Inertial Sensor Data Research Articles

How low can we go? Influence of sample rate on equine pelvic displacement calculated from inertial sensor data.

Low-cost sensor devices are often limited in terms of sample rate. Based on signal periodicity, the Nyquist theorem allows determining the minimum theoretical sample rate required to adequately capture cyclical events, such as pelvic movement in trotting horses. To quantify the magnitude of errors arising with reduced sample rates when capturing biological signals using the example of pelvic time-displacement series and derived minima and maxima used to quantify movement asymmetry in lame horses. Data comparison. Root mean square (RMS) errors between the 'reference' time-displacement series, captured with a validated inertial sensor at 100Hz sample rate, and down-sampled time-series (8Hz to 50Hz) are calculated. Accuracy and precision are determined for maxima and minima derived from the time-displacement series. Average RMS errors are <2mm at 50Hz sample rate, <4mm at 40Hz, <7mm between 25 and 35Hz, and increase to up to 20mm at 20Hz and below. Accuracy for maxima and minima is generally below 1mm. Precision is 1mm at 50Hz sample rate, 3mm at 40Hz and ≥9mm at 20Hz and below. Only sample rate, no other sensor parameters were investigated. Sample rate related errors for inertial sensor derived time-displacement series of pelvic movement are <2mm at 50Hz, a rate that many low-cost loggers, smartphones or wireless sensors can sustain hence rendering these devices valid options for quantifying parameters relevant for lameness examinations in horses.

Identification of Social Aspects by Means of Inertial Sensor Data

Today’s applications and providers are very interested in knowing the social aspects of users in order to customize the services they provide and to be more effective. Among the others, the most frequented places and the paths to reach them are information that turns out to be very useful to define users’ habits. The most exploited means to acquire positions and paths is the GPS sensor, however it has been shown how leveraging inertial data from installed sensors can lead to path identification. In this work, we present a Computationally Efficient algorithm to Reconstruct Vehicular Traces (CERT), a novel algorithm which computes the path traveled by a vehicle using accelerometer and magnetometer data. We show that by analyzing data obtained through the accelerometer and the magnetometer in vehicular scenarios, CERT achieves almost perfect identification for medium and small sized cities. Moreover, we show that the longer the path, the easier it is to recognize it. We also present results characterizing the privacy risks depending on the area of the world, since, as we show, urban dynamics play a key role in the path detection.

Deep Learning-Based Positioning of Visually Impaired People in Indoor Environments.

Wayfinding and navigation can present substantial challenges to visually impaired (VI) people. Some of the significant aspects of these challenges arise from the difficulty of knowing the location of a moving person with enough accuracy. Positioning and localization in indoor environments require unique solutions. Furthermore, positioning is one of the critical aspects of any navigation system that can assist a VI person with their independent movement. The other essential features of a typical indoor navigation system include pathfinding, obstacle avoidance, and capabilities for user interaction. This work focuses on the positioning of a VI person with enough precision for their use in indoor navigation. We aim to achieve this by utilizing only the capabilities of a typical smartphone. More specifically, our proposed approach is based on the use of the accelerometer, gyroscope, and magnetometer of a smartphone. We consider the indoor environment to be divided into microcells, with the vertex of each microcell being assigned two-dimensional local coordinates. A regression-based analysis is used to train a multilayer perceptron neural network to map the inertial sensor measurements to the coordinates of the vertex of the microcell corresponding to the position of the smartphone. In order to test our proposed solution, we used IPIN2016, a publicly-available multivariate dataset that divides the indoor environment into cells tagged with the inertial sensor data of a smartphone, in order to generate the training and validating sets. Our experiments show that our proposed approach can achieve a remarkable prediction accuracy of more than 94%, with a 0.65 m positioning error.

Evaluation of Inertial Sensor Data by a Comparison with Optical Motion Capture Data of Guitar Strumming Gestures.

