Basics of data import and cleaning in pathviewr

Vikram B. Baliga



Raw movement data, including those from motion capture systems, may have a variety of issues. These raw data often contain noise or artifacts from the recording session, which may not be easily removed via the recording software itself. Data may not be organized as “tidy” key-value pairs (making plotting more difficult), the axes and overall orientation of the environment may not conform to a standard, and individual movement trajectories may be ill-defined.

pathviewr provides functions in R to deal with such problems (i.e. “cleaning”). This vignette will cover the basics of how to import raw data and how to clean data to prepare it for visualization and/or statistical analyses.

What do movement data sets look like?

At minimum, movement data provide information on a subject or object’s position over time. These data are typically supplied in three dimensions (e.g. x, y, z), with position in each dimension sampled at a particular rate (e.g. 100 Hz). Different recording software may provide additional features, such as the ability to track multiple subjects simultaneously, information on subjects’ rotation, tracking of “rigid body” elements, or even the ability to apply Kalman filters.

A central goal of pathviewr is to take data from different sources (so far: Motive and Flydra), re-organize them into a common format that can be wrangled in R, clean them up a bit, and get them ready for visualization and/or statistical analyses. We’ll first cover what’s included in Motive and in Flydra data and how pathviewr handles these. Should you have data from another source, our as_viewr() function will allow you to bring it into the pathviewr framework.

Data import via pathviewr

Data can be imported via one of three functions:

We will showcase examples from each of these methods in this section.

Please feel free to reach out to the pathviewr authors via our Github Issues page should you have trouble with any of our data import options. We are happy to work with you to design custom read_ functions for file types we have not encountered ourselves.

We’ll start by loading pathviewr and a few of the packages in the tidyverse.

## If you do not already have pathviewr installed:
# install.packages("devtools")
# devtools::install_github("ropensci/pathviewr")


Motive CSV files

.csv files exported from Motive can be imported via read_motive_csv()

## Import the Motive example data included in 
## the package

motive_data <-
    system.file("extdata", "pathviewr_motive_example_data.csv",
                package = 'pathviewr')

## This produces a tibble 
#> # A tibble: 934 x 26
#>    frame time_sec device02_rotation_x device02_rotation_y device02_rotation_z
#>    <int>    <dbl>               <dbl>               <dbl>               <dbl>
#>  1 72210     722.              0.135               -0.977             -0.112 
#>  2 72211     722.              0.0819              -0.978             -0.0991
#>  3 72212     722.              0.211               -0.973             -0.0939
#>  4 72213     722.              0.196               -0.972             -0.128 
#>  5 72214     722.              0.131               -0.975             -0.121 
#>  6 72215     722.              0.0935              -0.975             -0.105 
#>  7 72216     722.              0.180               -0.975             -0.106 
#>  8 72217     722.              0.164               -0.972             -0.149 
#>  9 72218     722.              0.120               -0.973             -0.149 
#> 10 72219     722.              0.155               -0.970             -0.120 
#> # ... with 924 more rows, and 21 more variables: device02_rotation_w <dbl>,
#> #   device02_position_x <dbl>, device02_position_y <dbl>,
#> #   device02_position_z <dbl>, device02_mean_marker_error <dbl>,
#> #   device03_rotation_x <dbl>, device03_rotation_y <dbl>,
#> #   device03_rotation_z <dbl>, device03_rotation_w <dbl>,
#> #   device03_position_x <dbl>, device03_position_y <dbl>,
#> #   device03_position_z <dbl>, device03_mean_marker_error <dbl>,
#> #   device05_rotation_x <dbl>, device05_rotation_y <dbl>,
#> #   device05_rotation_z <dbl>, device05_rotation_w <dbl>,
#> #   device05_position_x <dbl>, device05_position_y <dbl>,
#> #   device05_position_z <dbl>, device05_mean_marker_error <dbl>

A key thing to note is that these data, as stored in Motive CSVs, are not “tidy”. Each frame occupies one row, but what that also means is that the rotation and position values for the various subjects take up 24 columns! This format not only makes plotting data more difficult in base R, ggplot2, and rgl, but also makes other aspects of data wrangling more difficult. In a later step, we will ‘gather’ these data into key-value pairs so that e.g. all length-wise position values are in one column, all width-wise are in another…etc.

