Biomechanics Technology: 3D Markerless Motion Capture and Biomechanical Modelling

Introduction and Learning Objectives

• Institutional Recognition: The University of Sydney acknowledges the Traditional Owners of Australia, specifically of the Gadigal people of the Eora Nation, and pays respect to Elders past, present, and emerging.

• Presenting Faculty: Associate Professor (A/Prof) Suzi Edwards, The University of Sydney.

• Learning Objectives (Week 5):

  • LO1: Understand the principles of data collection for biomechanics, specifically focusing on the principles of filming movements for quantitative analysis.

  • LO6: Gain the ability to conduct biomechanical assessments of movement techniques and communicate findings effectively to a lay audience.

3D Markerless Motion Capture (Mocap) and Pose Estimation

• Definition and Methodology: Markerless Mocap uses pose estimation via one or multiple synchronized cameras. These cameras record movements which are then processed through deep learning algorithms to perform pose estimation analyses for participants.

• Dependency on Training Data: The validity of biomechanical outcomes is strictly dependent on the information used to train the pose estimation algorithms. If the training data does not contain the specific movements of interest, pose detection accuracy may be significantly impaired.

• Primary Technologies:

  • High-speed video.

  • Red, Green, Blue, and Depth (RGBD) cameras.

  • Inertial Measurement Units (IMUs).

Red, Green, Blue & Depth (RGBD) Cameras: HumanTrak

• Hardware and Operation: Utilizes a laser pattern projector combined with two cameras. One camera records video while the second measures depth using infrared time-of-flight calculations.

• Kinematic Analysis: Acquires $3D$ joint centers to determine both sagittal and frontal plane angles. It uses "direct kinematics," where joint angles are calculated from segment vectors defined by estimated joint center positions.

• Application: Used by practitioners in clinical environments and strength and conditioning due to its portability and ease of setup.

• Cost: Lower than other systems, approximately $\$15,000$.

• Limitations:

  • Measures joint angles only at discrete events.

  • Low repeatability with location errors ranging from $30-50\,mm$ compared to marker-based joint centers.

  • Poor agreement with Vicon systems (ICC < 0.5).

• HumanTrak Pipeline and Time Requirements:

  • Preparation: Attaching camera to tripod ($2\,min$) and connecting to software ($2\,min$). Total: $4\,min$ (staff).

  • Participant Set-up: Not applicable ($0\,min$).

  • Data Collection: Loading profile ($0.5\,min$ per participant).

  • Data Analysis: Exporting data ($1\,min$).

  • Total Study Times: Active work time for a study of $22$ recorded trials is estimated at $135\,min$; total additional staff time per session is $5\,min$.

Inertial Measurement Units (IMUs)

• Components: Comprised of a gyroscope, magnetometer, and accelerometer (e.g., Xsens systems).

• Application: Used in sports, clinical settings, and for military/tactical personnel occupational demands.

• Commercial Example: APDM Opal IMUs, associated with Moveo Explorer software.

• Cost: Moderate, approximately $\$40,000$.

• Error Sources:

  • Higher measurement errors can result from unwanted sensor movement at attachment points.

  • Gyroscope data suffers from "drift" over time.

• APDM IMU Pipeline and Time Requirements:

  • Preparation: Synchronizing IMUs and configuring body sensor locations ($7\,min$ staff).

  • Participant Set-up: Attaching sensors with elastic bands and double-sided tape ($15\,min$ staff and participant).

  • Data Collection: Includes profile loading ($1\,min$), static calibration pose ($3\,s$ per trial), wireless transfer ($0.5\,min$ per trial), and resynchronization ($1\,min$ per trial). Total per trial: $1.55\,min$.

  • Data Analysis: Exporting for further analysis ($1\,min$).

  • Total Study Times: Active work time approximately $1046\,min$ for a typical study; additional participant time is $16\,min + 1.55\,min$ per trial.

High-Speed Video Markerless Systems: Theia

• Equipment: Requires $6$ to $8$ high-speed video cameras (e.g., Vicon FLIR, Qualisys Miqus).

• Calibration: Performed using a mocap wand or a proprietary Theia3D checkered calibration board.

• Cost: Moderate/high, approximately $\$80,000$.

• Users: Mayo Clinic, NBA, and various universities.

• Theia Data Collection and Analysis Pipeline:

  • Preparation: Camera setup on tripods ($20\,min$) and checkered board calibration ($10\,min$). Total: $30\,min$ staff.

  • Participant Set-up: Loading profile ($1\,min$).

  • Data Collection: $0\,min$ extra setup.

  • Data Analysis: Exporting video ($0.75\,min$ per trial), creating file paths ($1\,min$), and pose detection/export to Visual3D ($10\,min$ per trial).

  • Total Study Times: Active work time is $960\,min$. Passive computation time is significantly higher, estimated at $14,835\,min$ (approx. $10.75\,min$ passive work per trial).

Low-Cost and Specific Sports Applications

• OpenCap (Smartphone Video):

  • Open-source software developed at Stanford University using $2$ smartphones.

  • Cost: Free (Web application).

  • Caution: Uses cloud-based data upload; caution is advised for research as it may violate human ethics or USYD cybersecurity policies.

• Hawk-eye (Video):

  • Uses RGB video cameras to create stick figure models with $29$ body points.

  • Utilized by Major League Baseball (MLB) and Driveline Baseball.

• NBA Biomechanics Program Movements:

  • Double Leg Squat ($3$ reps).

  • Forward Lunge for max ankle motion ($2$ reps per side).

  • Single Leg (SL) Balance ($1$ rep, $20\,s$ each side, eyes closed).

  • Countermovement Jump (CMJ) for max height ($2$ reps).