Computing technologies have opened up a myriad of possibilities for expanding the sonic capabilities of acoustic musical instruments. Musicians nowadays employ a variety of rather inexpensive, wireless sensor-based systems to obtain refined control of interactive musical performances in actual musical situations like live music concerts. It is essential though to clearly understand the capabilities and limitations of such acquisition systems and their potential influence on high-level control of musical processes. In this study, we evaluate one such system composed of an inertial sensor (MetaMotionR) and a hexaphonic nylon guitar for capturing strumming gestures. To characterize this system, we compared it with a high-end commercial motion capture system (Qualisys) typically used in the controlled environments of research laboratories, in two complementary tasks: comparisons of rotational and translational data. For the rotations, we were able to compare our results with those that are found in the literature, obtaining RMSE below 10° for 88% of the curves. The translations were compared in two ways: by double derivation of positional data from the mocap and by double integration of IMU acceleration data. For the task of estimating displacements from acceleration data, we developed a compensative-integration method to deal with the oscillatory character of the strumming, whose approximative results are very dependent on the type of gestures and segmentation; a value of 0.77 was obtained for the average of the normalized covariance coefficients of the displacement magnitudes. Although not in the ideal range, these results point to a clearly acceptable trade-off between the flexibility, portability and low cost of the proposed system when compared to the limited use and cost of the high-end motion capture standard in interactive music setups.

CNN-Based Multistage Gated Average Fusion (MGAF) for Human Action Recognition Using Depth and Inertial Sensors

Convolutional Neural Network (CNN) provides leverage to extract and fuse features from all layers of its architecture. However, extracting and fusing intermediate features from different layers of CNN structure is still uninvestigated for Human Action Recognition (HAR) using depth and inertial sensors. To get maximum benefit of accessing all the CNN's layers, in this paper, we propose novel Multistage Gated Average Fusion (MGAF) network which extracts and fuses features from all layers of CNN using our novel and computationally efficient Gated Average Fusion (GAF) network, a decisive integral element of MGAF. At the input of the proposed MGAF, we transform the depth and inertial sensor data into depth images called sequential front view images (SFI) and signal images (SI) respectively. These SFI are formed from the front view information generated by depth data. CNN is employed to extract feature maps from both input modalities. GAF network fuses the extracted features effectively while preserving the dimensionality of fused feature as well. The proposed MGAF network has structural extensibility and can be unfolded to more than two modalities. Experiments on three publicly available multimodal HAR datasets demonstrate that the proposed MGAF outperforms the previous state-of-the-art fusion methods for depth-inertial HAR in terms of recognition accuracy while being computationally much more efficient. We increase the accuracy by an average of 1.5% while reducing the computational cost by approximately 50% over the previous state-of-art.

A Method of Human Activity Recognition in Transitional Period

Human activity recognition (HAR) has been increasingly used in medical care, behavior analysis, and entertainment industry to improve the experience of users. Most of the existing works use fixed models to identify various activities. However, they do not adapt well to the dynamic nature of human activities. We investigated the activity recognition with postural transition awareness. The inertial sensor data was processed by filters and we used both time domain and frequency domain of the signals to extract the feature set. For the corresponding posture classification, three feature selection algorithms were considered to select 585 features to obtain the optimal feature subset for the posture classification. And We adopted three classifiers (support vector machine, decision tree, and random forest) for comparative analysis. After experiments, the support vector machine gave better classification results than other two methods. By using the support vector machine, we could achieve up to 98% accuracy in the Multi-class classification. Finally, the results were verified by probability estimation.

