Fig 1.
All sensors travel the same distance in the ground-plane over the duration of one stride. Therefore, we expect the calculated stride length to be identical. However, the frequency response of the attachment modulates the recorded signal in a way, that for some sensor positions large integration errors become more likely. We demonstrated this effect with six different sensor positions on a dataset with 9000 strides.
Table 1.
Participant demographic.
Fig 2.
Sensor placement of the six shoe sensors.
The insole sensor (center) was embedded in a purpose-built insole. The cavity sensor was placed in a cavity in the sole of the custom shoe (right). All other sensors were attached with 3D-printed clips.
Fig 3.
Visual representation of the two-step alignment procedure.
The image shows the alignment of the instep sensor, but the same procedure is applied to all sensors. First, the local coordinate system is rotated, so that the z-axis aligns with the global direction of gravity during the foot-flat posture. In the second step, the rotation around the new z-axis is aligned with the cavity sensor by calculating the difference in phase between the angular velocity vectors ωxy in the x-y-plane. From the regions where the angular velocity is larger than 150 deg/s (gray areas in the plot) the median angle between the two sensors Δφ is extracted to fully align the signals of the sensors.
Fig 4.
The average signal over all strides of the left foot from the 2×20m walk test at normal speed of the participant 4d91. All strides were interpolated to 200 samples before averaging. The individual signals are rotated to match the primary anatomical axis during mid-stance. The start and the end of the stride conform with the labeled-stride definition. The vertical lines indicate the approximate position of the detected gait events.
Table 2.
Overview over the stride categories and the respective number of strides.
Fig 5.
The stride length error of each sensor position over all strides.
The image on the left shows all straight strides. The image on the right has all outliers (> Q3 + 1.5IQR and < Q1 − 1.5IQR) removed for visual clarity. Values larger than 0 indicate an overestimation by the IMU. The black triangles mark the means of the distributions. They are omitted in the left image for visual clarity.
Fig 6.
The absolute stride length error of each sensor position.
On the left the error is calculated over all strides and right for the individual speed categories. All outliers (> Q3 + 1.5IQR) are removed from the visualization. The black triangles mark the means of the distributions.
Fig 7.
The raw data differences for all sensor axes.
The 3D Difference (top) and the Norm Difference (bottom) over all strides for the accelerometer signal (left) and the gyroscope signal (right). The annotations provide the mean and the IQR over all strides. Note that the color for the Norm Difference is independent of the sign of the value.
Fig 8.
The STD over the different sensors per sample of a stride.
The curves are the averages over all strides from the 2×20m walk test at normal speed. The STD of each stride was interpolated to 200 samples before averaging. The individual signals are rotated to match the primary anatomical axis during mid-stance. The approximate regions around the IC and TC are marked in gray.
Fig 9.
The raw data features over all speeds.
The residual energy during the mid-stance is shown on the left and the peak acceleration amax during the gait cycle is shown on the right. Both boxplots contain all strides of each of the self-selected speed categories. Note the different y-scales for the individual plots. The black triangles mark the means of the distributions. All outliers (> Q3 + 1.5IQR) are removed from the visualization.
Fig 10.
The average power spectral density for the high and low frequency bins.
The average spectral power of the accelerometer (bottom row) and the gyroscope norm (top row) in the frequency ranges 020 Hz (left column) and 80102.4 Hz (right column). Each datapoint used in the boxplots corresponds to a single stride. The black triangles mark the means of the distributions. All outliers (> Q3 + 1.5IQR) are removed from the visualization.
Fig 11.
The correlation between and the absolute stride length error.
On the left the correlation on a gait test level is shown. For that the median PSD and stride length error for each of the six straight walk tests are calculated per participant and sensor. On the right the correlation on a stride level is shown for the same tests. The histograms in the margins show the relative distribution of values across all sensors along the respective axis.