An elaborate data set on human gait and the effect of mechanical perturbations

Participants

Fifteen able bodied subjects including four females and eleven males with an average age of 24 ± 4 years, height of 1.75 ± 0.09 m, mass of 74 ± 13 kg participated in the study. The study was approved by the Institutional Review Board of Cleveland State University (# 29904-VAN-HS) and written informed consent was obtained from all participants. The data has been anonymized with respect to the participants’ identities and a unique identification number was assigned to each subject. A selection of the meta data collected for each subject is shown in Table 1.

Equipment

The data were collected in the Laboratory for Human Motion and Control at Cleveland State University, using the following equipment:

• A R-Mill treadmill which has dual 6 degree of freedom force plates, independent belts for each foot, along with lateral translation and pitch rotation capabilities (Forcelink, Culemborg, Netherlands).

• A 10 Osprey camera motion capture system paired with the Cortex 3.1.1.1290 software (Motion Analysis, Santa Rosa, CA, USA).

• USB-6255 data acquisition unit (National Instruments, Austin, Texas, USA).

• Four ADXL330 Triple Axis Accelerometer Breakout boards attached to the treadmill (Sparkfun, Niwot, Colorado, USA).

• D-Flow software (versions 3.16.1 to 3.16.2) and visual display system, (Motek Medical, Amsterdam, Netherlands).

The Cortex software delivers high accuracy 3D marker trajectories from the cameras along with data from the force plates and analog sensors (e.g., EMG/Accelerometer) through a National Instruments USB-6255 data acquisition unit. D-Flow then receives streaming data from Cortex and any other digital sensors. It is also responsible for controlling the treadmill’s motion (lateral, pitch, belts). D-Flow can process the data in real time and/or export data to file.

Our motion capture system’s coordinate system is such that the X coordinate points to the right, the Y coordinate points upwards, and the Z coordinate follows from the right-hand-rule, i.e., points backwards with respect to the walking direction. The camera’s coordinate system is aligned to an origin point on the treadmill’s surface during camera calibration. The same point is used as the origin of the ground reaction force measuring system. Figure 1 shows the layout of the equipment.

Early on, we discovered that the factory setup of the R-Link treadmill had a vibration mode as low as 5Hz that was detectable in the force measurements; this was likely due to the flexible undercarriage and pitch motion mechanism. Trials 6–8 are affected by this vibration mode. During trials 9–15 the treadmill was stabilized with wooden blocks. During the remaining trials (>15) the treadmill was stabilized with metal supports; both with ones we fabricated in-house and ones supplied by the vendor. These supports aimed to improve the stiffness and frequency response of the force plate system. See the Data Limitations Section for more details.

The acceleration of the treadmill base was measured during each trial by the ADXL330 accelerometers placed at the four corners of the machine. These accelerometers were intended to provide information for inertial compensation purposes when the treadmill moved laterally or in pitch, but are extraneous for trials greater than number 8 due to the treadmill being stabilized in those degrees of freedom by the aforementioned supports.

Protocol

The experimental protocol consisted of both static measurements and walking on the treadmill for 10 min under unperturbed and perturbed conditions. Before a set of trials on the same day, the motion capture system was calibrated using the manufacturer’s recommended procedure. Before each subject’s gait data were collected, the subject changed into athletic shoes, shorts, a sports bra, a baseball cap,¹ and a rock climbing harness. All 47 markers were applied directly to the skin at the landmarks noted in Table 2 except for the heel, toe, and head markers, which were placed on the respective article of clothing.² Then the subject self-reported their age, gender, and mass. Finally, their height was measured by the experimentalist and four reference photographs (front, back, right, left) were taken of subject’s marker locations.

After obtaining informed consent and a briefing by the experimentalist on the trial protocol, the subject followed the verbal instructions of the experimentalist and the on-screen instructions from the video display. The final protocol for a single trial was as follows:

1. The subject stepped onto the treadmill and markers were identified with Cortex.

2. The safety rope was attached loosely to the rock climbing harness such that no forces were acting on the subject during walking, but so that the harness would prevent a full fall.

