SUMMATIVE STATEMENT:
This paper studies the use of motion capture to record hand motions and the use of the random forest machine learning algorithm for classification of motion capture data into categories and subcategories of the HandPak ergonomics evaluation method.
KEYWORDS:
Ergonomics, Motion capture, Posture recognition, Hand evaluation, Random forest.
PROBLEM STATEMENT:
Nowadays, different technologies are available for ergonomics evaluations in the workplace. The use of technologies, such as camera-based or inertial motion unit sensors-based motion capture systems, facilitates measuring and digitalizing postures of humans over time. These motion capture systems are now being integrated in the processes of performing ergonomics evaluations in production systems to evaluate the well-being of the workers in a more efficient and objective manner (Rybnikár, Kačerová, Hořejší, & Šimon, 2023).
Ergonomics evaluation methods are commonly used in order to assess worker well-being. The use of motion capture systems enables automating assessments of posture related exposure criteria in ergonomics evaluation methods. However, the most commonly used ergonomics evaluation methods are based on observation. The observational ergonomics evaluation methods were initially created with the intention to provide a structure for risk assessment based on observations for ergonomists (Takala et al., 2010). The observational ergonomics evaluation methods therefore rely on the assessment made by ergonomist, and the criteria are often defined as a subjective measurement in the ergonomics evaluation method, often leading to subjective assessments not being consistent between different ergonomists (Nyman et al., 2023). At the same time, these subjective definitions of the criteria make it difficult to automate the criteria assessment with the data obtained from motion capture systems.
One ergonomics evaluation method for quantifying acceptable forces and torques on the forearm, wrist and hand is HandPak (Potvin, 2024). Performing a HandPak evaluation requires selecting one of the nine categories depending on the grip and force that the worker applies. Inside each category, it is necessary to define sub-categories of hand posture. For example, for the category “Torque: Wrist Flexion or Extension” there is a subcategory of “Type of Grip/Pinch” that needs to be classified in “Power Grip”, “Lateral Pinch” or “Pull Pinch”. This classification and subclassification is not easily quantifiable and cannot be defined as a logical set of rules from the joint angles of the fingers.
OBJECTIVE/QUESTION:
The objective of this article is to automatically recognize hand postures from motion capture data to help with the categorization of the hand postures in the HandPak (Potvin, 2024) ergonomics evaluation method.
METHODOLOGY:
In this paper we study the use of motion capture systems to record hand motions, and the use of the random forest (Cutler, Cutler, & Stevens, 2012) machine learning algorithm for classification of motion capture data into categories and subcategories of the HandPak ergonomics evaluation method. We created random forests for the categorization of three different hand postures based on a dataset of more than 10.000 data points.
RESULTS:
The study is ongoing and the results will be added in the full paper.DISCUSSION:The study shows that random forests can be used to classify hand postures based on hand joints angle data, coming from a motion capture system, into subcategories of the HandPak ergonomics evaluation method, without the overfitting issues that decision trees usually present. The study is limited in that it only considers three subcategories in the HandPak ergonomics evaluation method. Other subcategories in HandPak, such as the frequency or duration, present difficulties to be automated without manual input. In addition to that limitation, the training and test data was obtained from two subjects (a male and a female). Adding more subjects to consider variation of postures could improve the accuracy of the random forest model.
CONCLUSIONS:
The use of machine learning for categorization of hand postures enables partial automation of evaluation of criteria in ergonomics evaluation methods of hands such as HandPak (Potvin, 2024) that would otherwise require manual input. Reducing the need of manual input is argued to make the use ergonomics evaluation methods faster and less subjective.
REFERENCES:
Cutler, A., Cutler, D. R., & Stevens, J. R. (2012). Random Forests. In C. Zhang & Y. Ma (Eds.), Ensemble Machine Learning: Methods and Applications (pp. 157–175). New York, NY: Springer.
Douwes, M., & de Kraker, H. (2012). HARM overview and its application: Some practical examples. Work (Reading, Mass.), 41 Suppl 1, 4004–4009.
Nyman, T., Rhén, I.-M., Johansson, P. J., Eliasson, K., Kjellberg, K., Lindberg, P., Fan, X., et al. (2023). Reliability and Validity of Six Selected Observational Methods for Risk Assessment of Hand Intensive and Repetitive Work. International Journal of Environmental Research and Public Health, 20(8), 5505.
Potvin, J. R. (2024). HandPak. Retrieved March 11, 2024, from https://potvinbiomechanics.com/handpak/
Rybnikár, F., Kačerová, I., Hořejší, P., & Šimon, M. (2023). Ergonomics Evaluation Using Motion Capture Technology—Literature Review. Applied Sciences, 13(1), 162. Multidisciplinary Digital Publishing Institute.
Takala, E.-P., Pehkonen, I., Forsman, M., Hansson, G.-A., Mathiassen, S. E., Neumann, W. P., Sjøgaard, G., et al. (2010). Systematic evaluation of observational methods assessing biomechanical exposures at work. Scandinavian Journal of Work, Environment & Health, 36(1), 3–24.