![Figure 1 shows landing and takeoff timing errors for double leg jumps. 766 events were captured. For takeoff timing error Q1 was -2 ms, the median was 0ms, Q3 was 2 ms, (range [-8ms, 8ms]). For landing timing error Q1 was -2ms, the median was 0ms, Q3 was 2ms, (range [- 8ms, 8ms]).](https://s3-us-west-2.amazonaws.com/secure.notion-static.com/65a4787c-ebb5-4bcb-a1b3-df68c655d512/1.png)
The purpose of this investigation was to determine the accuracy of the improved Ground Interaction algorithm, which is used to classify when the foot is on and off the ground.
23 recreationally active subjects volunteered to perform 11 activities on a dual force plate system while wearing insoles containing Inertial Measurement Units (IMUs) from Plantiga Technologies. These IMUs had a sampling frequency of 416 Hz. Insole data was analyzed to determine the off/on-ground location of each foot and find exact takeoff and landing times.
The data were separated into 4 groups of activities based on their similarities (single-leg jumps, double-leg jumps, walking/running, in-place activities). Data analysis showed the following results: 1) For single-leg jumps, results showed median errors of 0ms (IQR: 16ms) and 0ms (IQR: 16ms) for takeoff and landing, respectively (N=309). 2) For double-leg jumps, results showed median errors of 0ms (IQR: 16ms) and 0ms (IQR: 16ms) for takeoff and landing, respectively (N=383). 3) For walking and running, results showed median errors of 0ms (IQR: 16ms) and 0ms (IQR: 16ms) for toe-off and heel strikes, respectively (N=1628). 4) For in-place activities, results showed median errors of 0ms (IQR: 16ms) and 0ms (IQR: 6ms) for takeoff and landing, respectively (N=212).
For jumping and standing activities, two side-by-side force plates (sampling frequency: 1000 Hz and sampling frequency: 1500 Hz) were used as a gold standard for on and off ground force measurements. For walking and running activities, a dual force plate treadmill (sampling frequency: 100 Hz) was used as a gold standard for force measurement and finding heel strikes and toe offs.
We collected different activities and grouped them into the following categories for the analysis: 1) single-leg jumps (lateral, forward-backward, and height), 2) double-leg jumps (cyclic and consecutive countermovement jumps), 3) walking and running, 4) in-place activities (high knee, idle). These activities were selected to collect all types of data that our users collect on a daily basis.
Prior to collecting any movement data, each subject was weighed on each force plate individually, and then on both force plates together to ensure no offset between the plates. Before each activity, the force plates were zeroed. The subject stood still for five seconds, after which they stomped their left foot on the left force plate. This stomp allowed for syncing of force plate and IMU data. They then stood still for another five seconds and proceeded to complete the activity. All activities were started with the left leg, for consistency. After completing the activity, the subject remained still for five seconds, stomped their left foot, stood still for five seconds, and both the Plantiga and force plate timers were stopped at the same time.
The ground interaction (GRIN) model was designed to classify each IMU data point as an on-ground or off-ground. This feature enables us to perform an accurate analysis of gait parameters and jump activities for both single and double leg jumps. The GRIN model was developed by using a supervised deep learning method and a convolutional neural network (CNN) architecture.
The input of the model is the IMU data for each foot (6 axis time series for accelerometer and gyroscope) and on-ground/off-ground labels created from force plate data. The output of the model is a one-dimensional binary array (0 for off-ground and 1 for on-ground).
Labels (on-ground or off-ground) for each data point were generated using the force plate data. The collected data were separated into training, testing, and validation data. This report shows the performance of the model for validation data. This data has not been seen by the model and can show the robustness of the analysis for different data types.
From 67 activities, a total of 5276 events were collected. 13 activities had either false positive(s) (FP) or false negative(s) (FN). For individual activities, the maximum number of false positives was 4 and the maximum number of false negatives was 1. Overall, the ground interaction model missed 10 events (FNs) and found an extra 24 events (FPs). To process the timing accuracy of events, activities with either FPs or FNs were removed from the data. The remaining data points (5064 events) were separated by their activity type.
Figure 1: Double Leg Jump Error
Figure 1 shows landing and takeoff timing errors for double leg jumps. 766 events were captured. For takeoff timing error Q1 was -2 ms, the median was 0ms, Q3 was 2 ms, (range [-8ms, 8ms]). For landing timing error Q1 was -2ms, the median was 0ms, Q3 was 2ms, (range [- 8ms, 8ms]).
Figure 2: Single Leg Jump Error
![Figure 2 shows landing and takeoff timing errors for single-leg jumps. 618 events were captured. For takeoff timing error Q1 was -2ms, the median was 0ms, Q3 was 2ms, (range [-8ms, 8ms]). For landing timing error Q1 was -2ms, the median was 0ms, Q3 was 2ms, (range [-8ms, 8ms]).](https://s3-us-west-2.amazonaws.com/secure.notion-static.com/54329329-cc72-4d75-b0bf-3cb76f2cbc81/2.png)
Figure 2 shows landing and takeoff timing errors for single-leg jumps. 618 events were captured. For takeoff timing error Q1 was -2ms, the median was 0ms, Q3 was 2ms, (range [-8ms, 8ms]). For landing timing error Q1 was -2ms, the median was 0ms, Q3 was 2ms, (range [-8ms, 8ms]).
Figure 3: Walk and Run Error