Optimization of a wearable speed monitoring device for welding applications

The heat input is a parameter used to ensure quality of submerged metal arc welding, one of the most common metal joining procedures in the manufacturing industry. This parameter is dependent on the current and voltage, defined by the specific welding procedure, along with the welding speed, a feature that is strongly dependent on the hand movement of the welding operator. A wearable speed monitoring device (SMD) was developed by fusing optical and inertial sensor data to measure the handheld device speed of the welder. This work aims at tuning the parameters affecting the performance of SMD prototype by testing it with four batches of common electrodes in a real industrial environment. The outcomes were compared with the results from the traditional manual calculation, namely, a true coarse average (TCA) velocity. At first, observations were performed to identify any scope of improvement for the SMD. Secondly, an optimization of the main parameters was performed by minimizing a cost function involving the TCA velocities for each batch. A final assessment was then performed with the optimized parameters evaluating the achieved results; a comparative analysis showed a reduction of the relative average errors of 32%, 1.4%, and 4.2% for 2.5, 4, and 5 mm electrode sets, respectively. The system was found already optimized for the 3.25-mm electrodes, using baseline parameters with an error of about 12%. The methodology employed in this study can be further utilized in scenarios in which speed monitoring is employed for light-driven applications with a moving reference.

Inertial Sensor-Based Instrumented Cane for Real-Time Walking Cane Kinematics Estimation.

Falls are among the main causes of injuries in elderly individuals. Balance and mobility impairment are major indicators of fall risk in this group. The objective of this research was to develop a fall risk feedback system that operates in real time using an inertial sensor-based instrumented cane. Based on inertial sensor data, the proposed system estimates the kinematics (contact phase and orientation) of the cane. First, the contact phase of the cane was estimated by a convolutional neural network. Next, various algorithms for the cane orientation estimation were compared and validated using an optical motion capture system. The proposed cane contact phase prediction model achieved higher accuracy than the previous models. In the cane orientation estimation, the Madgwick filter yielded the best results overall. Finally, the proposed system was able to estimate both the contact phase and orientation in real time in a single-board computer.

A Novel Approach for Fall Risk Prediction Using the Inertial Sensor Data From the Timed-Up-and-Go Test in a Community Setting

Post-stroke patients usually suffer from a higher fall risk. Identifying potential fallers and giving them proper attention could reduce their chance of a fall that results in severe injuries and decreased quality of life. In this study, we introduced a novel approach for fall risk prediction that evaluates Short-form Berg Balance Scale scores via inertial measurement unit data measured from a 3-meter timed-up-and-go test. This approach used sensor technology and was thus easy to implement, and allowed a quantitative analysis of both gait and balance. The results showed that elastic net logistic regression achieved the best performance with 85% accuracy and 88% area under the curve compared with support vector machine, least absolute shrinkage and selection operator (LASSO), and stepwise logistic regression. This paper provides a framework for using sensor-based features together with a feature-selection strategy for screening and predicting the fall risk of post-stroke patients in a convenient setup with high accuracy. The findings of this study will not only enable the assessment of fall risk among post-stroke patients in a cost-effective manner but also provide decision-making support for community care providers and medical professionals in the form of sensor-based data on gait performance.

A Wearable Magnetometer-Free Motion Capture System: Innovative Solutions for Real-World Applications

Although inertial and magnetic wearable sensors are promising tools to develop novel technologies for human motion capture, their diffusion is being limited by their fair accuracy. In indoor applications, most of the inaccuracy comes from the magnetic disturbances contained in the magnetometer data. Besides, non-technicians might easily fail in properly calibrating the magnetometers. Hence, a novel magnetometer-free motion capture system was developed in this work, in the attempt to increase sensors accuracy and usability. Innovative strategies, invisible to the user, were introduced to compensate for the lack of the magnetometer: (a) an accurate linear Kalman filter-based 3D orientation estimator; (b) a simple sensor-to-segment calibration procedure; (c) efficient error compensations for soft-tissue artifact; (d) a simple solution to guarantee a common reference to all body segments. Ten expert Yoga practitioners were asked to perform a common sequence (sun salutation) while inertial sensor data and ground-truth data were acquired for the upper-body through a customized body area network and a stereophotogrammetric system, respectively. As expected, the overall joint angles accuracy proved to be dependent on the acquisition time window. The achieved worst-case accuracy at 90 s was 6 and 14 deg, depending on the main rotational axis involved (vertical or horizontal axis, respectively). Given the large and unconstrained angular workspace explored at each joint (up to 200 deg), these errors make the proposed solution (which can easily be extended to the full human body) an eligible alternative to current motion capture systems relying on magnetometers.

