Continuous Analysis of Running Mechanics by Means of an Integrated INS/GPS Device

paper describes a single body-mounted sensor that integrates accelerometers, gyroscopes, compasses, barometers, a GPS receiver, and a methodology to process the data for biomechanical studies. The sensor and its data processing system can accurately compute the speed, acceleration, angular velocity, and angular orientation at an output rate of 400 Hz and has the ability to collect large volumes of ecologically-valid data. The system also segments steps and computes metrics for each step. We analyzed the sensitivity of these metrics to changing the start time of the gait cycle. Along with traditional metrics, such as cadence, speed, step length, and vertical oscillation, this system estimates ground contact time and ground reaction forces using machine learning techniques. This equipment is less expensive and cumbersome than the currently used alternatives: Optical tracking systems, in-shoe pressure measurement systems, and force plates. Another advantage, compared to existing methods, is that natural movement is not impeded at the expense of measurement accuracy. The proposed technology could be applied to different sports and activities, including walking, running, motion disorder diagnosis, and geriatric studies. In this paper, we present the results of tests in which the system performed real-time estimation of some parameters of walking and running which are relevant to biomechanical research. Contact time and ground reaction forces computed by the neural network were found to be as accurate as those obtained by an in-shoe pressure measurement system. Keywords: gait analysis; INS/GPS; machine learning; neural networks; sports equipment; velocity measurement 1. Introduction Contemporary systematic approaches for the analysis of walking and running mechanics use a combination of high-speed

relevant to biomechanical research. Contact time and ground reaction forces computed by the neural network were found to be as accurate as those obtained by an in-shoe pressure measurement system. Keywords: gait analysis; INS/GPS; machine learning; neural networks; sports equipment; velocity measurement 1. Introduction Contemporary systematic approaches for the analysis of walking and running mechanics use a combination of high-speed video analysis and ground reaction force and EMG measurements, which can provide relatively detailed gait information including the involved forces, pressure distribution, joint angles, running velocity, and other characteristics. However, the limitations of such approaches include high cost, lack of portability, lengthy analysis time, and the need for trained personnel. A strong need exists for small and unobtrusive equipment that still provides meaningful information about athletic performance in the eld without unnecessarily encumbering the athlete or constraining natural movement. It should also be possible to make such information rapidly available to coaches and Sensors2019,19, 1480; doi:10.3390/s19061480

Sensors2019,19, 1480 2 of 20 athletes, presented in a form that is practically relevant and easy to interpret [1]. It is well-established that such equipment enables more frequent and regular collection of data, providing a continuous and comprehensive picture of individual-speci c training adaptation, rather than just occasional snapshots. The ability to collect data in the eld would also expand the horizons of movement analysis for clinical applications and physical rehabilitation treatments [2,3]. Application of inertial measurement sensors is a promising approach for movement analysis [4–7], which can provide an alternative to the traditional tools. Previous studies have used inertial measurement units (IMU) to capture human motion [4], compute stride (a gait cycle, two consecutive steps) parameters [5], assess running performance [6], and detect the type of foot strike (rearfoot or forefoot) [7]. Some approaches require foot-mounted IMUs [8] or multiple IMUs [9,10]. Inertial sensors can be used, for example, to detect temporal or spatial features of gait, such as the sharp peaks occurring when the foot hits the ground. Most publications consider only simple parameters, such as the stride time, that can be detected by accelerometers attached to the foot [8], thigh [11], and waist [12], or a combination of foot, shank, and thigh accelerometers [13]. Some studies have also used gyroscopes attached to the foot and shank to estimate the stride duration [14]. When sensors are placed on both legs, it is possible to analyze both step time and gait symmetry [15]. Some gait features, such as stride length, ground contact time (GCT), and ground reaction force (GRF) can be estimated indirectly based on the time series of the measured body and foot acceleration. These methods usually use regressions or arti cial neural networks to compute the relationships between the acceleration vector and gait features [16–19]. However, due to the physiological differences between people, these methods require calibration or training data for each person to estimate gait features with reasonable accuracy [20]. Consumer devices for monitoring running performance include GPS tracking and accelerometer-based systems to count strides, step length, ground contact time, and vertical oscillation. These simple devices

