Fig 1.
Marker-based motion capture (Mocap) versus video-based (OpenCap) analysis of human movement dynamics.
(Top row) Marker-based movement analysis usually occurs in a motion capture laboratory, and a comprehensive study of musculoskeletal dynamics typically requires more than two days of an expert’s time and equipment worth more than $150,000. (Bottom row) OpenCap enables the study of musculoskeletal dynamics in less than 10 minutes of hands-on time and with equipment worth less than $700 (assuming users need to purchase new mobile devices). OpenCap can be used anywhere with internet access and requires a minimum of two iOS devices (e.g., iPhones or iPads). (Right panel) OpenCap enables the estimation of kinematic, dynamic, and musculotendon parameters, many of which were previously only accessible using marker-based motion capture and force plate analysis.
Fig 2.
OpenCap comprises a smartphone application, a web application, and cloud computing.
To collect data, users open an application on two or more iOS devices and pair them with the OpenCap web application. The web application enables users to record videos simultaneously on the iOS devices and to visualize the resulting 3-dimensional (3D) kinematics. In the cloud, 2D keypoints are extracted from multi-view videos using open-source pose estimation algorithms. The videos are time synchronized using cross-correlations of keypoint velocities, and 3D keypoints are computed by triangulating these synchronized 2D keypoints. These 3D keypoints are converted into a more comprehensive 3D anatomical marker set using a recurrent neural network (LSTM) trained on a large motion capture dataset. 3D kinematics are then computed from marker trajectories using inverse kinematics and a musculoskeletal model with biomechanical constraints. Finally, dynamic measures are estimated using muscle-driven dynamic simulations that track 3D kinematics.
Table 1.
Mean absolute error (MAE) in kinematics and kinetics from OpenCap compared to laboratory-based motion capture and force plates.
Fig 3.
Medial knee loading during walking.
We evaluated how accurately OpenCap estimates the knee adduction moment (KAM), a measure of medial knee loading that predicts knee osteoarthritis progression, and how knee loading changes with a modified walking pattern. Participants (n = 10) walked naturally and with a trunk sway gait modification. (A) OpenCap estimated the early-stance peak KAM with r2 = 0.80, compared to an analysis using marker-based motion capture and force plates (Mocap). The KAM is normalized by bodyweight (BW) and height (ht). (B) The mean (bar) and standard deviation (error bar) across participants (circles) are shown for the changes in the peak KAM and peak medial contact force (MCF), which is a more comprehensive measure of medial knee loading, from natural to trunk sway walking (*P < .05). OpenCap detected the reductions in peak KAM and MCF (P < .006, t test and Wilcoxon signed rank test) that were measured with Mocap (P < .016, t tests). Furthermore, OpenCap correctly identified the one individual who did not reduce KAM or MCF as estimated by Mocap (filled circles).
Fig 4.
Distribution of lower-extremity joint moments when rising from a chair.
To evaluate OpenCap’s ability to detect between-group differences in dynamics, we computed differences in lower-extremity joint moments while rising from a chair that commonly exist between young and older adults. Individuals (n = 10) stood naturally and with increased trunk flexion, a strategy used by individuals with knee extensor weakness that shifts muscle demand to the hip extensors and ankle plantarflexors. (A) The mean (bar) and standard deviation (error bar) across participants (open circles) are shown for the changes in knee extension, hip extension, and ankle plantarflexion moments, averaged over the rising phase, from the natural to trunk flexion condition (*P < .05). Moments are normalized to percent bodyweight (BW) times height (ht). OpenCap identified the changes in joint moments (P = .004–.024, t tests) that were identified with motion capture and force plates (Mocap, P = .001–.002, t tests). (B) The rising-phase-averaged knee extension moment values for each participant and condition are shown. OpenCap estimated the knee extension moment with r2 = 0.65 compared to simulations that used motion capture and force plate data as input (Mocap).
Fig 5.
Asymmetry in vasti muscle activation during squatting.
To assess the utility of OpenCap for informing rehabilitation decisions, we sought to identify between-limb asymmetries in knee extensor muscle (vasti) function that indicate incomplete rehabilitation and relate to poor post-surgical functional outcomes. Participants (n = 10) performed squats naturally, then asymmetrically, where they were instructed to reduce the force under the left foot. (A, B) The mean (line) and standard deviation (shading) across participants are shown for the vasti muscle activation of the left (unweighted) leg measured with electromyography (EMG) and estimated using OpenCap. Muscle activations are normalized by the maximum value for each participant and measurement modality. (C) OpenCap identified peak vasti activation asymmetry between the left and right leg (asymmetry defined from EMG and clinically relevant symmetry threshold), with area under the receiver operator characteristic curve (AUC) of 0.83 and accuracy of 75%. This was similar to the performance of simulations that used marker-based motion capture and force plate data as input (Mocap sim., AUC = 0.82, accuracy = 70%).
Fig 6.
To demonstrate the practical utility of OpenCap for tracking rehabilitation progress, we enrolled 100 participants in a clinician-led field experiment. Participants performed symmetric squats and asymmetric squats, where they were instructed to reduce the force under the left foot, which likely resulted in an asymmetry between the left and right knee extension moments. We first evaluated the utility of OpenCap as a screening tool to detect peak knee extension moment asymmetries. (A) The distributions of knee extension moment symmetry indices for both squat conditions are shown, with a symmetry index larger than one indicating a lower peak knee extension moment for the left (unweighted) leg compared to the right leg. (B) OpenCap’s symmetry index estimates classified between natural and asymmetric squats with an area under the receiver operator characteristic curve (AUC) of 0.90 and accuracy of 85%. We then evaluated the utility of OpenCap for detecting changes in peak knee extension moment symmetry that would be expected to occur over time during rehabilitation. (C) The distributions of the average difference in the symmetry index between the asymmetric and natural conditions (i.e., hypothetical improved symmetry over time; red) and the average difference in the symmetry index between the three trials in the asymmetric condition (i.e., hypothetical unchanged symmetry over time; gray) are shown. (D) OpenCap detected improvements in symmetry from the asymmetric to the natural condition with AUC = 0.93 (improved compared to unchanged distributions from (C) and accuracy = 89%.