An Intelligent In-Shoe System for Real-Time Gait Monitoring and Analysis

: The deterioration of gait can be used as a biomarker for ageing and neurological diseases. Continuous gait monitoring and analysis are essential for early deficit detection and personalized rehabilitation. The use of mobile and wearable inertial sensor systems for gait monitoring and analysis have been well explored with promising results in the literature. However, most of these studies focus on the technologies for gait characteristic assessment, few of them have considered the data acquisition bandwidth of the sensing system. Inadequate sampling frequency will sacrifice signal fidelity, thus leading to an inaccurate estimation especially for spatial gait parameters. In this work, we developed an inertial sensor based in-shoe gait analysis system for real-time gait monitoring and investigated the optimal sampling frequency to capture all the information on walking patterns. An exploratory validation study was performed using an optical motion capture system on four healthy adult subjects, where each person underwent five walking sessions, giving a total of 20 sessions. Percentage mean absolute errors (MAE%) obtained in stride time, stride length, stride velocity, and cadence while walking were 1.19%, 1.68%, 2.08%, and 1.23%, respectively. In addition, an eigen-analysis based graphical descriptor from raw gait cycle signals was proposed as a new gait metric that can be quantified by principal component analysis to differentiate gait patterns, which has great potential to be used as a powerful analytical tool for gait disorder diagnostics. of abnormal gait when compared to healthy subjects. It proves a valuable metric for tracking gait progress remotely throughout rehabilitation, and will help provide an appropriate treatment during early stages rehabilitation.


Introduction
Human locomotion is one of the most important abilities that must be acquired and maintained to perform activities of daily life (ADL) and, despite requiring little thought, implies a complex series of coordinated events within the body. This involves the communication of intricate sensory information, which is integrated in the nervous system and results in motor commands that control muscle activation and, finally, joint movement. Gait analysis provides important insight into the health and state of these Verbal descriptions of rehabilitation progress throughout the gait cycle tend to be imprecise. A comprehensive understanding and an objective data analysis subject to quantitative gait analysis is urgently needed.
In this paper, a miniaturized, low-power, cost-effective, highly mobile, and userfriendly in-shoe system embedded with inertial sensors, i.e., accelerometer, gyroscope and magnetometer, is developed to address hardware limitations particularly related to signal bandwidth and to devise new gait metrics. The integrated in-shoe system is able to collect long-term gait data without intervention and inconvenience for the subjects in a real-world setting. The aim of this study is three-fold: (1) develop a shoe-integrated inertial sensor-based gait analysis platform to perform reliable measurements and capture all relevant gait behaviors, investigating and optimizing system data acquisition bandwidth for gait analysis, (2) extend the well-established zero-velocity update technique with a gradient descent-based complementary Madgwick filter and heuristic techniques to identify and quantify spatiotemporal gait parameters, and (3) introduce an eigen-analysis based graphical metric quantified by principal component analysis, which can intuitively identify temporal characteristics of the gait cycles.

System setup
To perform gait analysis, a pair of motion sensing shoes called Nushu is developed in this study for data acquisition. The system comprises a customized sensor unit inserted in the posterior portion of the outsoles of the shoes (usually thicker and less prone to bending), as shown in Figure 1A. The sensor unit is fixed in place with silicon glue, and the upper component of the shoe is glued on top with a heat-activated contact cement normally used in the shoe-manufacturing industry. This allows for a tight seal and consequently the possibility of prolonged indoor and outdoor testing. Each sensor unit is equipped with a battery, a 32-bit microcontroller (ARM Cortex-M4), a Bluetooth low energy network processor (BlueNRG-MS, STMicroelectronics, Geneva, Switzerland), a micro-SD card socket and on-board sensors (LSM6DSM and LSM303AGR, STMicroelectronics, Geneva, Switzerland), including a tri-axis accelerometer (with selectable dynamic ranges of: ±2g, ±4g, ±8g or ±16g), a tri-axis gyroscope (125 dps to 2000 dps) and 3-axis magnetometer (±50 gauss). The sensors can record data continuously during walking as schematically illustrated in Figure 1B and Figure 1C. In the meantime, the sensor data can be either logged onto a local flash memory or streamed in real time via Bluetooth wirelessly to a mobile device. Data processing is performed off-line immediately after walking. With estimated spatiotemporal gait parameters, a report of 13 gait parameters such as stride velocity, stride time, stride length, minimum foot clearance, strike angle, stance time, swing time, stance phase, swing phase, cadence, maximum angular velocity, symmetry, and variability can be generated autonomously, as shown in Figure 1D.

