Direct estimation of the knee flexion angle from a monocular image during walking

Abstract

The recent use of deep learning techniques in biomechanical applications has opened new avenues for non-invasive image-based motion capture and assessment. With the growing importance of deep learning-based computer vision models, the healthcare innovation company Emovi aims to replace KneeKG-based MoCap Marker-based methods with markerless-based strategies to estimate and assess certain anatomical landmarks of the lower limbs. This thesis proposes a novel deep regression model built with a ResNet backbone enhanced with feature map constraints to directly estimate lower limb anatomical landmarks, specifically knee flexion angles; by enforcing spatial consistency through internal feature map supervision; from monocular RGB images of defined functional movement sequences. A dataset was collected in a controlled laboratory environment using the KneeKG system, which provides clinically validated kinematic reference data. Unlike prior methods that rely on multi-view setups or complex marker-based systems, this work introduces an approach to infer knee joint angles from a RGB image. This works also introduces image inpainting strategies to help with a more focused model training. Quantitative evaluations demonstrate a test mean absolute error of 2.4◦ in flexion angle estimation, indicating a potential use for clinical applications where cost-effective and accessible motion analysis tools are needed.

Date	18 Aug 2025
Original language	American English
Awarding Institution	École de technologie supérieure
Supervisor	Carlos Vázquez (Supervisor)