3. The subject started by stepping on sides of treadmill so that feet did not touch the force plates and the force plate signals are zeroed. This corresponds to the “Force Plate Zeroing” event.

4. Once notified by the video display, the subject stood in the calibration pose: standing straight up, looking forward, arms out by their sides (approximately 45 degree abduction) and the event, “Calibration Pose,” was manually recorded by the operator.

5. A countdown to the first normal walking phase was displayed. At the end of the countdown the event “First Normal Walking” was recorded and the treadmill ramped up to the specified speed and the subject was instructed to walk normally, to focus on the “endless” road on the display, and not to look at their feet.

6. After 1 min of normal walking, the longitudinal perturbation phase begun and was recorded as “Longitudinal Perturbation.”

7. After 8 min of walking under the influence of the perturbations, the second normal walking phase begun and was recorded as “Second Normal Walking.”

8. After 1 min of normal walking, a countdown was shown on the display and the treadmill decelerated to a stop.

9. The subject was instructed to step off of the force plates for 10 s and the “Unloaded End” event was recorded.

10. The subject could then take a rest break before each additional trial.

Perturbation signals

As previously described, the protocol included a phase of normal walking, followed by longitudinal belt speed perturbations, and ended with a second segment of normal walking. Three pseudo-random belt speed control signals, with mean velocities of 0.8 m s^-1, 1.2 m s^-1 and 1.6 m s^-1, were pre-generated with MATLAB and Simulink (Mathworks, Natick, Massachusetts, USA) and are available for download from Zenodo (Hnat, Moore & Van den Bogert, 2015). The same control signal was used for all trials at that given speed.

To create the signals, we started by generating random acceleration signals, sampled at 100 Hz, using the Simulink discrete-time Gaussian white noise block followed by a saturation block set at the maximum belt acceleration of 15 m s^-2. The signal was then integrated to obtain belt speed and high-pass filtered with a second-order Butterworth filter to eliminate drift. One of the three mean speeds were then added to the signal and limited between 0 m s^-1 to 3.6 m s^-1. The cutoff frequencies of the high-pass filter, as well as the variance in the acceleration signal, were manually adjusted until acceptable standard deviations for each mean speed were obtained: 0.06 m s^-1, 0.12 m s^-1 and 0.21 m s^-1 for the three speeds, respectively. These ensured that the test subjects were sufficiently perturbed at each speed, while remaining within the limits of our equipment and testing protocol. To ensure that the treadmill belts could accelerate to the desired values, the high performance mode in the D-Flow software was enabled. The MATLAB script and Simulink model produce a comma-delimited text file of with the desired belt speed signals indexed by the time stamp.

Figure 2 gives an idea of the effect of the treadmill and controller dynamics by plotting the measured speed of the treadmill belts from loaded trials (76, 77, 78) against the commanded treadmill control input signal. The system introduces a delay and seems to act as a low pass filter. The standard deviations of the measured speeds do not significantly differ from those of the commanded speeds: 0.05 m s^-1, 0.12 m s^-1 and 0.2 m s^-1 for the three speeds, respectively.

Figure 3 gives a frequency domain view of the effects of the treadmill dynamics. These spectral density plots were created by averaging a spectrogram of a twenty second Hamming window. For all speeds, the frequency content of the commanded and measured time series show similarity below 4 Hz and attenuation in the measured spectral density above 4 Hz.

When belt speed is not constant, the inertia of the rollers and motor will likely induce error in the force plate x axis moment, and hence, the anterior-posterior coordinate (z axis) of the center of pressure that is measured by the instrumentation in the treadmill. This error may or may not be pertinent to different analyses. If needed, this error can be partially compensated by a linear model as shown in Hnat & Van den Bogert (2014). The model coefficients can be identified from the unloaded trials given in Table 3. The error due to inertia is random and does not affect the averaged joint moments presented in Fig. 5. Compensation should, however, be done if joint moments from individual gait cycles are of interest rather than the ensemble average.