Sensor Parameters Optimization
Aliasing effects have a large impact on the signal fidelity of current commercial inertial sensors. Inadequate sampling frequency can lead to critical aliasing problems, causing the digital representation of the analogue signal to be erroneous. Erroneous digital representation may not alter the estimation accuracy of the temporal parameters, but it can have a significant impact on the estimated spatial parameters, e.g., stride length, stride velocity due to the accumulation of the error caused by aliasing effect during the integration process. Therefore, optimizing the sampling frequency of the sensor is critical for improving their performance and also increasing the accuracy of the gait analysis results.
An anti-aliasing filter reduces undesired-frequency components above the Nyquist frequency prior to digital sampling. The anti-aliasing filter characteristics of the IMU sensor (BNO055, Bosch) was experimentally obtained in a setup where the sensor configured to a sample rate of 100 Hz was fixed on the surface of a speaker. Sinusoidal audio waves spanning a duration of one minute with different frequencies of 120 Hz, 125 Hz, and 130 Hz, which were slightly above the IMU sampling frequency, were generated using a computer (using Matlab R2017a, Mathworks). Those audio waves were played through the speaker and recorded by the accelerometer of the IMU sensor. The digitally sampled accelerometer data recorded in the time-domain are mathematically transformed into the frequency domain by the Fast Fourier Transform (FFT) for spectral analysis shown in Figure 2. The actual IMU digitized the data at around 99 Hz, low frequency components at 21 Hz, 26 Hz and 31 Hz were observed on the spectrum, illustrating the absence of the appropriate anti-aliasing filter before analogue to digital conversion in this commercialized IMU sensor. Moreover, even if the anti-aliasing filtering is present in IMU sensors, valuable highfrequency information above the cut-off frequency of anti-aliasing filter, such as fast variations in the acceleration, could still be lost [34]. Therefore, knowing the highest frequency component within a walking time series in advance can help deciding the optimal sampling rate, which is sufficient to precisely estimate gait parameters, but not too high so that the system remains power and memory efficient [35].
Nushu system was evaluated against three different sensors to determine the optimal sampling frequency and to investigate the validity of the choice of 100 Hz as the sampling frequency, which is a common choice for gait analysis [28,[36][37][38]. The first sensor is a wired, high-sensitivity piezoelectric accelerometer sensor (DJB). It can sample up to 5000 Hz with an anti-aliasing filter applied priori. The second sensor is a wireless commercial sensor (Axivity) with only accelerometer and a maximum sampling rate of 100 Hz, with no anti-aliasing filter priori. Nushu system, which has both accelerometers and gyroscopes was set to sample at 100 Hz, with no anti-aliasing filter priori during these measurements. All three sensors were attached to the subject's instep position of the shoes using straps. The subject was asked to walk in a straight line about 6m and walk back, followed by a series of jumps for 3 times. Three collected sets of time series acceleration data were normalized individually after removing DC bias (by means of selecting a short segment (5s) of the stationary signals, and subtracting the mean of the stationary signals). Each set of the obtained normalized acceleration data is denoted as ( ∆ ), ( ∆ ), ( ∆ ) , where ∆ = , is the sampling frequency, ∈ Ν , and ∈ {1, 2, 3} represents for three different sensors. Since no clock was shared between those three acquisition systems, a synchronization process was implemented as follows. The magnitude of acceleration was first calculated for each set as Then an anti-aliasing filter was applied to ( ∆ ) and down-sampled ( ∆ ) to min such that all signals have the same time interval ∆ . Time series ( ∆ ) and ( ∆ ) were shifted along the time axis such that their cross-correlation with ( ∆ ) was maximized. According to the obtained time lag based on the maximum correlation time point, the raw time series data ( ∆ ), ( ∆ ), ( ∆ ) were shifted along their time axes for synchronization. Using the synchronized data, each walking stride was detected and separated for spectral analysis with the help of the gyroscope from Nushu inertial sensor.
