Deep learning applied to 2D video data for the estimation of clamp reaction forces acting on running prosthetic feet and experimental validation after bench and track tests

Carbon fiber Running Specific Prostheses (RSPs) have allowed athletes with lower extremity amputations to recover their functional capability of running. RSPs are designed to replicate the spring-like nature of biological legs: they are passive components that mimic the tendons elastic potential energy storage and release during ground contact. The knowledge of loads acting on the prosthesis is crucial for evaluating athletes’ running technique, prevent injuries and designing Running Prosthetic Feet (RPF). The aim of the present work is to investigate a method to estimate forces acting on a RPF based on its geometrical configuration. Firstly, the use of kinematic data acquired with 2D videos was assessed, to understand if they can be a good approximation to the golden standard represented by motion capture (MOCAP). This was done by evaluating steps acquired during two running sessions (OS1 and OS3) with elite paralympic athletes. Then, the problem was formulated using a deep learning approach, training a neural network over data collected from in vitro bench tests, carried out on a hydraulic test bench. Two models were built: the first one was trained over data from standard procedures and validated on two steps of OS1; then, in order to improve the performance of the prototype, a second model was built and trained with data from newly studied procedures. It was then validated on three steps from OS3.