In addition to the longitudinal perturbations, lateral perturbations were also prescribed for a duration of four minutes in the pilot trials 3–8. Figure 4 shows an example of the measured lateral deviation of the treadmill base. These signals were generated in a similar manner using MATLAB and Simulink in which a Gaussian white noise block was twice integrated to obtain the lateral deviation. The signal was then high-pass filtered with a second-order Butterworth filter to eliminate drift and then saturated so that the signal remained within the 5 cm lateral range of the physical hardware. The same perturbation signal was used for each of the three trials.

Raw data

The raw data consists of a set of ASCII tab delimited text files output from both the “mocap” and “record” modules in D-Flow in addition to a manually generated YAML³ file that contains all of the necessary meta data for the given trial. These three files are stored in a hierarchy of directories with one trial per directory. The directories are named in the following fashion T001/ where T stands for “trial” and the following three digits provide a unique trial identification number.

Processed data

We developed a toolkit for data processing, GaitAnalysisToolKit v0.1.2 (Moore et al., 2014) for common gait computations and provide an example processed trial to present the nature of the data. The tool was developed in Python, is dependent on the SciPy Stack [NumPy (Walt, Colbert & Varoquaux, 2011), SciPy (Jones et al., 2001), matplotlib (Hunter, 2007), Pandas (McKinney, 2010), etc] and Octave (Octave community, 2014), and provides two main classes: one to do basic gait data cleaning from D-Flow’s output files, DFlowData, and a second to compute common gait variables of interest, GaitData.

The DFlowData class collects and stores all the raw data presented in the previous section and applies several “cleaning” operations to transform the data into a usable form. The cleaning process follows these steps:

1. Load the meta data file into a Python dictionary.

2. Load the D-Flow mocap module TSV file into Pandas DataFrame.

3. Relabel the column headers to more meaningful names if this is specified in the meta data.

4. Optionally identify the missing values in the mocap marker data and replace them with numpy.nan.

5. Optionally interpolate the missing marker values and replaces them with interpolated estimates using a variety of interpolation methods.

6. Load the D-Flow record module TSV file into a Pandas DataFrame.

7. Extract the events and create a dictionary mapping the event names in the meta data to the events detected in the record module file.

8. Inertially compensate the ground reaction loads based on whether the meta data indicates there was treadmill motion.

9. Merge the data from the mocap module and record module into one data frame at the maximum common constant sample rate.

Once the data is cleaned there are two methods that allow the user to extract the cleaned data: either extract sections of the data bounded by the events recorded in the record-xxx.txt file or save the cleaned data to disk. These operations are available as a command line application and as an application programming interface (API) in Python. An example of the DFlowData API in use is provided in Listing 2.

The GaitData class is then used to compute gait events (toe off and heel strike times), basic 2D inverse kinematics and dynamics, and to store the data into a Pandas Panel with each gait cycle on the item axis at a specified sampling rate. This object can also be serialized to disk in HDF5 format. An example of using the Python API is shown in Listing 3.

A similar work flow was used to produce Fig. 5 which compares the mean and standard deviation of sagittal plane joint angles and torques from the perturbed gait cycles to the unperturbed gait cycles computed from trial 20. This gives an idea of the more highly variable dynamics required to walk while being longitudinally perturbed.

For more insight into the difference in the unperturbed and perturbed data, Fig. 6 compares the distribution of a few gait cycle statistics. One can see that the perturbed strides have a much larger variation in frequency and length and even larger variation in stride width. It is also interesting to note that the coupled nature of the system’s degrees of freedom can be exploited to increase the stride width with only longitudinal perturbations, although not relatively as much as is in the other statistics.

Data limitations

The data is provided in good faith with great attention to detail but as with all data there are anomalies that may affect the use and interpretation of results emanating from the data. The following lists give various notes and warnings about the data that should be taken into account before use.