Deep Learning-based Decoupling Method for 3D Plantar Force Detection Relevant to Human Gait

Plantar pressure and shear stresses are the 3D contact forces during foot-ground interactions, serving as the core parameters for dynamic gait analysis. Although many force measurement systems have been applied in clinical environment, only few can detect both plantar pressure and shear stress simultaneously. The main technical challenge is that the existing force platforms cannot decouple multidimensional forces effectively. Our recent work on the vision-based 3D plantar forces detection platform offers new possibilities. By utilizing deep learning (DL) techniques for forces decoupling our force platform is capable of achieving decoupling of the complex pressure and shear stresses. However, to date, there has been no in-depth research on performance of different DL approaches for decoupling such 3D forces relevant to gait conditions. This paper compared the forces decoupling capabilities of different DL models using a customized dataset of 4, 300 groups, which was collected by our dedicated force collection system. We evaluate the effectiveness of three classic image segmentation models such as FCN, U-Net, PSPNet and four image classification models such as ResNet, AlexNet, VGG16, and GoogleNet. A comparative analysis of all the above models was performed, encompassing various metrics such as accuracy and linearity of the decoupled multidimensional forces. After calibration, the highest measurement accuracy was down to 0.12 N/cm<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> for pressure measurement within the 42 N/cm<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> range and 0.058 N/cm<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> for shear stress within the 6 N/cm<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> range. In terms of linearity between predicted and true values, the VGG16 model demonstrates the best performance. Specifically, the R-squared (R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup>) value for pressure is 0.99, and 0.86 for shear stress. The findings of this study will lay the foundation for future research in optimizing force decoupling methods for vision-based 3D force detection systems.