Accurate measurement of lower-limb kinematics in outdoor sports is challenging due to the limitations of marker-based motion capture and inertial sensors. This study evaluates a stereo camera-based 3D pose estimation pipeline using real-time multi-person one-stage (RTMO) detection to reconstruct sagittal knee and hip angles during running, sprinting and jumping. Twelve adult football players were recorded with two synchronised GoPro cameras, while inertial measurement units (IMUs) provided reference joint angles. RTMO extracted 2D keypoints from both views, which were triangulated to 3D joint positions for angle computation. During running at 2-4 m, agreement with IMUs was strong (r > 0.90, RMSE < 10 degrees). At 8 m, accuracy declined (RMSE 23.74 degrees, r = 0.80). Sprinting produced higher errors than running and jumping yielded the lowest error (RMSE 5.11 degrees, r = 0.98). Findings indicate that stereo-based pose estimation can capture joint kinematics outdoors, with accuracy dependent on distance and activity type.

Understanding the factors contributing to running-related injuries is important, since such injuries are common among runners. Vertical ground reaction forces (vGRFs) can help quantify biomechanical load during running, but are typically measured using force plates in laboratory settings. To allow continuous monitoring outdoors, wearable sensors such as inertial measurement units (IMUs) may offer a practical alternative. These sensors are often also embedded in smartwatches and heart rate monitor belts. This study aimed to predict vGRF using IMUs placed on the wrist and sternum-mimicking smartwatch and heart rate strap positions-combined with smartwatch-derived variables. Eleven rearfoot strike runners completed twelve 90-s treadmill trials at four speeds (8, 10, 12, 14 km/h) and three cadences (preferred ±10%). IMUs captured 3D acceleration and angular velocity; pressure insoles provided vGRF estimates used for outdoor validation. Subsequently, participants ran outdoors on an athletics track at the same four speeds. A long short-term memory (LSTM) neural network was trained to predict vGRF from IMU data. Feature importance analysis showed the sternum IMU contributed most to prediction accuracy. Leave-one-subject-out cross-validation, using all features, yielded an RMSE of 0.10 ± 0.024 body weight compared to indoor force plate measurements. Outdoor validation showed no significant performance drop. These results suggest that wearable devices enable meaningful vGRF monitoring during outdoor running.

This study presents a modular adaptation of the Drift-Free 3D Orientation and Displacement (DFOD) estimation method for estimating lower-extremity kinematics using independent inertial measurement units (IMU) during steady-state walking and running without calibration procedures or biomechanical models. The adapted DFOD was evaluated in ten healthy recreational runners during walking (2.5 and 5 km/h) and running (9, 11, and 13 km/h), using optical motion capture as reference. Good accuracy was achieved for the feet and lower legs in the sagittal plane (orientation) and forward direction (displacement), with mean errors below 7.2∘ and 4.2 cm, respectively, and Pearson correlations above 0.97. Accuracy was lower in other movement directions and for the upper legs, with mean errors up to 12.8∘ and 6.5 cm. However, Pearson correlations for the upper legs in the sagittal plane (orientation) and forward direction (displacement) exceeded 0.87, suggesting that waveform characteristics can be captured. Statistical analysis confirmed that movement axis and body segment were the dominant factors explaining estimation accuracy, with speed having a significant but smaller effect. Overall, larger errors were observed for movements with smaller RoM, for the upper legs, and for the slowest walking speed (2.5 km/h). These findings indicate that the adapted DFOD provides promising single-IMU-based estimates of lower-extremity orientation and displacement during walking and running for distal segments and primary movement directions.