Metadata are stored as attributes. We won’t go through all of these, but here are a couple important ones.

## E.g. to see the header of the original file:
attr(motive_data, "header")
#>                 metadata                      value
#> 1         Format Version                       1.23
#> 2              Take Name  sept-18_mixed-group_16-30
#> 3             Take Notes                           
#> 4     Capture Frame Rate                 100.000000
#> 5      Export Frame Rate                 100.000000
#> 6     Capture Start Time 2019-09-18 PM
#> 7   Total Frames in Take                     190951
#> 8  Total Exported Frames                     190951
#> 9          Rotation Type                 Quaternion
#> 10          Length Units                     Meters
#> 11      Coordinate Space                     Global

## Names of all marked objects:
attr(motive_data, "subject_names_simple")
#> [1] "device02" "device03" "device05"

## Types of data included
attr(motive_data, "data_types_simple")
#> [1] "Rigid Body"

## Frame rate
attr(motive_data, "frame_rate")
#> [1] 100

Storing such metatdata in the attributes is a key feature of pathviewr. These metadata may not be as immediately as important as the time series of position or rotation, but they can provide important experimental information such as the date & time of capture and the units of the position data (here, meters).

Flydra Matlab files

.mat files exported from Flydra can be imported via read_flydra_mat().

Note that you must supply a subject_name for Flydra data, as subject names are not embedded in the .mat files. Only one name can be added and it will be used throughout the resultant tibble.

## Import the Flydra example data included in 
## the package
flydra_data <-
                package = 'pathviewr'),
    subject_name = "birdie_wooster"

## Similarly, this produces a tibble with important 
## metadata as attributes
#> # A tibble: 7,744 x 6
#>    frame time_sec subject        position_length position_width position_height
#>    <dbl>    <dbl> <chr>                    <dbl>          <dbl>           <dbl>
#>  1   746     0    birdie_wooster           0.869       -0.00417            1.31
#>  2   747     0.01 birdie_wooster           0.864       -0.00614            1.31
#>  3   748     0.02 birdie_wooster           0.863       -0.00695            1.31
#>  4   749     0.03 birdie_wooster           0.862       -0.00672            1.31
#>  5   750     0.04 birdie_wooster           0.862       -0.00644            1.31
#>  6   751     0.05 birdie_wooster           0.862       -0.00619            1.31
#>  7   752     0.06 birdie_wooster           0.863       -0.00667            1.31
#>  8   753     0.07 birdie_wooster           0.864       -0.00712            1.31
#>  9   754     0.08 birdie_wooster           0.865       -0.00727            1.31
#> 10   755     0.09 birdie_wooster           0.865       -0.00760            1.31
#> # ... with 7,734 more rows

attr(flydra_data, "frame_rate")
#> [1] 100

Note that unlike the example Motive data, the Flydra data are already organized into key-value pairs. Because rotation is not captured by Flydra, such data are also not included.

Data from other sources

Data from another format can be converted to a viewr object via pathviewr::as_viewr(). Although this function does not handle data import per se, it allows data that you may already have imported into R as a tibble or data.frame to then be reformatted for use with pathviewr functions.

We’ll run through a quick example with simulated data:

## Create a dummy data frame with simulated (nonsense) data
df <-
    frame = seq(1, 100, by = 1),
    time_sec = seq(0, by = 0.01, length.out = 100),
    subject = "birdie_sanders",
    z = rnorm(100),
    x = rnorm(100),
    y = rnorm(100)

## Use as_viewr() to convert it into a viewr object
test <-
    frame_rate = 100,
    frame_col = 1,
    time_col = 2,
    subject_col = 3,
    position_length_col = 5,
    position_width_col = 6,
    position_height_col = 4

## Some metadata are stored as attributes
attr(test, "frame_rate")
#> [1] 100

We also welcome you to request custom data import functions, especially if as_viewr() does not fit your needs. We are interested in expanding our data import functions to accommodate additional file types. Please feel free to file a request for additional import functions via our Github Issues page.