between the acceleration vector and gait features [16–19]. However, due to the physiological differences between people, these methods require calibration or training data for each person to estimate gait features with reasonable accuracy [20]. Consumer devices for monitoring running performance include GPS tracking and accelerometer-based systems to count strides, step length, ground contact time, and vertical oscillation. These simple devices provide useful information about running duration, distance, and velocity, but lack accurate and detailed information about many key aspects of running techniques [21]. The IMU-based methods don't provide all of the parameters that are required for running mechanics and kinematic analysis (e.g., these methods cannot compute speed). Their typical accuracy, in measured orientation (1–2 ), is insuf cient for kinematic analysis [22,23]. Additionally, in wireless IMUs, data transfer to the host computer is not reliable [9]. The aim of this paper is to present a measurement set-up based on an inertial navigation system (INS) combined with GPS (INS/GPS) to perform continuous 3D analysis of gait mechanics and technique during a session of outdoor running and walking on level terrain. This approach provides a range of important parameters that could have numerous applications in biomechanics research, such as detailed step characteristics, in-depth running mechanics, running ef ciency, and inter-individual responses to fatigue. The current version works only for walking and running on level terrain. Other activities and types of terrains are excluded from the consideration. The data segments recorded during irregular motion (neither walking nor running) are discarded. The main novelty of our approach lies in two key areas: (a) The comprehensive range of parameters that we compute gives more extensive information about running performance, compared to existing methods such as motion analysis and force platforms alone, and allows unimpeded movement; and (b) we compute these parameters with high precision and at high frame rates, making the approach suitable for research-grade experiments, as opposed to many commercial devices which only give a low-resolution overview of exercise performance.

and (b) we compute these parameters with high precision and at high frame rates, making the approach suitable for research-grade experiments, as opposed to many commercial devices which only give a low-resolution overview of exercise performance.

Sensors2019,19, 1480 3 of 20 2. Methods This section describes the measurement equipment and data processing issues, including gait segmentation, computation of running/walking metrics, vertical velocity, and displacement computation, and the prediction of GCT and GRF using only body-mounted sensors. 2.1. Measurement Setup The measurement setup consists of a Raspberry Pi 3 model B board running a Linux OS, a Vectornav VN-200 GPS-aided inertial navigation system (INS/GPS), a GPS antenna, and a 4200 mAh power bank (see Figure). The VN-200 INS/GPS is connected to the board through an UART (universal asynchronous receiver/transmitter) serial connection. The GPS antenna is located in close proximity to the INS, making the lever arm error negligible. All components are packaged in a 3D-printed case. The data from the INS/GPS are stored on a memory card. Figure 1. The data logger includes a Vectornav VN-200 inertial navigation system (INS)/GPS, a GPS antenna, a Raspberry Pi 3, and a battery. The battery is under the board. The data logger placement during eld tests is on the right image. The GPS antenna is pointing upwards. After the exercise, the data is transmitted to cloud storage using a 4G/LTE USB modem connected to the Raspberry Pi. The data analysis is done on a computer after the exercise. It is technically possible to send the data to the running coach during the exercise by 4G/LTE, but this is not implemented in the current version of our data logger. This is a small (150 75 48 mm 3 , about 400 g), completely autonomous, and self-contained data logger that can be used in different eld tests to measure many types of movement, including walking and running. The standard 4200 mAh power bank can provide power to the data logger for 4–5 h. The cost of components in our prototype is about 2500 euro. However, the cost (and size) can be reduced if, instead of the Raspberry Pi, a system on chip (SoC) is used and, instead of the rugged version of the VN-200, we use a surface mounted device or another less-expensive alternative. The size is mostly limited by the battery.