Monitoring training load is an important aspect of optimizing performance and preventing overuse injuries in runners. This is the first study comparing physiological, biomechanical and subjective load between typical outdoor training sessions, contributing to the transfer of methodologies from the gait laboratory to real-world conditions with the final goal of improving athlete monitoring. Twelve experienced runners participated in distinct sessions: an endurance run, a submaximal effort, and interval training, which varied in perceived exertion. Using heart rate monitors, inertial measurement units and questionnaires, estimated cumulative load and its correlation with session Rate of Perceived Exertion (sRPE) and physiological load calculated via Training Impulse (TRIMP) were analysed. sRPE significantly distinguished between session types, while TRIMP and cumulative biomechanical load did not. Furthermore, correlations between the three training load metrics were weak to moderate (sRPE vs. TRIMP: r = 0.49; sRPE vs. weighted cumulative load: r = 0.25; weighted cumulative load vs. TRIMP: r = 0.35), where only sRPE and TRIMP correlated significantly (p < 0.05). This suggests that the different measures capture different aspects of load or that the measures could be inadequate to capture load. Objective physiological and biomechanical metrics alone may not adequately reflect athletes’ perceived exertion when training includes different session types. This highlights the importance of using a multifactorial approach to training load monitoring in running.

The moment of inertia (MoI) is an important parameter in biomechanical modeling that can influence the accuracy of kinetic estimations. The most commonly used approach to estimate MoI relies on anthropometric tables derived from a limited number of subjects, which may not account for subject-specific variability. This study evaluated the previously proposed angular momentum technique for estimating subject-specific body MoI in two healthy adult males. We were unable to obtain realistic MoI estimates using the angular momentum method, with MoI values being up to 36 times larger than reference values. Initial investigations revealed two promising alternative methods that yielded more realistic MoI estimates.

Running is a popular sport that offers health benefits but also poses a high risk of overuse injuries, often leading to a temporary or permanent cessation of running. These injuries are often caused by repetitive mechanical loading, emphasizing the need to quantify training loads to understand injury development. While such quantification is feasible in controlled settings like gait labs, capturing accurate data in natural outdoor conditions is challenging. This requires a practical sensor setup that is feasible for daily use and robust modelling approaches to relate sensor-derived data to mechanical load. We address these challenges by quantifying biomechanical loads during outdoor running using a minimal sensor setup. A key achievement is a model that uses inertial sensors to estimate vertical ground reaction forces (GRFs) used to estimate load. Moreover, we show that runners' perceived exertion is often misaligned with sensor-derived load measurements, highlighting the importance of combining subjective and objective metrics for a holistic view of training loads. The discrepancy also underscores the need to quantify structure-specific loads, such as forces on the lower legs, to better understand their potential role in mechanical fatigue and injury risk. We present a neural network model to estimate 3D GRFs needed to quantify these structure-specific loads accurately. These results advance biomechanics with novel data methods for quantifying training loads in outdoor environments with a feasible setup for many runners, enabling large future studies to better understand injury mechanisms. The study highlights the difficulty of data analysis when moving from a controlled situation in the laboratory to a large study outside the laboratory.

Inverse dynamics is a method to estimate joint forces and external moments needed for movement by analysing kinematics and ground reaction forces (GRF). In a bottom-up inverse dynamics analysis using a full-body inertial measurement unit (IMU) setup, the Centre of Pressure (CoP) is the only missing variable to complete the calculation. This study aimed to estimate the anteroposterior CoP from the tibia IMU orientation to calculate the sagittal ankle moment and tibial bone load (TBL) in rearfoot strikers running at 2.5, 3.1, and 3.6 m/s, using both tibia and the sternum IMU. This achieved strong correlations (≥0.90) for the CoP, sagittal ankle moment, and TBL compared with a marker/force plate reference. While the CoP estimate had fair accuracy, the sagittal ankle moment (rRMSE ≤ 12.9 %) and TBL (rRMSE ≤ 10.2 %) showed high accuracy. No significant differences were found between the IMU-only method and the reference for maximum ankle plantar flexion moment and TBL across all speeds. Future work should explore the multidimensional CoP, the inclusion of 3D GRF, and validation for non-rearfoot strike runners. These findings highlight the potential of using both tibia and the sternum IMU to monitor lower extremity forces and moments during running, independent of measurement location.