Doppler Shift Mitigation in Acoustic Positioning Based on Pedestrian Dead Reckoning for Smartphone

Acoustic localization is a viable solution for high-precision indoor positioning with current built-in smartphone sensors. However, the Doppler shift produces an additional error in the time difference of arrival (TDOA) estimation when the smartphone is in motion, resulting in degradation in the acoustic localization performance. In this article, we present an indoor localization system for smartphones that overcomes the effects of the Doppler shift by combining inertial and acoustic sensors data. First, we establish an error correction model related to the current position and velocity of the smartphone, which can estimate and mitigate the errors caused by the Doppler shift in the TDOA measurements. Second, a uniform acceleration dynamic model and an improved pedestrian dead reckoning (PDR) algorithm are designed to obtain the model parameter. Experiments show that the two proposed algorithms reduce the positioning error by 49.1% and 71.9% under continuous acoustic signal conditions and by 33.3% and 69.8% under occasional abnormal signal conditions. The presented system can achieve a positioning accuracy of about 0.2 m (RMS).

Adherence Tracking With Smart Watches for Shoulder Physiotherapy in Rotator Cuff Pathology: Protocol for a Longitudinal Cohort Study.

BackgroundPhysiotherapy is essential for the successful rehabilitation of common shoulder injuries and following shoulder surgery. Patients may receive some training and supervision for shoulder physiotherapy through private pay or private insurance, but they are typically responsible for performing most of their physiotherapy independently at home. It is unknown how often patients perform their home exercises and if these exercises are performed correctly without supervision. There are no established tools for measuring this. It is, therefore, unclear if the full benefit of shoulder physiotherapy treatments is being realized.ObjectiveThe proposed research will (1) validate a smartwatch and machine learning (ML) approach for evaluating adherence to shoulder exercise participation and technique in a clinical patient population with rotator cuff pathology; (2) quantify the rate of home physiotherapy adherence, determine the effects of adherence on recovery, and identify barriers to successful adherence; and (3) develop and pilot test an ethically conscious adherence-driven rehabilitation program that individualizes patient care based on their capacity to effectively participate in their home physiotherapy.MethodsThis research will be conducted in 2 phases. The first phase is a prospective longitudinal cohort study, involving 120 patients undergoing physiotherapy for rotator cuff pathology. Patients will be issued a smartwatch that will record 9-axis inertial sensor data while they perform physiotherapy exercises both in the clinic and in the home setting. The data collected in the clinic under supervision will be used to train and validate our ML algorithms that classify shoulder physiotherapy exercise. The validated algorithms will then be used to assess home physiotherapy adherence from the inertial data collected at home. Validated outcome measures, including the Disabilities of the Arm, Shoulder, and Hand questionnaire; Numeric Pain Rating Scale; range of motion; shoulder strength; and work status, will be collected pretreatment, monthly through treatment, and at a final follow-up of 12 months. We will then relate improvement in patient outcomes to measured physiotherapy adherence and patient baseline variables in univariate and multivariate analyses. The second phase of this research will involve the evaluation of a novel rehabilitation program in a cohort of 20 patients. The program will promote patient physiotherapy engagement via the developed technology and support adherence-driven care decisions.ResultsAs of December 2019, 71 patients were screened for enrollment in the noninterventional validation phase of this study; 65 patients met the inclusion and exclusion criteria. Of these, 46 patients consented and 19 declined to participate in the study. Only 2 patients de-enrolled from the study and data collection is ongoing for the remaining 44.ConclusionsThis study will provide new and important insights into shoulder physiotherapy adherence, the relationship between adherence and recovery, barriers to better adherence, and methods for addressing them.International Registered Report Identifier (IRRID)DERR1-10.2196/17841

Eating Episode Detection with Jawbone-Mounted Inertial Sensing.