Data cleaning

As noted above, raw data often suffer the following:
- contain noise or artifacts from the recording session
- not organized as “tidy” key-value pairs
- axes and overall orientation of the environment may not conform to a standard
- individual movement trajectories may be ill-defined

Several functions to clean and wrangle data are available, and we have a suggested pipeline for how these steps should be handled. The rest of this vignette will cover these steps.

All of the steps in the suggested pipeline are also covered by two all-in-one functions: clean_viewr() and import_and_clean_viewr(). See the section at the very end of this vignette for details.

And speaking of pipes, all functions in pathviewr are pipe-friendly. We will detail each step separately, but each of the subsequent steps may be piped, e.g. data %>% relabel_viewr_axes() %>% gather_tunnel_data() etc etc

Relabeling axes, gathering data columns, and trimming outliers

Axis labels (x, y, z) may be applied in arbitrary ways by software. A user might intuitively think the z axis represents height, but the original software may label it as the y axis instead.

relabel_viewr_axes provides a means to relabel axes with “tunnel_length”, “tunnel_width”, and “tunnel_height”. These axis labels will be expected by subsequent functions, so skipping this step is ill-advised.

Typically, axes from Motive data will need to be relabled, but axes in data imported from Flydra will not.

motive_relabeled <-
  motive_data %>%
    tunnel_length = "_z",
    tunnel_width = "_x",
    tunnel_height = "_y",
    real = "_w"

#>  [1] "frame"                      "time_sec"                  
#>  [3] "device02_rotation_width"    "device02_rotation_height"  
#>  [5] "device02_rotation_length"   "device02_rotation_real"    
#>  [7] "device02_position_width"    "device02_position_height"  
#>  [9] "device02_position_length"   "device02_mean_marker_error"
#> [11] "device03_rotation_width"    "device03_rotation_height"  
#> [13] "device03_rotation_length"   "device03_rotation_real"    
#> [15] "device03_position_width"    "device03_position_height"  
#> [17] "device03_position_length"   "device03_mean_marker_error"
#> [19] "device05_rotation_width"    "device05_rotation_height"  
#> [21] "device05_rotation_length"   "device05_rotation_real"    
#> [23] "device05_position_width"    "device05_position_height"  
#> [25] "device05_position_length"   "device05_mean_marker_error"

Akin to the behavior of dplyr::gather(), gather_tunnel_data() will take all data from a given session and organize it so that all data of a given type are within one column, i.e. all position lengths are in position_length, as opposed to separate length columns for each rigid body. These column names will be expected by subsequent functions, so skipping this step is also ill-advised if you are using data from Motive. Should you have data from Flydra, this step should be skipable.

Use trim_tunnel_outliers() to remove extreme artifacts and other outlier data. What this function does is create a (virtual) boundary box according to user-specification, and any data outside that boundary are removed. For example, if you know your arena measures 10m x 10m x 10m and your data were calibrated to range from 0-10m in each dimension, you can be reasonably sure that extreme values such as 45m on a given axis are bogus. This step is entirely optional, and should only be used when the user is confident that data outside certain ranges are artifacts or other bugs. Data outside these ranges are then filtered out. Best to plot data beforehand and check!!

