A Deep Learning Model for Markerless Pose Estimation Based on Keypoint Augmentation: What Factors Influence Errors in Biomechanical Applications?

In biomechanics, movement is typically recorded by tracking the trajectories of anatomical landmarks previously marked using passive instrumentation, which entails several inconveniences. To overcome these disadvantages, researchers are exploring different markerless methods, such as pose estimation networks, to capture movement with equivalent accuracy to marker-based photogrammetry. However, pose estimation models usually only provide joint centers, which are incomplete data for calculating joint angles in all anatomical axes. Recently, marker augmentation models based on deep learning have emerged. These models transform pose estimation data into complete anatomical data. Building on this concept, this study presents three marker augmentation models of varying complexity that were compared to a photogrammetry system. The errors in anatomical landmark positions and the derived joint angles were calculated, and a statistical analysis of the errors was performed to identify the factors that most influence their magnitude. The proposed Transformer model improved upon the errors reported in the literature, yielding position errors of less than 1.5 cm for anatomical landmarks and 4.4 degrees for all seven movements evaluated. Anthropometric data did not influence the errors, while anatomical landmarks and movement influenced position errors, and model, rotation axis, and movement influenced joint angle errors.

Positioning errors of anatomical landmarks identified by fixed vertices in homologous meshes.

BACKGROUND: Human movement analysis is usually achieved by tracking markers attached to anatomical landmarks with photogrammetry. Such marker-based systems have disadvantages that have led to the development of markerless procedures, although their accuracy is not usually comparable to that of manual palpation procedures. New motion acquisition systems, such as 3D temporal scanners, provide homologous meshes that can be exploited for this purpose. RESEARCH QUESTION: Can fixed vertices of a homologous mesh be used to identify anatomical landmarks with an accuracy equivalent to that of manual palpation? METHODS: We used 3165 human shape scans from the CAESAR dataset, with labelled locations of anatomical landmarks. First, we fitted a template mesh to the scans, and assigned a vertex of that mesh to 53 anatomical landmarks in all subjects. Then we defined a nominal vertex for each landmark, as the more centred vertex out of the set assigned for that landmark. We calculated the errors of the template-fitting and the nominal vertex determination procedures, and analysed their relationship to subject's sex, height and body mass index, as well as their size compared to manual palpation errors. RESULTS: The template-fitting errors were below 5 mm, and the nominal vertex determination errors reached maximum values of 24 mm. Except for the trochanter, those errors were the same order of magnitude or smaller than inter-examiner errors of lower limb landmarks. Errors increased with height and body mass index, and were smaller for men than for women of the same height and body mass index. SIGNIFICANCE: We defined a set of vertices for 53 anatomical landmarks in a homologous mesh, which yields location errors comparable to those obtained by manual palpation for the majority of landmarks. We also quantified how the subject's sex and anthropometric features can affect the size of those errors.

Accuracy of a 3D temporal scanning system for gait analysis: Comparative with a marker-based photogrammetry system.

BACKGROUND: Combining the accuracy of marker-based stereophotogrammetry and the usability and comfort of markerless human movement analysis is a difficult challenge. 3D temporal scanners are a promising solution, since they provide moving meshes with thousands of vertices that can be used to analyze human movements. RESEARCH QUESTION: Can a 3D temporal scanner be used as a markerless system for gait analysis with the same accuracy as traditional, marker-based stereophotogrammetry systems? METHODS: A comparative study was carried out using a 3D temporal scanner synchronized with a marker-based stereophotogrammetry system. Two gait cycles of twelve healthy adults were measured simultaneously, extracting the positions of key anatomical points from both systems, and using them to analyze the 3D kinematics of the pelvis, right hip and knee joints. Measurement differences of marker positions and joint angles were described by their root mean square. A t-test was performed to rule out instrumental errors, and an F-test to evaluate the amplifications of marker position errors in dynamic conditions. RESULTS: The differences in 3D landmark positions were between 1.9 and 2.4 mm in the reference pose. Marker position errors were significantly increased during motion in the medial-lateral and vertical directions. The angle relative errors were between 3% and 43% of the range of motion, with the greatest difference being observed in hip axial rotation. SIGNIFICANCE: The differences in the results obtained between the 3D temporal scanner and the marker-based system were smaller than the usual errors due to lack of accuracy in the manual positioning of markers on anatomical landmarks and to soft-tissue artefacts. That level of accuracy is greater than other markerless systems, and proves that such technology is a good alternative to traditional, marker-based motion capture.