Furthermore, to find the optimal sampling rate for gait analysis, the signals with the highest sampling frequency (5000 Hz) from the piezoelectric sensor were cut off by low pass filters at different cut-off frequencies. The percent root mean squared errors (RMSE (%)) between the filtered data and the original data were compared after applying different cut-off frequencies. Then the optimal sampling rate was determined by balancing the RMSE (%) and the cut-off frequency.

Gait Analysis Method
To assess the gait performance, each gait cycle is divided into multiple gait phases such that quantitative gait features, including temporal and spatial gait parameters, can be evaluated. The gait cycle is typically divided into two main phases by gait events, stance and swing, as shown in Figure 1B. Those two phases can further be subdivided into eight functional phases, five during stance and three during swing. The first two phases, initial contact and loading response, occur during the weight acceptance with double feet support. Mid-stance and terminal stance go on during the single foot support, followed by pre-swing phase where the forward limb motion starts. Afterwards, the swing phase commences with the initial swing, in which the hip and knee start to bend in tandem with ankle dorsiflexion. The mid-swing immediately follows when the swinging leg is aligned with the standing leg. Finally, the terminal swing occurs when the leg decelerates by contraction of the hamstrings and prepares for ground contact [3,39]. The three key gait events to anchor these phases are heel strike (HS), when the heel strikes the ground at initial contact, flat foot (FF), when the foot is flat on the ground, and toe off (TO), when the toes leave the ground.
With the defined gait phases, 13 gait metrics are configured to be estimated by the Nushu system, as listed in Table 1. Additional gait parameters can be customized based on user requirements. Given the nature of the gait metrics, raw kinematic data from the inertial sensors are first transferred into the global reference frame to detect characteristic events during the walking cycle and estimate the orientation of the sensors. This allows for the extraction of spatial features through integration of the measured accelerations and angular velocities.

Attitude Estimation and Heuristic Techniques
The orientation of the foot is estimated through sensor fusion, which is a technique often used for Attitude and Heading Reference Systems (AHRS) that makes use of gyroscope, accelerometer, and magnetometer measurements through specific filters. Among many that have been implemented in previous work, our system uses the algorithm derived by Madgwick [40]. This gradient descent based complementary filter employs a quaternion representation and uses the gravitational acceleration and the earth's magnetic field picked up by the accelerometer and magnetometer to estimate the orientation error that arises from naively integrating the raw angular data from the gyroscope. Compared to other filters, e.g. Kalman or extended Kalman filters, Madgwick's algorithm compensates the magnetic distortion to eliminate the need for predefining the reference magnetic field direction [40]. It is also computationally less expensive and offers the potential of real-time data processing on Nushu's microcontroller.
If the sensor moves arbitrarily in space, the gravitational acceleration will become mixed with other linear accelerations. Heuristic techniques are required for the correct implementation of the Madgwick filter and the post-processing algorithm. These techniques exploit the fact that gait is a periodic series of alternate stances and swings. It is widely accepted that during the stance phase there is a short period of time, referred to as mid-stance, where the foot is stationary, while the shank pivots around the ankle, leading to linear accelerations of zero on the foot. Assuming that during mid-stance, the accelerometer measures only the gravitational acceleration, these stationary periods are used as flags for the AHRS algorithm.