Recent work in Automated Dietary Monitoring (ADM) has shown promising results in eating detection by tracking jawbone movements with a proximity sensor mounted on a necklace. A significant challenge with this approach, however, is that motion artifacts introduced by natural body movements cause the necklace to move freely and the sensor to become misaligned. In this paper, we propose a different but related approach: we developed a small wireless inertial sensing platform and perform eating detection by mounting the sensor directly on the underside of the jawbone. We implemented a data analysis pipeline to recognize eating episodes from the inertial sensor data, and evaluated our approach in two different conditions: in the laboratory and in naturalistic settings. We demonstrated that in the lab (n=9), the system can detect eating with 91.7% precision and 91.3% recall using the leave-one-participant-out cross-validation (LOPO-CV) performance metric. In naturalistic settings, we obtained an average precision of 92.3% and a recall of 89.0% (n=14). These results represent a significant improvement (>10% in F1 score) over state-of-the-art necklace-based approaches. Additionally, this work presents a wearable device that is more inconspicuous and thus more likely to be adopted in clinical applications.

Automatic clinical gait test detection from inertial sensor data.

The analysis of gait data is one approach to support clinicians with the diagnosis and therapy of diseases, for example Parkinson's disease (PD). Traditionally, gait data of standardized tests in the clinic is analyzed, ensuring a predefined setting. In recent years, long-term home-based gait analysis has been used to acquire a more representative picture of the patient's disease status. Data is recorded in a less artificial setting and therefore allows a more realistic perception of the disease progression. However, fully unsupervised gait data without additional context information impedes interpretation. As an intermediate solution, performance of gait tests at home was introduced. Integration of instrumented gait test requires annotations of those tests for their identification and further processing. To overcome these limitations, we developed an algorithm for automatic detection of standardized gait tests from continuous sensor data with the goal of making manual annotations obsolete. The method is based on dynamic time warping, which compares an input signal with a predefined template and quantifies similarity between both. Different templates were compared and an optimized template was created. The classification scored a F1-measure of 86.7% for evaluation on a data set acquired in a clinical setting. We believe that this approach can be transferred to home-monitoring systems and will facilitate a more efficient and automated gait analysis.

Quantifying Y Balance Test performance with multiple and single inertial sensors.

A growing body of evidence has highlighted that inertial sensor data can increase the sensitivity and clinical utility of the Y Balance Test, a commonly used clinical dynamic balance assessment. While early work has demonstrated the value of a single lumbar worn inertial sensor in quantifying dynamic balance control, no research has investigated if alternative (shank) or combined (lumbar and shank) sensor mounting locations may improve the assessments discriminant capabilities. Determining the optimal sensor set-up is crucial to ensuring minimal cost and maximal utility for clinical users The aim of this cross-sectional study was to investigate if single or multiple inertial sensors, mounted on the lumbar spine and/or shank could differentiate young (18-40 years [n = 41]) and middle-aged (40-65 years [n = 42]) adults, based on dynamic balance performance. Random-forest classification highlighted that a single lumbar sensor could classify age-related differences in performance with an accuracy of 79% (sensitivity = 81%; specificity = 78%). The amalgamation of shank and lumbar data did not significantly improve the classification performance (accuracy = 73-77%; sensitivity = 71-76%; specificity = 73-78%). Jerk magnitude root-mean-square consistently demonstrated predictor importance across the three reach directions: posteromedial (rank 1), anterior (rank 3) and posterolateral (rank 6).

Parametric evaluation of deep brain stimulation parameter configurations for Parkinson's disease using a conformal wearable and wireless inertial sensor system and machine learning.