h. The cost of components in our prototype is about 2500 euro. However, the cost (and size) can be reduced if, instead of the Raspberry Pi, a system on chip (SoC) is used and, instead of the rugged version of the VN-200, we use a surface mounted device or another less-expensive alternative. The size is mostly limited by the battery. If the power consumption is optimized, it can be reduced to about one watt, so an 1800 mAh battery will be enough to provide power for 5 h. In our outdoor walking tests, the unit was attached to the torso and the test route was on a level outdoor track. The INS/GPS data are computed in the geographical coordinate frame with North, East, and Down axes. The output parameters are sampled at 400 Hz and include position, velocity, acceleration, orientation, angular velocity, and ground track (the path on the Earth's surface). Each individual VN-200 INS/GPS sensor undergoes a robust calibration and acceptance testing process at VectorNav's manufacturing facility. The accuracy speci cations for real-time applications provided by the manufacturer are [24]: Velocity accuracy: 0.05 m/s Heading, true inertial: 0.3 RMS

Sensors2019,19, 1480 4 of 20 Pitch/Roll: 0.1 RMS Angular resolution:<0.05 Repeatability:<0.1 This accuracy can be achieved when there is good GPS signal without multipath. Accurate heading estimation requires movement at a speed greater than approximately 1.5–2 m/s. At stand-still, the heading accuracy drops to approximately 1–2 , depending on the magnetic environment. The computation of orientation in an integrated INS/GPS is completely different from that in an IMU. The fusion of INS and GPS is usually implemented using an error-state extended Kalman lter (EKF), whose states include at least errors of position, velocity, orientation, gyro drifts, and accelerometer biases (15 states). In INS, the orientation errors are observable through the velocity measurement orientation errors, and the gyro drifts can be estimated using GPS velocity measurements. The IMUs that are usually used in biomechanical research are, in fact, attitude and heading reference systems (AHRS). In AHRS, the vertical angles (pitch and roll) are usually corrected using the gravity vector, and heading is corrected using a magnetometer. In a typical high-end MEMS (microelectromechanical systems) IMU/AHRS (e.g., Xsens MTx), the orientation errors are about 1–2 . These errors can be larger when strong magnetic disturbances are present or during continuous high-dynamic applications. An integrated INS/GPS does not exhibit these problems. For the same accelerometers and gyros, the error in pitch and roll is about 0.1–0.2 , and 0.3 for heading. Magnetic disturbances don't affect the accuracy. High dynamics have little impact on the performance. As is standard for human gait analysis, velocity and acceleration are computed in the anatomical frame, in which thex-axis is pointing in the anterior direction (direction of progression), thez-axis is vertical (parallel to the eld of gravity) and points upwards, and they-axis is perpendicular to thex- andz-axes and completes a right-hand triple. Some parameters, such as ground contact time (GCT) and ground reaction forces (GRF), can be computed using machine learning methods and, therefore, require additional in-shoe pressure measurement equipment or multiple IMUs to create datasets for training and validation. The general system architecture, including instrumented insoles and/or force plates, is shown in Figure. Figure 2. The system architecture.

a right-hand triple. Some parameters, such as ground contact time (GCT) and ground reaction forces (GRF), can be computed using machine learning methods and, therefore, require additional in-shoe pressure measurement equipment or multiple IMUs to create datasets for training and validation. The general system architecture, including instrumented insoles and/or force plates, is shown in Figure. Figure 2. The system architecture. An optional in-shoe pressure measurement system provides a dataset for training and validation of machine learning methods that can be used for indirect estimation of ground contact time (GCT) and ground reaction forces (GRF).