Here, we introduce an approach to detecting dynamic phases and stationary phases by the following method. We define a new motion signal to be where ( ) and where F( ) is the binarized flag. The dynamic and stationary phase differentiation result is shown in Figure 3. The blue regions depict stationary phases, and the white regions represent the dynamic phases. With these stationary flags, the orientation of the global frame relative to the sensor frame is estimated through sensor fusion by Madgwick's filter, as shown in Figure 4. The filter gain , which represents gyroscope measurement error, in gradient descent is calibrated as 0.1. The acceleration data can thus be transformed from sensor frame into the global reference frame, as shown in Figure 5. Afterwards, each gait cycle is collected by segmenting the midpoint of the stationary regions for further processing.

Event Detection Algorithm
During a gait cycle, the rotation of the foot around the mediolateral axis (flexion/extension axis of the ankle) is the most prominent movement comparing to supination/pronation and inversion/eversion. Therefore, the key gait events HS, FF and TO can be detected by inspecting the angular velocity aligned with the mediolateral axis of the foot ( ) [36]. For a normal gait cycle, the foot rotation around the mediolateral axis changes as follows: a gait cycle begins with the foot flat on the ground; the foot starts to rotate forward with the toe as the contact point, the heel leaves the ground first with the rotation speeding up; then the toe leaves the ground, the rotation slows down; as the foot leans forward, the rotation reverses; near the end of the foot forward movement, the toe elevation exceeds the heel elevation, the rotation reverses again after the heel strike the ground; in the end of the gait cycle, the toe lower down with the heel as the contact point until the foot is flat again [36]. In Figure 6, the low pass filtered normal walking signal from gyroscope y axis is presented. The signals of walking cycles display a characteristic pattern with well-defined features (peaks and plateaus) that are associated with the gait events. The rules for event detection from the angular velocity are predefined according to the ground truth provided by the motion capture system. In particular, the toe off is associated with the strongest local maximum within the cycle, whereas heel strike corresponds to the following zero-crossing within the updated dynamic region. The FF event is observed as the first local maximum between heel strike and toe off. This instant resides within a plateau region of the gyroscope signal and generally represents the angular velocity which is closest to 0 dps. The detection algorithm uses a set of dynamic thresholds and local peak-identification techniques to recognize the sequence of the events. In order to further process the data, only valid strides (which contain the correct sequence <FF -TO -HS -FF>) have been considered. Furthermore, the event detection algorithm is used to extract temporal parameters such as swing time, stance time, swing phase and stance phase (as described in Table 1). Figure 6. The signal from the gyroscope aligned with the medio-lateral axis of the foot is used for gait event detection. Each stride is characterized by a sequence of FF, TO, HS and FF as indicated.
Drift accumulates due to the numerical integration of the acceleration errors. This is addressed by using the Zero-velocity Update technique (ZUPT), whose main assumptions are that accelerations and velocities are equal to zero during mid-stances, and that integration drift within a gait cycle is accumulated linearly [41,42]. Therefore, numerical integration can be carried out piecewise at every gait-cycle (using stationary intervals as reset points) to avoid the propagation of drift throughout the signal, and linear de-drifting is applied to each cycle to remove drift and discontinuities. Through the implementation of the AHRS and ZUPT algorithms, the accelerations are integrated in the global reference frame to extract spatial features such as stride velocity and stride length.
The gait analysis procedure discussed in this section is summarized in Figure 7.