Drift remains a significant challenge in using inertial measurement units (IMUs) for human movement analysis. Drift reduction in orientation estimation is of particular interest, since inaccuracies in these estimates will negatively affect estimated linear kinematics (e.g., position). To address this, we developed an open-source toolbox designed to investigate the effects of signal characteristics, sampling frequencies, and integration orders on orientation estimation accuracy. The toolbox uses Taylor series approximations to estimate the change in orientation from angular velocity and contains two pipelines: a reference-based (RB) pipeline that compares the estimated orientation against a known ground-truth orientation, and a reference-free (RF) pipeline that does not rely on a ground-truth orientation. We demonstrate the toolbox's capabilities with three case studies. These case studies are also used to investigate the effects of signal characteristics, sampling frequency, and integration order on orientation estimation accuracy. Results show improved orientation estimation for slower movements, higher sampling frequencies, and higher integration orders. Additionally, through the case studies, we highlight how the toolbox can guide decisions on sampling frequencies and data processing strategies for specific application scenarios.

The net joint moment is a commonly investigated kinetic quantity in running but currently requires force plates and optical motion capture. This study proposes a physics-based top-down inverse dynamics method to estimate net sagittal knee and ankle moment across three speeds using only inertial measurement units (IMUs). This method does not require musculoskeletal modelling, machine learning, pressure insoles, or centre of pressure. The top-down method was validated against a 2D IMU-driven/3D marker-driven OpenSim model and an IMU-based bottom-up inverse dynamics approach. Strong correlations were found for the top-down net sagittal knee (0.87-0.96) and ankle moment (0.83-0.90) during stance. Maximum knee extension moment showed similar values during stance compared to IMU-based references, while maximum ankle plantar flexion moment was significantly higher. The marker-driven OpenSim model showed overall significantly lower values. This study highlights the potential of top-down inverse dynamics in calculating net sagittal knee moment during running using only IMUs, while the sagittal ankle moment was less accurate and needs a different approach. This method could potentially be used for running (i.e. providing feedback) during training sessions. However, a deeper understanding of upper body kinematics and kinetics is needed, as the top-down method is highly dependent on upper body movement.

To increase understanding in development of running injuries, the biomechanical load over time should be studied. Ground reaction force (GRF) is an important parameter for biomechanical analyses and is typically measured in a controlled lab environment. GRF can be estimated outdoors, however, the repeatability of this estimation is unknown. Repeatability is a crucial aspect if a measurement is repeated over prolonged periods of time. This study investigates the repeatability of a GRF estimation algorithm using inertial measurement units during outdoor running. Twelve well-trained participants completed 3 running sessions on different days, on an athletics track, instrumented with inertial measurement units on the lower legs and pelvis. Vertical accelerations were used to estimate the GRF. The goal was to assess the algorithm's repeatability across 3 sessions in a real-world setting, aiming to bridge the gap between laboratory and outdoor measurements. Results showed a good level of repeatability, with an intraclass correlation coefficient (2, k) of .86 for peak GRF, root mean square error of .08 times body weight (3.5%) and Pearson correlation coefficients exceeding .99 between the days. This is the first study looking into the day-to-day repeatability of the estimation of GRF, showing the potential to use this algorithm daily.

Approximately 40% of runners experience a running-related injury (RRI) each year. RRIs predominantly affect the lower extremities, and are often attributed to the repeated impacts that occur while running. Peak tibial acceleration (PTA) has been used as an indirect measure of these impacts experienced during running, and has been linked to RRIs in the lower extremities. Widespread ambulatory monitoring of runners using reliable measurements of impacts could enhance our understanding of RRIs. However, previous research suggests that PTA might be sensitive to factors such as the sampling frequency of the measurement device. The impulse has been suggested as a more reliable, alternative measure. In the current study, we evaluated the effect of sampling frequency on the determined PTA and impulse values using data collected from a high-rate, custom inertial measurement unit (IMU) placed on the tibia of one runner. The results suggested that both PTA and impulse are underestimated at lower sampling frequencies, with impulse being less affected by sampling frequency than PTA. Furthermore, both PTA and impulse showed an increase in underestimation for higher running speeds. These findings underscore the need for additional research to ensure accurate interpretation of ambulatory movement data, thereby potentially enabling a more comprehensive understanding of RRIs.