Deep brain stimulation enables highly specified patient-unique therapeutic intervention ameliorating the symptoms of Parkinson's disease. Inherent to the efficacy of deep brain stimulation is the acquisition of an optimal parameter configuration. Using conventional methods, the optimization process for tuning the deep brain stimulation system parameters can intrinsically induce strain on clinical resources. An advanced means of quantifying Parkinson's hand tremor and distinguishing between parameter settings would be highly beneficial. The conformal wearable and wireless inertial sensor system, such as the BioStamp nPoint, has a volumetric profile on the order of a bandage that readily enables convenient quantification of Parkinson's disease hand tremor. Furthermore, the BioStamp nPoint has been certified by the FDA as a 510(k) medical device for acquisition of medical grade data. Parametric variation of the amplitude parameter for deep brain stimulation can be quantified through the BioStamp nPoint conformal wearable and wireless inertial sensor system mounted to the dorsum of the hand. The acquired inertial sensor signal data can be wirelessly transmitted to a secure Cloud computing environment for post-processing. The quantified inertial sensor data for the parametric study of the effects of varying amplitude can be distinguished through machine learning classification. Software automation through Python can consolidate the inertial sensor data into a suitable feature set format. Using the multilayer perceptron neural network considerable machine learning classification accuracy is attained to distinguish multiple parametric settings of amplitude for deep brain stimulation, such as 4.0 mA, 2.5 mA, 1.0 mA, and 'Off' status representing a baseline. These findings constitute an advance toward the pathway of attaining real-time closed loop automated parameter configuration tuning for treatment of Parkinson's disease using deep brain stimulation.

Classification Of Lower Limb Frontal Plane Excursion During Single-legged Landing Using Principal Component Analysis Of Inertia Sensor Data

PURPOSE: This study aimed to evaluate the temporal pattern of the frontal-plane lower limb translation during single-legged landing using the principal component analysis (PCA) of the inertial sensor data. METHODS: Eighteen legs from nine healthy women were tested. Participants performed a total of six single-legged landings for each leg. The wearable inertial sensor was fixed at the participant’s tibial tuberosity. The x-axis of the inertia sensor was aligned to the transverse axis of the shank segment, and the mediolateral acceleration of the proximal end of the shank segment was measured at 200Hz. The time window of interest was 0 to 75 ms after foot contact (15 data points). The acceleration data matrix (108 trials * 15 data point) from all trials of all participants was constructed for PCA. Eigenvalues and eigenvectors of the covariance matrix of the acceleration data matrix were calculated to identify the principal components (PC). RESULTS: The first two PCs explained more than 90% of total data variation (1st PC 69.7%, and 2nd PC 21.7%). The 1st PC represented the medially directed abrupt acceleration of the proximal shank, and the 2nd PC showed the slow but oscillated knee motion. The athlete of the highest PC score of the 1st PC showed a rapid medial knee motion just after foot impact, whereas the lowest one showed a stable knee motion after landing. CONCLUSIONS: PCA using acceleration signals successfully classified the degree of the medial knee translation after single-legged landing and screened the athletes who had a rapid medial knee translation. The abrupt increase of the medial knee translation has reported being a risk of non-contact anterior cruciate ligament (ACL) injury; therefore, the PCA based evaluation would be a useful technology for risk prediction of the non-contact ACL injury.