Validation Method
To evaluate whether gait parameters could be accurately estimated by Nushu, four healthy adults (age 25-64 years old), each perform five walking sessions, twenty trials in total were measured and their data compared between optical motion capture system and Nushu in a validation experiment. Sixteen reflective markers were attached to each subject's shoes and lower limbs, i.e., anterior superior iliac spine (ASIS), posterior iliac spine (PSIS), distal lateral thigh, lateral femur epicondyle, distal lateral shank, lateral Malleoli, above the toe (second Metatarsal phalangeal joint), behind the calcaneus (the same height as the toe marker). This is a standard lower limb gait marker set. A highly accurate optical 3D gait analysis (3DGA) system (Vicon Oxford Metrics, Ltd. UK) with fourteen infrared cameras, which served as the reference system, was used to track the instantaneous position of markers located on subject's body segments during the walking experiment. The lower limb skeleton model constructed from 3DGA markers is presented in Figure 8. Both the optical motion capture system and the optimized Nushu system recorded walking data simultaneously at the optimal frequency , which is determined by the sampling optimization described in section 2.1. After the system set up, each healthy adult subject was instructed to complete five walking sessions along the same 10-meter long, straight line route, which was marked on the floor with yellow tape. During 20 walking sessions, each subject stood stationary for five seconds and then walked to the end of the marked area at his/her comfortable speed under the guidance of a computer-generated voice. The Nushu system was reset between each subject's walking test. Each walking session was also video recorded for later analysis in case of unexpected events. After experiments, the position of calcaneus marker was used to detect the heel strike and estimate the foot displacement. The marker above the second metatarsal bone was used to detect the toe off. The velocity was estimated by differentiating the marker locations as where is the position of 3DGA markers. The stride length was calculated as

Sampling Frequency Optimization
In Figure 9, a spectrum of acceleration signals from one normalized average stride captured by three different sensors sampled at 5000 Hz (blue: DJB), 100 Hz (red: Nushu), 100 Hz (yellow: Axivity) is presented. The amplitude indicates the strength of the frequency components relative to noise. We can see that the amplitude flatness is maintained up to 25 Hz for all X, Y, Z components. Subsequently, amplitudes start to decline until approximately 120 Hz, while they are still significantly larger than the amplitude of the high-frequency noise, which indicates that there are useful frequency components between 25 Hz and 120 Hz. Two sensors sampled at 100 Hz cannot capture useful information between 50 Hz and 120 Hz as shown in Figure 9. This shows that a sampling rate of 100 Hz is not adequate to capture all the relevant information for gait analysis. To find the optimal sampling frequency for gait analysis, the normalized detected stride data from the highest resolution piezoelectric sensor data was first processed with a series of low pass filters (LPFs). Those LPFs have stopband frequencies ranging from 50 Hz to 2500 Hz. The RMSE (%) between raw data and filtered data of X, Y, Z axis as a function of LPF cut-off frequency is shown in Figure 10. As the LPF cut-off frequency increases from 25 Hz to 125 Hz, the RMSE (%) decreases significantly. However, the further increase of LPF cut-off frequency above 125 Hz has no significant effect on the RMSE (%). Based on the trade-off between RMSE (%) and LPF cut-off frequency, the sampling frequency 250 Hz (twice of the stride frequency 125 Hz) was chosen as the optimal frequency , which could not only capture the most stride information but also save energy and memory. This result also shows a good agreement with the study in [34], which has found that the lowest sampling frequency for gait analysis lies between 200 Hz and 300 Hz.

Performance Evaluation
Eight parameters, stride time, stride length, swing time, stance time, velocity, cadence, swing phase and stance phase, were calculated using Nushu and compared to the ones calculated based on reference data from 3DGA. For each stride, the differences between Nushu and 3DGA of these eight parameters across all subjects' strides are shown in the normalized histogram ( Figure 11). The error metrics, mean absolute error (MAE), MAE (%), RMSE, MAE standard deviation (SD), averaged by each session for all subjects are reported in Table 2. Percentage MAEs for stride time, stride length, stride velocity, cadence, swing time and stance time were 1.19%, 1.68%, 2.08%, 1.23%, 3.02%, and 2.59%, respectively. Compared to other wearable sensors for healthy subjects' gait assessment in recent studies, the developed Nushu system has demonstrated high accuracy in estimated spatio-temporal gait metrics as shown in Table 3.