Training Load Management (TLM) is crucial for achieving optimal athletic performance and preventing chronic sports injuries. Current sports trackers provide runners with data to manage their training load. However, little is known about the extent and the way sports trackers are used for TLM. We conducted a survey (N=249) and interviews (N=24) with runners to understand sports tracker use in TLM practices. We found that runners possess some understanding of training load and generally trust their trackers to provide accurate training load-related data. Still, they hesitate to strictly follow trackers’ suggestions in managing their training load, often relying on their intuitions and body signals to determine and adapt training plans. Our findings contribute to SportsHCI research by shedding light on how sports trackers are incorporated into TLM practices and providing implications for developing trackers that better support runners in managing their training load.

Peak tibial acceleration (PTA) is a widely used indicator of tibial bone loading. Indirect bone loading measures are of interest to reduce the risk of stress fractures during running. However, tibial compressive forces are caused by both internal muscle forces and external ground reaction forces. PTA might reflect forces from outside the body, but likely not the compressive force from muscles on the tibial bone. Hence, the strength of the relationship between PTA and maximum tibial compression forces in rearfoot-striking runners was investigated. Twelve runners ran on an instrumented treadmill while tibial acceleration was captured with accelerometers. Force plate and inertial measurement unit data were spatially aligned with a novel method based on the centre of pressure crossing a virtual toe marker. The correlation coefficient between maximum tibial compression forces and PTA was 0.04 ± 0.14 with a range of -0.15 to +0.28. This study showed a very weak and non-significant correlation between PTA and maximum tibial compression forces while running on a level treadmill at a single speed. Hence, PTA as an indicator for tibial bone loading should be reconsidered, as PTA does not provide a complete picture of both internal and external compressive forces on the tibial bone.  .

Recent advances in wearable sensing and machine learning have created ample opportunities for “in the wild” movement analysis in sports, since the combination of both enables real-time feedback to be provided to athletes and coaches, as well as long-term monitoring of movements. The potential for real-time feedback is useful for performance enhancement or technique analysis, and can be achieved by training efficient models and implementing them on dedicated hardware. Long-term monitoring of movement can be used for injury prevention, among others. Such applications are often enabled by training a machine learned model from large datasets that have been collected using wearable sensors. Therefore, in this perspective paper, we provide an overview of approaches for studies that aim to analyze sports movement “in the wild” using wearable sensors and machine learning. First, we discuss how a measurement protocol can be set up by answering six questions. Then, we discuss the benefits and pitfalls and provide recommendations for effective training of machine learning models from movement data, focusing on data pre-processing, feature calculation, and model selection and tuning. Finally, we highlight two application domains where “in the wild” data recording was combined with machine learning for injury prevention and technique analysis, respectively.

Running is a popular activity that contributes to better physical and mental health. Running is, however, associated with high numbers of running related injuries (RRI) [1]. The etiology of these injuries is assumed to be multifactorial but still not fully clear. A recent systematic review showed that training load management and kinematic and kinetic factors are associated with RRIs, but few direct causal relations were found [2]. It is assumed that accumulating fatigue also plays an important role, negatively impacting the preferred movement pattern of a runner, leading to overload [3]. The main reason that the etiology of RRIs is still unclear is that most studies have been performed in the standardized and controlled gait laboratory. The limited measurement volume and/or the cyclical nature of the treadmill are not representative of the setting where people typically run and get injured. Therefore, experiments need to be performed in the real world to get a better understanding of running RRIs to eventually prevent them. Inertial magnetic measurement units (IMUs) allow for three dimensional assessment of (changes in) running technique in the sport specific setting.

Inertial measurement units are used in ambulatory human movement analysis, but their use is often subject to domain- or application-specific assumptions and methods to compensate for drift in kinematic estimations. Here, we propose and evaluate a generic drift reduction technique for orientation estimation. A second order Taylor approximation for 3D orientation estimation from sampled angular velocity is derived and evaluated on a publicly available dataset containing angular velocity data of 4 runners. The use of a second order Taylor approximation substantially reduces drift when the angular velocity has considerable contributions along all three axes. The second order Taylor approximation could therefore facilitate the use of minimal sensor setups for the study of dynamic 3D human movements.