Reliability Of IMU Derived Kinematic Measures During The Forward Lunge

The integration of inertial sensor data into the clinical environment and beyond, demonstrates the potential of improved clinical interpretation of an individual’s functional movement capacity. Previously this has only been possible through the use of expensive laboratory-based motion capture systems. The forward lunge, a lower limb functional movement which exaggerates the gait cycle is commonly used by clinicians to assess lower limb strength, flexibility and balance. Combining Inertial measurement units with the forward lunge exercise could provide a quantified measure which was previously not possible in the clinical setting. PURPOSE: To determine the intrasession-reliability of kinetic measures derived from shank based inertial sensors during a forward lunge. METHODS: Twenty-three healthy participants took part in the study (12 Male, 11 Female, 30.8 ± 8.6yrs, 1.7 ±0.9m, 65.3 ± 10.8kg). Each participant performed 3 sets of 5 lunges bilaterally, each with a 10 min rest period adequate to establish intra-session reliability. Lunge distance and stance was set as 100% of leg length and hip width respectively (± 5%). IMUs were worn on the lateral aspect of each shank. The lunge was segmented into initiation, initial contact, mid-point, and termination. Peak & root mean squared (RMS) of total acceleration signals of the shank based IMU were taken for all lunges. Intraclass correlation coefficients (ICCs) were calculated based on a mean rating (k=3), absolute agreement two-way mixed effects model. Intra-session reliability was defined as poor (ICC<0.5), moderate (0.5-0.75), good (0.75-0.9) or excellent (>0.9). RESULTS: ICC values ranged from 0.916 to 0.981 for left limb peak acceleration and 0.903 to 0.978 for right limb peak acceleration. ICCs for RMS of left and right limb ranged from 0.908 to 0.979 and 0.899 to 0.977 respectively. CONCLUSIONS: The IMUs showed good to excellent reliability for both peak and RMS total acceleration across both limbs during the forward lunge. This demonstrated the potential for their integration as a clinical tool to provide quantified measures of an individual’s forward lunge performance.

Derivative Based Gait Event Detection Algorithm Using Unfiltered Accelerometer Signals.

Wearable sensors have been investigated for the purpose of gait analysis, namely gait event detection. Many types of algorithms have been developed specifically using inertial sensor data for detecting gait events. Though much attention has turned toward machine learning algorithms, most of these approaches suffer from large computational requirements and are not yet suitable for real-time applications such as in prostheses or for feedback control. Current rules-based algorithms for real-time use often require fusion of multiple sensor signals to achieve high accuracy, thus increasing complexity and decreasing usability of the instrument. We present our results of a novel, rules-based algorithm using a single accelerometer signal from the foot to reliably detect heel-strike and toe-off events. Using the derivative of the raw accelerometer signal and applying an optimizer and windowing approach, high performance was achieved with a sensitivity and specificity of 94.32% and 94.70% respectively, and a timing error of 6.52 ± 22.37 ms, including trials involving multiple speed transitions. This would enable development of a compact wearable system for robust gait analysis in real-world settings, providing key insights into gait quality with the capability for real-time system control.

CNN-Based Estimation of Sagittal Plane Walking and Running Biomechanics From Measured and Simulated Inertial Sensor Data.

Machine learning is a promising approach to evaluate human movement based on wearable sensor data. A representative dataset for training data-driven models is crucial to ensure that the model generalizes well to unseen data. However, the acquisition of sufficient data is time-consuming and often infeasible. We present a method to create realistic inertial sensor data with corresponding biomechanical variables by 2D walking and running simulations. We augmented a measured inertial sensor dataset with simulated data for the training of convolutional neural networks to estimate sagittal plane joint angles, joint moments, and ground reaction forces (GRFs) of walking and running. When adding simulated data, the root mean square error (RMSE) of the test set of hip, knee, and ankle joint angles decreased up to 17%, 27% and 23%, the RMSE of knee and ankle joint moments up to 6% and the RMSE of anterior-posterior and vertical GRF up to 2 and 6%. Simulation-aided estimation of joint moments and GRFs was limited by inaccuracies of the biomechanical model. Improving the physics-based model and domain adaptation learning may further increase the benefit of simulated data. Future work can exploit biomechanical simulations to connect different data sources in order to create representative datasets of human movement. In conclusion, machine learning can benefit from available domain knowledge on biomechanical simulations to supplement cumbersome data collections.