Loading...

Human Whole-Body Static 3D Posture Prediction in One- and Two-Handed Lifting Tasks from Different Load Positions using Machine Learning

Mohseni, Mahdi | 2023

74 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 55951 (08)
  4. University: Sharif University of Technology
  5. Department: Mechanical Engineering
  6. Advisor(s): Arjmand, Navid
  7. Abstract:
  8. Biomechanical models require body posture to evaluate the risk of musculoskeletal injuries during daily/occupational activities like manual material handling (MMH). The procedure to measure body posture via motion-analysis techniques is complex, time-consuming, and limited to equipped laboratories. This study aims to develop an easy-to-use yet accurate model that predict human whole-body static posture (3D body coordinates and anatomical joint angles) during different MMH activities. Twenty healthy male right-handed individuals with body mass index between 18 and 26 performed 204 symmetric and asymmetric MMH activities. Each person reached (i.e., without any load in hands) the destinations located at fifteen different horizontal and seven vertical (0, 30, 60, 90, 120, 150, and 180 cm from the floor) locations with one and two hands (i.e., load-handling technique) using all the possible lifting techniques (i.e., upright standing, stoop, semi-squat, and full-squat). Whole-body posture was measured via Vicon motion capture system and extended plug-in gait marker placement standard via 41 skin markers. Two multi-task artificial neural networks (ANNs) were trained to predict the 3D coordinates of these markers and 3D anatomical angles of 15 body joints considering seven simple input parameters including 3D hand coordinates, adapted lifting/load-handling techniques, and body height/weight using the collected in vivo dataset. Meanwhile, the performance of the developed ANNs were evaluated with both random hold-out (RH) validation and leave-one-subject-out (LOSO) cross-validation methods. Root-mean-square-error (RMSE), normalized-RMSE (nRMSE, i.e., RMSE calculated by dividing each output error to its in vivo measured range), and R2 of trained ANNs were, respectively, 25.7 mm, 5.1%, and 0.996 for 3D coordinates and 10.5°, 3.4 %, and 0.928 for 3D anatomical joint angles, using RH validation method. These values, in the same order, were 33.8 mm, 10.5%, and 0.990 for 3D coordinates and 15.1°, 10.2%, and 0.813 for 3D anatomical joint angles using the LOSO cross-validation method. Results indicate that the RMSE of ANNs for predicting body 3D coordinates and anatomical joint angles reduced by, respectively, ~36% and 61% compared to the previous studies. Using these novel posture prediction ANNs, human body motion analysis during MMH activities in real occupational environments is facilitated via a user-friendly approach
  9. Keywords:
  10. Artificial Neural Network ; Machine Learning ; Posture Prediction ; Manual Material Handeling ; Human Movment Analysis ; Applied Machine Learning ; Body Poses

 Digital Object List

 Bookmark

No TOC