## First gather and show the new column names
motive_gathered <-
  motive_relabeled %>%

#>  [1] "frame"             "time_sec"          "subject"          
#>  [4] "position_length"   "position_width"    "position_height"  
#>  [7] "rotation_length"   "rotation_width"    "rotation_height"  
#> [10] "rotation_real"     "mean_marker_error"

## Now trim, using ranges we know to safely include data
## and exclude artifacts. Anything outside these ranges 
## will be removed.
motive_trimmed <-
  motive_gathered %>%
    lengths_min = 0,
    lengths_max = 3,
    widths_min = -0.4,
    widths_max = 0.8,
    heights_min = -0.2,
    heights_max = 0.5

Standardization of tunnel position and coordinates

The coordinate system of the tunnel itself may require adjustment or standardization. For example, data collected across different days may show slight differences in coordinate systems if calibration equipment was not used in identical ways. Moreover, the user may want to redefine how the coordinate system itself is defined (i.e. change the location of (0, 0, 0) to another place within the tunnel.

Note that having (0, 0, 0) set to the center of the region of interest (covered in the next section of this vignette) is required for all subsequent pathviewr functions to work.

pathviewr offers three main choices for such standardization:

Two quick examples will follow, using our Motive and Flydra data:

## Rotate and center the motive data set:
motive_rotated <-
  motive_trimmed %>% 
    perch1_len_min = -0.06,
    perch1_len_max = 0.06,
    perch2_len_min = 2.48,
    perch2_len_max = 2.6,
    perch1_wid_min = 0.09,
    perch1_wid_max = 0.31,
    perch2_wid_min = 0.13,
    perch2_wid_max = 0.35

In the above, virtual perches are defined by the user using the arguments shown. The center of each perch is then found and then the locations of the two perch centers are then used to 1) set (0, 0, 0) to the point that is equidistant from the perches (i.e. the middle of the tunnel) and 2) rotate the tunnel about the height axis so that both perch width coordinates are at 0. This may be easier to understand through plotting:

## Quick (base-R) plot of the original data
     asp = 1)

## Quick (base-R) plot of the rotated data
     asp = 1)

Differences due to rotation may be extremely subtle, but the redefining of (0, 0, 0) to the middle of the tunnel should be clear from contrasting the axes of the plots.

Flydra data typically do not need to be rotated, so we will instead use redefine_tunnel_center() to adjust the location of (0, 0, 0):

## Re-center the Flydra data set:
flydra_centered <-
  flydra_data %>%
  redefine_tunnel_center(length_method = "middle",
                         height_method = "user-defined",
                         height_zero = 1.44)

Here, we are using length_method = "middle" to use the middle of the range of “length” data to set length = 0 (a translation), making no change to the width axis, and then specifying that height = 0 should be equal to the value 1.44 from the original data (another translation). Again, plotting may help; note that this time, we’ll plot length x height (instead of width):

## Quick (base-R) plot of the original data
     asp = 1)

## Quick (base-R) plot of the redefined data
     asp = 1)

Selecting a region of interest

This required step has benefits that are twofold: 1) treatment effects on animal movement may only manifest over certain regions of the tunnel, and 2) focusing on a subset of the data makes it easier to define explicit trajectories and run computations faster.

The region of interest is defined via the function select_x_percent(). Once tunnel coordinates have been standardized (via one of the function in the previous section), select_x_percent() then selects the middle x percent (along the length axis) of the tunnel as the region of interest. For example, selecting 50 percent would start from the center of the tunnel and move 25% of the tunnel along positive length and 25% along negative length values to then select the middle 50% of the tunnel.

Quick examples:

## Motive data: select the middle 50% of the tunnel as the region of interest
motive_selected <-
  motive_rotated %>% select_x_percent(50)

## Quick plot:
     asp = 1)

## Flydra data:
flydra_selected <-
  flydra_centered %>% select_x_percent(50)

## Quick plot:
     asp = 1)

Isolating each trajectory

The pathviewr standard for defining a trajectory is: continuous movement from one side of the tunnel to the other over the span of the region of interest. Note that this definition does not strictly require linear movement from one end to the other; an animal could make several loops inside the region of interest within a given trajectory.

Isolating trajectories is handled via the separate_trajectories() function in pathviewr. Note that a region of interest must be selected beforehand via select_x_percent().