Sensors2019,19, 1480 5 of 20 2.2. Data Processing Additional data processing includes transformation of velocity and acceleration to the anatomical frame, step segmentation, computation of running metrics, and optional velocity and orientation accuracy improvement in post-processing. For running and walking applications, transformation of velocity and acceleration to the anatomical frame is accurately approximated by computing horizontal speed and the direction of movement, which is sometimes called the ground track. The following equations can be used to approximate speed and ground track: V(t) = q VN(t) 2 +VE(t) 2 (1) Y(t) =atan(VE(t)/VN(t)), (2) whereVE(t),VN(t) are the East and North velocity components,V(t)is speed, andY(t)is the ground track. In a similar way, acceleration is transformed to forward acceleration by computing the magnitude of the horizontal acceleration components. Examples of computed speed, ground track, and forward and vertical acceleration are shown in Figures. Figure 3. Motion parameters computed by the INS/GPS system for walking: Speed, forward and vertical accelerations, and ground track. Vertical lines show the beginning of the gait cycle.

Sensors2019,19, 1480 6 of 20 Figure 4.Motion parameters computed by the INS/GPS system for running: Speed, forward and vertical accelerations, and ground track. Vertical lines show the beginning of the gait cycle. The advantages of INS/GPS for velocity computation, compared to a typical single-frequency GPS receiver with 5 Hz output rate, are demonstrated in Figures. In Figure, the GPS velocity error is typical of the kind of consumer-grade receivers used in wearable devices for sport and well-being monitoring, when signal conditions are optimal. In Figure caused by multipath errors or the obstruction of some of the GPS satellites. In both cases, the combined INS/GPS solution has better accuracy and time-resolution: It can accurately follow the actual movement at an output rate of 400 Hz. The combined solution accurately computes motion features that may be useful in gait analysis. These features are valuable when machine learning (especially deep learning) methods are used to indirectly estimate some gait parameters, such as initial contact and toe-off. The accuracy of combined INS/GPS can be improved in post-processing, for example, by applying the Bayesian smoothing procedure described in [25]. Figure 5. Walking speed computed by the INS/GPS integrated system and by the GPS receiver only. The plot shows the typical velocity accuracy for a consumer-grade single frequency GPS receiver.

Sensors2019,19, 1480 7 of 20 Figure 6. Walking speed computed by the INS/GPS integrated system and by the GPS receiver only. The plot shows degraded velocity accuracy for the same GPS receiver. 2.3. Gait Segmentation There are many different approaches to human gait segmentation [26,27]. These approaches usually use body- or foot-mounted accelerometer measurements [12,28] to identify gait phases, such as touchdown or toe-off. The timing accuracy of these approaches is about 10–15 ms [29]. However, their reliability and robustness with respect to changing walking and running conditions or different gait styles is not very good. Our approach towards gait segmentation is based on the periodicity of the vertical velocity. We de ne the beginning of each step to be the instant when vertical velocity changes from positive (center of mass (CoM) moving upwards) to negative (i.e., the instant when the CoM is at its highest). When vertical velocity falls to zero again and the CoM position is highest, this instant is close toe-off, and this is especially accurate in running. Figures The algorithm for gait segmentation requires pre-processing to remove the low-frequency drift in the vertical velocity caused mainly by accelerometer bias or due to walking/running on an inclined surface. We used a high-pass lter to remove the low-frequency velocity drift. This was implemented using the MATLAB Signal Processing Toolbox highpass function, which uses a minimum-order lter with de ned stopband attenuation to compensate for the delay introduced by the lter. The following function parameters were chosen: Normalized passband frequency (wpass),0.005prad/sample; attenuation (stopbandattenuation), 30 dB; and steepness (steepness), 0.7. The input to the gait segmentation algorithm is a data window of the ltered vertical velocity data. In our experiments, the length of this window varied signi cantly, from 4 to 450 s, and the step segmentation accuracy was not affected. However, this data window must not include any irregular motion (i.e., motion that is neither walking nor running). The rst stage is to determine the points where the vertical velocity is close to zero; that is, below the pre-de ned threshold. During one step, the vertical velocity is equal

The study presents a sensor system for analyzing running mechanics.