A New Gait Metric
To characterize human gait in a simple and intuitive way, a new gait hodograph descriptor obtained from the geometric gait features, is proposed. This gait descriptor reflects the characteristics of gait patterns without complex calculations and detects abnormal gait patterns intuitively. This is in contrast to advanced gait analysis algorithms, which can achieve relatively high performance, but are computationally expensive and cannot avoid errors in estimating the quantitative gait metrics.
For each gait cycle, a gait hodograph is formed by neglecting the time dimension and directly plotting the low pass filtered motion data from the accelerometer and gyroscope sensors. The low pass filter with a cut off frequency 20 Hz is applied to get the main characteristics, such that the shape characteristic is not noisy, but still contains the dominant information in the low frequencies. Among all motion signals obtained from the triaxial accelerometer, the triaxial gyroscope and the triaxial magnetometer, the trajectories of three most prominent signals  Figure 12 shows 2-D planar projections of walking cycle trajectories from three healthy persons' left ( Figure 12A, C, E) and right foot ( Figure 12B, D, F), and one stroke patient's (recording performed in a different experiment) left ( Figure  12G) and right ( Figure 12H) foot. The stroke patient's walking data on a treadmill are collected by Nushu in this section only for metric comparison. The corresponding walking movements are shown in Supplementary Movie 1-3 for healthy persons and Supplementary Movie 4 for stroke patient. In these hodographs, the closed planar shape is encircled once in the counterclockwise direction for each gait cycle. Green stars mark the HS, blue circles mark the TO. The green lines represent stance phases, blue lines describe swing phases. Subtle differences of shapes and positions are observed among three different healthy persons, for example, the variability in gait cycles of the first person ( Figure 12A-B) is smaller than the other two healthy persons ( Figure 12C-F). However, all hodographs of healthy individuals have the same well-recognizable shape characteristics. From the stroke patient's hodograph ( Figure 12G, H), the gait patterns of both feet significantly differ from healthy individuals. Recognizable shape characteristic for the right foot are not apparent. In addition, the range of ( ), ( ), ( ) for the stroke patient's right foot is significantly smaller than the ones for healthy individuals, which reflects reduced flexibility of the ankle dorsiflexion and knee flexion caused by stroke. This also agrees well with the fact that stroke affected this patient's right body more severely than the left side.
In Figure 13, a planar 2-D projection of ( ), ( ) and { ( ), ( )} are presented to provide information about foot movement forward-and-backward, left-and-right in the horizontal plane, and the rotation of the ankle joints around raw and yaw axis but does not indicate up-and-down foot movement and rotation around pitch axis during walking. When healthy people walk along a straight path, the feet do not significantly sway from side to side. Therefore, ( ) is limited to a small range. The 2-D projection hodograph of ( ), ( ) should be in an elongated ellipsoid. For the projection of { ( ), ( )}, the ankle joint's rotation around raw and yaw axes is limited, and the signal should be dispersed around the original point of the coordinate.
Afterwards, PCA was performed on these matrices. The PCA results of from three healthy people (yellow, blue, green) and one stroke patient (red) for left and right side are shown in Figure 14:   In those plots, the data of the three healthy subjects show different principal components, indicating their personal walking patterns. The data of the stroke patient shows a significant different principal component compared to the healthy subjects. As shown in Figure 14A and B, the singular values of stroke patient are much smaller than that of healthy subjects. Moreover, the principal component vectors of stroke patient in Figure 14B even lie in different quadrants from the vectors of the healthy subjects. This agrees with the fact that the right body side of the stroke patient is more severely compromised than the left side. Similar differences can also be observed in Figure 14C and D. The left side (less affected) shows a smaller confidence ellipse which reveals that the ankle rotation variation of the stroke patient is much smaller than healthy subjects. This also demonstrates that during walking, the patient's ankle joint is less flexible and the variability of spatio-temporal kinematic joint parameters is much smaller than healthy subjects [47,48]. The right foot (more affected) of the stroke patient demonstrates the largest deviation from normal, with the principal component vector manifesting in a different orientation than in healthy subjects.
