATA.matrix),
#        cex=0.6,adj=c(-0.1,1.5))

# Based on the plotted figure, connect the line segments
CONNECTOR<-c(1,2,2,3,3,4,3,5,3,7,5,6,7,8)
segments3d(x=x[CONNECTOR],y=y[CONNECTOR],z=z[CONNECTOR],col="red")

Then, we have this:

To create an animation, we can put all this into a function and use lapply.

movement.points<-function(DATA,time.point,CONNECTOR,body.points,coordinates){

  DATA.time<-DATA[time.point,]

  DATA.time<-matrix(DATA.time,c(body.points,coordinates),byrow = TRUE)

  x<-unlist(DATA.time[,1])
  y<-unlist(DATA.time[,2])
  z<-unlist(DATA.time[,3])

  # I used next3d instead of open3d because now I want R to plot 
  # several plots on top of our original, creating the animation

  next3d(reuse=FALSE)
  points3d(x=x,y=y,z=z,size=6,col="blue")
  segments3d(x=c(x,x[CONNECTOR]),y=c(y,y[CONNECTOR]),z=c(z,z[CONNECTOR]),col="red")
# You can control the "velocity" of the animation by changing the 
# parameter below. Smaller = faster
  Sys.sleep(0.5)
}

I know this solution is not elegant, but it works.

enter image description here

ogustavo
  • 546
  • 4
  • 12
1

Judging by a quick search on RSeek, there isn't a motion capture package available for R. It looks like you'll need to find equivalents for each function. The more general ones should be fairly easy to find (interpolation, subsetting, transformation/ projection, time-series analysis, pca, matrix analysis etc) and the very process of writing your own custom functions for specific things like estimating instantaneous kinetic energy is probably the best way to learn!

You may find plyr useful for knocking the data into shape and the animation package for visualising motion.

RobinGower
  • 928
  • 6
  • 14