Because cameras may occasionally drop frames, we allow the user to permit some relaxation of how stringent the “continuous movement” criterion is. This is handled via the max_frame_gap argument within separate_trajectories(). For more details, please see the vignette Managing frame gaps with pathviewr.

In our Motive example, we’ll use the automated feature built into the function to guesstimate the best max_frame_gap allowed. When frame gaps larger than max_frame_gap are encountered, the function will force the defining of a new trajectory. But if frame gaps smaller than max_frame_gap are encountered, keeping observations within the same trajectory is permitted.

In the Flydra example, we’ll simply set max_frame_gap to 1 so that no frame gaps are allowed (movement must be continuous with no dropped frames).

motive_labeled <-
  motive_selected %>% 
  separate_trajectories(max_frame_gap = "autodetect")
#> autodetect is an experimental feature -- please report issues.

flydra_labeled <-
  flydra_selected %>% 
  separate_trajectories(max_frame_gap = 1)

Retain only complete trajectories

Now that trajectories have been isolated and labeled, the final cleaning step is to retain only the trajectories that completely span from one end of the region of interest to the other.

This final step is handled via pathviewr’s get_full_trajectories().

There is a built-in “fuzziness” feature: because trajectories may not have observations exactly at the beginning or the end of the region of interest, it may be necessary to allow trajectories to be slightly shorter than the range of the selected region of interest. The span parameter of this function handles this. By supplying a numeric proportion from 0 to 1, a user may allow trajectories to span that proportion of the selected region. For example, setting span = 0.95 will keep all trajectories that span 95% of the length of the selected region of interest. Setting span = 1 (not recommended) will strictly keep trajectories that start and end at the exact cut-offs of the selected region of interest.

For these reasons, spans of 0.99 to 0.95 are generally recommended. The best choice ultimately depends on your capture frame rate as well as your own judgment. Should you desire to set it lower (which you can), you may instead consider using a smaller region of interest (i.e. set the desired_percent parameter in select_x_percent() to be lower).

## Motive
motive_full <-
  motive_labeled %>% 
  get_full_trajectories(span = 0.95)

     asp = 1, col = as.factor(motive_full$file_sub_traj))

## Flydra
flydra_full <-
  flydra_labeled %>% 
  get_full_trajectories(span = 0.95)

     asp = 1, col = as.factor(flydra_full$file_sub_traj))

All-in-one cleaning functions

All of the above steps can also be done by using pathviewr’s designated all-in-one functions. import_and_clean_viewr() imports raw data and allows the user to run through all of the cleaning steps previously covered in this vignette. clean_viewr() does the same on any object already imported into the R environment (i.e. it skips data import).

For both of these functions, all of the cleaning steps are set to TRUE by default, but may be turned off by using FALSE. Argument names correspond to standalone functions in pathviewr, and if the user wants to use non-default values for corresponding arguments, they should also be supplied for any steps that are set to TRUE.

For example, if the user keeps select_x_percent = TRUE as an argument in clean_viewr(), the select_x_percent() function is run internally. This means that should the user desire to select a region of interest that does not match the default value of 33 percent, an additional argument should be supplied to clean_viewr() as if it were being supplied to select_x_percent() itself, e.g.: desired_percent = 50.

All additional arguments should be written out fully and explicitly.

Here’s an example using what we previously covered:

motive_allinone <-
  motive_data %>%
    relabel_viewr_axes = TRUE,
    gather_tunnel_data = TRUE,
    trim_tunnel_outliers = TRUE,
    standardization_option = "rotate_tunnel",
    select_x_percent = TRUE,
    desired_percent = 50,
    rename_viewr_characters = FALSE,
    separate_trajectories = TRUE,
    max_frame_gap = "autodetect",
    get_full_trajectories = TRUE,
    span = 0.95
#> autodetect is an experimental feature -- please report issues.

## Quick plot
     asp = 1, col = as.factor(motive_allinone$file_sub_traj))

That’s all!