These results suggest that the principal components can be a useful metric for quantifying personal gait characteristics and allows deficit detection of abnormal gait when compared to healthy subjects. It proves a valuable metric for tracking gait progress remotely throughout rehabilitation, and will help provide an appropriate treatment during early stages rehabilitation.

Discussion and Conclusion
In the presented study, we developed a wearable in-shoe gait analysis system embedded with on-board inertial sensors, i.e., accelerometer, gyroscope, magnetometer, that can transfer data to mobile devices wirelessly via Bluetooth in real-time or store data locally on an SD card without causing inconvenience for the users. This cost-effective, portable, and user-friendly in-shoe system enables real-time gait monitoring even during outdoor settings.
We examined the aliasing effect of an inertial sensor and identified the optimal sampling frequency for walking data acquisition. It is important to identify the frequency distribution of the useful signal information, especially for fast varying signals such as accelerometer. Inadequate sampling frequency will prohibit capturing high frequency gait signals, and leading to significant drift for spatial gait parameter estimation due to the continuous integration of the sensory data. Despite the advances in the field, this aspect is usually overlooked in the literature. We believe the research in the field will benefit from a standardization effort in determining optimal data acquisition parameters.
To estimate spatiotemporal gait parameters, a well-established ZUPT gait analysis technique was extended with a gradient descent-based Madgwick's filter and heuristic techniques. The estimation results of spatiotemporal gait parameters compared between the proposed system and the computer-based optical 3DGA system showed a good agreement, with percentage absolute errors of 1.19%, 1.68%, 2.08%, 1.23%, 3.02%, 2.59% for stride time, stride length, stride velocity, cadence, swing time and stance time respectively. For stride time, stride length, stride velocity and cadence, the validation results showed small errors. While for swing and stance time estimation, the estimation was relatively worse. Those two phases were separated by the transition from foot flexion to dorsal extension. While as the foot is not a rigid body segment, this definition may not hold for all strides and subjects, thus leading to a reduced accuracy of TO detection. We demonstrated our system in a small-scale study with healthy subjects. Having validated our hardware and software, we are planning to validate pin a larger set of young adults then move to older adults. We are also scheduling clinical studies for the validation of our approach for the diagnostics and monitoring of patient groups (such as Parkinson's, stroke and multiple sclerosis) and for the monitoring of specific population groups such as the elderly.
We introduced a new intuitive visualization method of describing spatiotemporal relations between collected gait data using hodographs, which captured clearly the variance in the walking data. It reflected the gait kinematics throughout every single gait cycle, which was different from traditional discrete metrics commonly used in the literature, such as walking speed, stride variability. Moreover, the hodographs clearly displayed important differences in walking characteristics between healthy subjects and stroke patients at the first glance without further processing the gait cycle data, which requires significant effort and introduces inevitable error when applying traditional gait metrics for abnormal gait recognition. Differences in gait patterns can be easily recognizable from graphical hodographs, even with a limited number of gait cycles, e.g. only 12 gait cycles in Figure 12. We complemented the visual hodographs with a new quantitative metric provided by a PCA algorithm. In the future, a shape recognition algorithm could further be developed for hodographs in order to perform automatic pathologic gait recognition and classification [49]. The combination of visual and quantitative metrics offers a great potential to become a diagnostic tool for differentiating gait patterns between individuals and tracking the progress of patients during their rehabilitation process.
Supplementary Materials: The following are available online at www.mdpi.com/xxx/s1, Video S1: Walking pattern of healthy subject 1. Video S2: Walking pattern of healthy subject 2. Video S3: Walking pattern of healthy subject 3. Video S4: Walking pattern of stroke patient.