  • SL CMJ for max height ($3$ reps per side).

  • Drop Vertical Jump ($30\,cm$ box) into rebound jump ($3$ reps).

  • Lateral Bound and return ($3$ reps each side).

Validation Metrics: ICC, RMSE, and Sampling Rates

• Accuracy Challenges: Difficult to achieve precise accuracy due to soft tissue movement artifact (STA) relative to bone movement and participant clothing obscuring landmarks.

• Intraclass Correlation Coefficient (ICC):

  • Defines how strongly groups of quantitative data resemble each other (between-day or between-system reliability).

  • An ICC > 0.9 indicates the measure is highly repeatable.

• Root Mean Squared Error (RMSE):

  • Represents the magnitude of error (accuracy) between time-domain signals.

  • Formula:     $RMSE = \sqrt{\frac{\sum_{i=1}^N (\text{predicted}_i - \text{observed}_i)^2}{N}}$

  • Lower RMSE indicates better accuracy. A joint angle $RMSE \le 5^\circ$ and $ICC \ge 0.75$ are usually considered valid for recommending future use.

• Empirical Finding: In one comparison, only two joint angle measures for Theia reached acceptable validity ($RMSE \le 5^\circ$, $ICC \ge 0.75$), while APDM and HumanTrak failed to reach these thresholds for any joint angles across recorded movements (e.g., lunge).

• Sampling Rate and Nyquist Theorem:

  • Comparison of $128\,Hz$ vs. $100\,Hz$. Insufficient sample rates lead to blurred images and violate the Nyquist sampling theorem.

  • Sample rates must be selected based on the fastest movement performed.

Smallest Worthwhile Change (SWC) and Minimal Detectable Change (MDC)

• Definition: The standard deviation multiplied by a Cohen effect size of $0.2$.

• Effect Size Scale: Small change ($0.2$), moderate ($0.6$), large ($1.2$).

• Formula:     $SWC = \text{standard deviation} \times 0.2$

• Numerical Example: If the standard deviation (SD) for the early Rate of Force Development (eRFD) is $2200\,N/s$:     $SWC = 2200\,N/s \times 0.2 = 440\,N/s$

• System Specifics:

  • Theia can evaluate sagittal plane ankle angles in lunges, but APDM and HumanTrak cannot.

  • APDM errors often stem from sensor-to-segment misalignment (mitigated by static or dynamic calibration).

  • HumanTrak’s frontal view limits sagittal plane accuracy but improves frontal plane tracking.

3D Biomechanical Modelling: Principles and Geometric Shapes

• Terminology: Modelling is necessary to interpret 3D optoelectronic (passive) mocap data.

• Historical Models: Helen Hayes, Vicon Plug-in-Gait (PiG), UWA Lower limb model (McLean & van den Bogert), Oxford Foot model ($50$ markers on lower limb/pelvis).

• Step 1: Geometric Modeling: The body is modeled as standard shapes:

  • Right circular ellipsoids.

  • Spheres.

  • Frustums of right circular cones (Hananan Geometric model).

• Step 2: Estimation of Inertial Properties: Each shape possesses:

  • Mass.

  • Length.

  • Center of Mass (COM).

  • Radius of Gyration ($k$).

• Study Foundations: Estimates based on cadaver studies from Dempster ($1955$), Zatsiorsky et al. ($1990$), de Leva ($1996$), and Pearsall et al. ($1996$).

Estimation of Segment Inertial Properties and Calculations

• Parallel Axis Theorem: Most body segments do not rotate around their COM but around a joint at either end, causing an "eccentric axis." Rotation about an eccentric axis increases the Moment of Inertia (MOI).

• Equation:     $I = I_0 + mx^2$

  • $I_0$: Moment of inertia about the center of mass.

  • $x$: Distance between the COM and the center of rotation.

• Segment Mass Calculations (as percentage of total body mass, e.g., for an $80\,kg$ participant):

  • Foot ($0.0145 \times 80\,kg = 1.16\,kg$).

  • Leg/Shank ($0.0465 \times 80\,kg = 3.72\,kg$).

  • Thigh ($0.100 \times 80\,kg = 8.0\,kg$).

• Segment Length: Calculated as a percentage of height.

• Location of COM: Calculated as a percentage of segment length.

  • COM of foot: $0.50 \times \text{segment length}$.

  • COM of leg: $0.433 \times \text{segment length}$.

  • COM of thigh: $0.433 \times \text{segment length}$.

• Moment of Inertia Equation ($I_0$):     $I_0 = mk^2$

  • Example for Leg (Length $0.435\,m$, Mass $3.72\,kg$):

    • About COM: $3.72\,kg \times (0.435\,m \times 0.302)^2 = 0.064\,kg \cdot m^2$.

    • Proximal end: $3.72\,kg \times (0.435\,m \times 0.528)^2 = 0.196\,kg \cdot m^2$.

    • Distal end: $3.72\,kg \times (0.435\,m \times 0.643)^2 = 0.291\,kg \cdot m^2$.

Muscle, Tendon, and Subject-Specific Modelling

• Force and Moment Arms: Can be estimated via regression equations (Herzog & Read $1993$, Menegaldo $2004$).

• Moment Arm Regression Equation: Expressed as a function of knee angle ($\theta$):     $\text{Moment Arm} = B_0 + B_1\theta + B_2\theta^2 + B_3\theta^3 + B_4\theta^4$

• Subject-Specific Modelling: Uses EMG and MRI imaging to capture unilateral musculoskeletal pathologies or geometric differences between limbs that generic models miss.

• HAMI (Hamstrings & Adductor Myotendinous Injuries): A specific project utilizing 3D Mocap and subject-specific modelling to account for these differences.