A general marker-less motion capture approach for the Microsoft Kinect Sensor v2 using subject-specific articulated models within an iterative closest point algorithm for improved anatomical accuracy.
Translated title
En general markør-fri motion capture metode til Microsoft Kinect Sensor v2 som benytter subjekt-specifikke artikulerede modeller i en iterativ tætteste punkt algoritme for at forbedre den anatomisk nøjagtighed.
Author
Gammelgaard, Mikkel Svindt
Term
4. semester
Education
Publication year
2016
Submitted on
2016-06-03
Pages
13
Abstract
Microsoft Kinect måler ledbevægelser ud fra en forenklet model af kroppen, som ikke har anatomisk nøjagtige ledcentre (de punkter, som knoglerne roterer omkring). Det gør målingerne svære at bruge direkte i detaljerede muskel‑skelet‑modeller. I dette studie foreslår vi en markørløs metode (uden refleksmarkører) med Kinect v2 til at indfange 3D‑bevægelsen af et enkelt ben. Vi byggede en personspecifik, ledet model af højre ben med anatomisk korrekte ledcentre ved at scanne segmenternes geometri med Kinect v2. Modellens stillinger blev tilpasset en dynamisk dybdemåling af en 90° hoftefleksion (løfte låret til ca. 90°) ved hjælp af en iterative closest point‑algoritme, dvs. en metode der justerer 3D‑former ved gentagne gange at matche nærmeste punkter, kombineret med en begrænset optimering. Vi sammenlignede den foreslåede metode med et markørbaseret system. Hos én forsøgsperson målte vi: hofte fleksion/ekstension, abduktion/adduktion (benet ud/ind til siden), indad-/udadrotation, knæ fleksion/ekstension samt ankel plantar-/dorsalfleksion (foden peger ned/op). Hofte fleksion/ekstension og indad-/udadrotation samt knæ fleksion/ekstension var i god overensstemmelse med markørsystemet, mens hofte abduktion/adduktion og ankel fleksion/ekstension viste dårlig overensstemmelse.
Microsoft Kinect measures joint motion using a simplified body model that lacks anatomically accurate joint centers (the points around which bones rotate). As a result, these measurements are not directly compatible with detailed musculoskeletal models. This study proposes a markerless Kinect v2 approach to capture the 3D motion of a single leg. We built a subject‑specific articulated model of the right leg with anatomically correct joint centers by scanning the segments’ geometry with Kinect v2. The model’s poses were aligned to a dynamic depth recording of a 90° hip flexion (raising the thigh to about a right angle) using an iterative closest point algorithm—i.e., repeatedly aligning 3D shapes by matching nearest points—combined with constrained optimization. We compared the proposed method with a marker‑based system. For one subject, we measured hip flexion/extension, abduction/adduction (leg moving out/in to the side), internal/external rotation, knee flexion/extension, and ankle plantar/dorsiflexion (foot pointing down/up). Hip flexion/extension and internal/external rotation, and knee flexion/extension showed good agreement with the marker‑based results, whereas hip abduction/adduction and ankle flexion/extension performed poorly.
[This abstract was generated with the help of AI]
Keywords
Documents