Ground reaction force (GRF) is an important parameter for biomechanical analyses in running. For example, GRFs are used as input for inversed dynamics to estimate joint load or powers, which can give insights in the aetiology of running injuries. However, GRF measurement is restricted to the laboratory setting. Machine learning methods have been used to estimate GRF from inertial measurement units (IMUs), but creating robust generic models is challenging. An ensembled model, where multiple models are combined to create an estimate, could enhance the performance of (3D) GRF estimation models. PURPOSE: Ensemble multiple models to improve the overall performance to estimate 3D GRFs in running. METHODS: 11 experienced heel strike runners (4 F, 7 M, 30.6 ± 8.3 years, 1.79 ± 0.11 m, 74.2 ± 17.4 kg) ran 9 trials on a force-instrumented treadmill at a combination of three velocities (10, 12 and 14 km/h) and stride frequencies (preferred, ±10%). Subjects were instrumented with IMUs (240 Hz) at both proximal tibias and pelvis. Using leave-one-subject-out cross validation, artificial neural networks (two layers, both 250 neurons) were trained to estimate 3D GRF where the remaining subjects were randomly assigned to the training (n = 6) or validation (n = 4) set to create 7 different models. The ensembled model averaged the estimate of the 7 different ensemble members. For each subject, the ensembled model is compared to the 7 ensemble members in terms of relative root mean squared error (rRMSE), which is the RMSE normalized by the full range of the measured forces in that direction, multiplied by 100. RESULTS: The ensembled estimates performed better than the average performance over the ensemble members (Table 1). Furthermore, for 21 out of 33 GRF estimates, the ensembled model had similar or better performance than the best ensemble member. CONCLUSION: An ensembled model increases the performance of machine learning models to estimate 3D GRFs in running based on IMU data.

To understand the mechanisms causing running injuries, it is crucial to get insights into biomechanical loading in the runners' environment. Ground reaction forces (GRFs) describe the external forces on the body during running, however, measuring these forces is usually only possible in a gait laboratory. Previous studies show that it is possible to use inertial measurement units (IMUs) to estimate vertical forces, however, forces in anterior-posterior direction play an important role in the push-off. Furthermore, to perform an inverse dynamics approach, for modelling tissue specific loads, 3D GRFs are needed as input. Therefore, the goal of this work was to estimate 3D GRFs using three inertial measurement units. Twelve rear foot strike runners did nine trials at three different velocities (10, 12 and 14 km/h) and three stride frequencies (preferred and preferred ± 10%) on an instrumented treadmill. Then, data from IMUs placed on the pelvis and lower legs were used as input for artificial neural networks (ANNs) to estimate 3D GRFs. Additionally, estimated vertical GRF from a physical model was used as input to create a hybrid machine learning model. Using different splits in validation and training data, different ANNs were fitted and assembled into an ensemble model. Leave-one-subject-out cross-validation was used to validate the models. Performance of the machine learning, hybrid machine learning and a physical model were compared. The estimated vs. measured GRF for the hybrid model had a RMSE normalized over the full range of values of 10.8, 7.8 and 6.8% and a Pearson correlation coefficient of 0.58, 0.91, 0.97 for the mediolateral direction, posterior-anterior and vertical direction respectively. Performance for the three compared models was similar. The ensemble models showed higher model accuracy compared to the ensemble-members. This study is the first to estimate 3D GRF during continuous running from IMUs and shows that it is possible to estimate GRF in posterior-anterior and vertical direction, making it possible to estimate these forces in the outdoor setting. This step towards quantification of biomechanical load in the runners' environment is helpful to gain a better understanding of the development of running injuries.

Medicine & Science in Sports & Exercise 55(9S):p 825, September 2023. | DOI: 10.1249/01.mss.0000987600.80403.80