Inverse Kinematic, Dynamic, and Muscle Simulation from Video Files: Toward Markerless 3D Human Pose, Torque, and Muscle Estimation

EasyChair Preprint 13337

Abstract

Optical skeletal motion capture technology is extensively employed across various domains, including entertainment, sports biomechanical analysis, healthcare, gaming, augmented reality, and human-computer interaction. Conventional commercial systems mandate the usage of marker suits resulting in usability limitations. In reaction, scholars have devised marker-less motion capture techniques enabling motion estimation in a broader range of settings through multi-view video analysis. Current leading pose estimation methods are heavily relying on deep learning techniques. These machine learning methods depend on pre-processed databases, which are typically generated through visual detection and manual annotation processes. The manual processes introduce inaccuracies, violation of biomechanical limits, variable body segment lengths, or leading to non-existent degrees of freedom (DoFs). This study presents a method for correcting the manual annotations by integrating a human musculoskeletal (MSK) model with International Society of Biomechanics standard.

A scalable MSK model using MapleSim was created to simulate human body dynamics, including skeletal dynamics and biomechanical joint torque. The model has 15 body segments with 40 DoFs and uses muscle torque generators (MTGs) with various characteristics. The simulation input involves videos with marker annotations, while the output includes joint coordinates, animation files, marker locations, and other biomechanical data. The model's accuracy in estimating joint positions and motions was evaluated, highlighting its potential for applications like motion analysis and healthcare. Physics-based methods like this are preferred for human pose estimation, avoiding unrealistic poses seen in purely kinematic approaches. Future work aims to enhance the simulation framework for diverse motions, possibly integrating neurological or control system elements for motion generation.

Keyphrases: human pose estimation, inverse dynamic simulation, multibody musculoskeletal modeling, optimization-based approaches

@booklet{EasyChair:13337,
  author    = {Ali Nasr and John McPhee},
  title     = {Inverse Kinematic, Dynamic, and Muscle Simulation from Video Files: Toward Markerless 3D Human Pose, Torque, and Muscle Estimation},
  howpublished = {EasyChair Preprint 13337},
  year      = {EasyChair, 2024}}