نوع مقاله : مقاله مروری
نویسندگان
1 دانشجوی دکتری بیومکانیک ورزشی، دانشکده علوم ورزشی، دانشگاه مازندران
2 دانشیار گروه بیومکانیک ورزشی، دانشکده علوم ورزشی، دانشگاه مازندران
3 دانشیار گروه آموزشی ریاضی، دانشکده ریاضی، دانشگاه مازندران
کلیدواژهها
موضوعات
عنوان مقاله English
نویسندگان English
Background and Purpose
In running analysis and evaluating athletes' performance, ground reaction force (GRF) components are essential parameters in estimating joint reaction forces and joint moments. In a biomechanics lab, Force platforms, instrumented treadmills and wearable sensors are the most common and reliable tools to measure GRF. However, using these tools require specific spaces such as motion analysis laboratories, skilled operators and they may alter foot-ground interaction and shoe stiffness. Furthermore, the mechanical and repeated stress applied to these sensors is high, so that the sensors can be easily worn or damaged. Because of these limitations, researchers have recently focused on indirect model in the estimation of GRF from kinematics data.
In biomechanical models, most of the proposed methods require modeling of the biomechanical system. This modeling, in turn, requires extensive knowledge of the subject's special parameters, such as mass of segments, dimensions, moment of inertia and etc. This inevitably brings inaccuracy and uncertainty. To calculate the mechanical parameters of each segment, extensive use of standard tables is required. Such statistical values may cause inaccuracies in the estimation of the desired values and they may not be generalizable to all societies.
In statistical models, many of the anatomical features of people are not considered. Therefore, this method may not be applicable for activities that involve repeated loads, such as running or during a training session. Furthermore, the accuracy of the estimated GRF by the statistical model is lower than that of other methods.
Recently, the use of machine learning methods, as a modern approach to the GRF estimation, are in the spotlight. These methods are based on the hypothesis that there is a relationship between the acceleration of each segment and the ground reaction forces. Machine learning methods do not require prior knowledge of model and build their own model using the training data obtained in previous experiments. Artificial neural network (ANN) is reported to be a good flexible tool for nonlinear modeling and very efficient for GRF estimation. Indeed, the use of neural networks simplifies modeling and data collection strategies. However, the disadvantages of ANNs include the sensitivity to the chosen input parameters, computationally expensive, as well as requiring a large amount of data for training the system to achieve acceptable accuracy.
According to the review conducted in the current study, most of the reviewed articles reported an acceptable estimation of the vertical component of GRF, while a few focused on its medial-lateral components. However, finding was attributed to the lower absolute values of the medial- lateral force components. Furthermore, while the correlation between body acceleration and vertical GRF was good in most cases, the absolute values of GRF were not correctly estimated. A combination of methods based on biomechanical modeling and machine learning, as well as the use of more sophisticated algorithms, seems to be a promising way to increase the overall accuracy, even in estimating the medial-lateral component of GRF.
Ground reaction force (GRF) components are critical parameters in running analysis, as they are essential for estimating joint reaction forces and joint moments, which are key to evaluating athletes' performance. Traditionally, GRF is measured using force platforms, instrumented treadmills, and wearable sensors, which are considered reliable tools in biomechanics laboratories. However, these methods require specialized spaces, skilled operators, and may alter natural foot-ground interactions and shoe stiffness. Additionally, the mechanical and repeated stress applied to these sensors can lead to wear and damage, further limiting their practicality. Due to these limitations, researchers have increasingly focused on indirect methods for estimating GRF from kinematic data.
Traditional biomechanical models for GRF estimation require detailed modeling of the biomechanical system, which involves extensive knowledge of subject-specific parameters such as segment mass, dimensions, and moments of inertia. This reliance on statistical data from standard tables can introduce inaccuracies and uncertainties, making the results less generalizable across different populations. Statistical models, on the other hand, often overlook many anatomical features, reducing their applicability for activities involving repeated loads, such as running or training sessions. Moreover, the accuracy of GRF estimation using statistical models is generally lower compared to other methods.
Recent advancements in machine learning (ML) methods have provided a promising alternative for GRF estimation. These methods are based on the hypothesis that a relationship exists between segmental accelerations and ground reaction forces. Unlike traditional approaches, machine learning methods do not require prior knowledge of the biomechanical model and can build their own models using training data from previous experiments. Artificial neural networks (ANNs), in particular, have been reported as flexible and efficient tools for nonlinear modeling, making them highly suitable for GRF estimation. However, ANNs are sensitive to the choice of input parameters, computationally expensive, and require large datasets for training to achieve acceptable accuracy.
A review of the current literature indicates that most studies report acceptable estimation accuracy for the vertical component of GRF, while fewer studies focus on the medial-lateral components. This discrepancy is attributed to the lower absolute values of medial-lateral forces, which are more challenging to estimate accurately. Although a strong correlation between body acceleration and vertical GRF has been observed in many cases, the absolute values of GRF are often not estimated correctly. To address these limitations, a combination of biomechanical modeling and machine learning methods, along with the use of more sophisticated algorithms, appears to be a promising approach for improving the overall accuracy of GRF estimation, including the medial-lateral components.
Machine learning methods, particularly ANNs, offer a promising solution to the limitations of traditional GRF estimation techniques. By leveraging the relationship between kinematic data and GRF, these methods can simplify modeling and data collection strategies. However, challenges such as sensitivity to input parameters, computational costs, and the need for large training datasets must be addressed to improve their accuracy and applicability. Future research should focus on integrating biomechanical modeling with machine learning and developing more advanced algorithms to enhance the estimation of both vertical and medial-lateral GRF components.
Article message
Machine learning methods, especially neural networks, are a promising alternative to traditional biomechanical models in estimating ground reaction force during running using kinematic data.
کلیدواژهها English