Video Gait Analysis: How AI Powers Markerless Motion Capture

Reading Time: 10 minutes

When evaluating software for a gait or movement lab, start with the decisions the data needs to support, not which system sounds most advanced.

A sports performance director comparing stride timing week to week does not need the same setup as a clinician-researcher studying joint moments. A teaching lab has different constraints again. The right software depends on how much measurement error your decisions can tolerate, how many people you need to assess each day, and which tools must work with the output.

This guide covers what motion analysis measures, where 2D video differs from 3D capture, and how marker-based and AI-driven markerless systems work. Theia3D provides a practical example of how multi-camera video becomes 3D movement data. A use-case shortlist and three-week pilot plan help you assess the options in your own setting.

Key Takeaways

• Cameras measure motion, not force directly. Video supports measurements of positions, angles, and timing. Force-related outputs require force sensors or modeling assumptions that need separate validation.
• 2D video is useful within limits. A single view can support timing and selected angle measurements, but it cannot fully resolve movement outside the image plane.
• Accuracy depends on the joint, plane, and task. Agreement during walking does not establish accuracy during running or cutting.
• Multi-camera capture needs calibration and synchronization. Theia3D combines calibrated, synchronized views to reconstruct movement in 3D.
• Test the export workflow. Check file formats, included variables, and compatibility with your biomechanics or statistics tools.
• Pilot the system before committing. Repeated trials on your own tasks can reveal measurement and workflow problems that a demonstration may miss.

What Motion Analysis Actually Measures

Motion analysis involves two broad categories of output. Knowing the difference helps you choose the right equipment and interpret its results.

Kinematics describes movement without explaining the forces that cause it. Outputs include joint positions, segment orientations, joint angles, movement speeds, step length, cadence, and stance time. Camera-based systems derive these measurements by tracking reflective markers or detecting anatomical landmarks in video.

Kinetics describes forces and moments, or turning effects, involved in movement. A key measure in gait is ground reaction force: the force the ground applies to the foot during contact. Force plates measure this directly. To calculate joint moments and powers, labs commonly combine motion data with force measurements through inverse dynamics, a calculation that also uses a model of the body.

Some video-based systems estimate forces and joint loading through modeling. These estimates are not direct force measurements, and their suitability depends on the assumptions and validation behind them.

If your decisions depend on measured loading or force asymmetry, plan for force-measuring hardware and software that aligns it with motion data. If you need angles and timing, a kinematics-only workflow may be sufficient.

2D Video vs 3D Motion Capture

Single-camera 2D video is a practical starting point for coaches and clinics. A consistent side view can support timing measures and selected sagittal-plane angles, such as knee flexion at initial contact. The sagittal plane describes forward-and-back movement, including bending and straightening. Distance measurements also need a suitable scale reference. Kinovea, a free, open-source video analysis application, supports basic movement annotation and measurement.

As Theia3D’s overview explains, a single camera projects movement onto a flat image. A knee that appears to bend cleanly from the side may also rotate or move inward in ways that view cannot resolve. Apparent angles and distances can change as a limb moves toward or away from the lens.

Multi-camera 3D motion capture combines calibrated views to reconstruct positions in three dimensions. This removes the single-plane constraint, but it does not make every measurement equally reliable. Accuracy still varies by joint, movement plane, and task.

2D may be enough when:

• The main purpose is education, movement demonstration, or basic coaching feedback.
• You need timing or selected visible-plane measurements rather than full 3D joint motion.
• The camera can view the movement square-on, with consistent positioning across sessions.
• You can clearly state the method’s limits when reporting results.

A front view can show some side-to-side movement, but it still provides a projection rather than a complete 3D measurement. If a decision depends on rotation or movement across several planes, use a validated 3D method.

Marker-Based vs AI Markerless

Traditional optical 3D capture uses reflective markers placed on anatomical landmarks and tracked by infrared cameras. Systems using Vicon Nexus, Qualisys Track Manager (QTM), and OptiTrack Motive have established workflows and are commonly used as comparators when evaluating newer methods.

The practical costs include participant preparation and skilled marker placement. Differences between operators can affect results, and skin movement can shift markers relative to the underlying bones. Hidden markers create gaps that may need processing. Participants usually need fitted clothing or exposed skin around marker locations.

AI-driven markerless systems use trained models to detect anatomical landmarks in video instead. This reduces marker-placement work and can allow participants to wear ordinary clothing. Theia3D uses synchronized multi-camera footage to reconstruct movement in 3D without attaching markers to the participant.

The tradeoff is that the software estimates landmark locations from appearance, the same inference problem that shows up wherever camera-based sensing replaces direct measurement. These estimates may differ systematically from marker-based definitions.  Performance depends on the model’s training, camera coverage, clothing, and how clearly each body segment is visible. Occlusion, when one body part or object blocks another, still matters.

Many validation studies report closer markerless and marker-based agreement for sagittal-plane angles than for side-to-side or rotational movement. However, the pattern is not a guarantee for every system or task. Check results for the joints and movements you intend to measure.

What not to assume about accuracy

A headline error figure does not apply to every joint. Hip rotation and knee flexion are different measurements.

Agreement for bending and straightening does not establish agreement for side-to-side movement or rotation.

Results for walking do not automatically apply to running, cutting, or stairs. Speed and visibility change the measurement challenge.

How AI Markerless Motion Capture Works

Understanding the main processing steps helps you assess what drives measurement quality. Theia3D’s documentation provides a useful example of a multi-camera markerless workflow.

1. Multi-camera capture. Theia3D uses synchronized cameras positioned around the capture space. Camera count and placement should follow the recommended layout for the volume and tasks. Each body segment needs to remain visible from several angles throughout the movement.

2. Camera calibration. Intrinsic calibration describes each camera’s lens properties, including distortion. Extrinsic calibration establishes the cameras’ positions and orientations in a shared coordinate system. Theia3D’s calibration guidance addresses both, supporting the reconstruction of consistent 3D coordinates. In practice, calibration requires deliberate coverage of the space where participants will move, not just its center.

3. Time synchronization. Views used together must represent the same moment. Hardware synchronization coordinates capture; simply starting recordings together is not enough. Theia3D’s synchronization tools help identify frame offsets between views so timing problems can be checked before downstream analysis.

4. 2D landmark detection. A trained model identifies anatomical landmarks in each video frame. Theia3D detects these keypoints across camera views, providing the image locations used for 3D reconstruction.

5. Reconstruction in 3D. The system combines corresponding detections from calibrated, synchronized views to estimate landmark positions in three dimensions. Multiple clear views help constrain each estimate.

6. Model fitting and smoothing. A skeletal model is fitted to the reconstructed landmarks to calculate segment positions and joint angles. Filtering can reduce frame-to-frame noise, but it can also affect peaks and timing. Record the settings and keep them consistent across comparable sessions.

7. Export. Theia3D supports C3D, FBX, and JSON exports, along with a handoff to Visual3D for further biomechanical analysis. FBX export requires the pose to be solved with the full body model, so confirm that fits your workflow if animation output matters. Check which outputs your intended analysis requires.

Theia3D’s video gait analysis overview offers further detail on the capture and reconstruction workflow. Pair that workflow guidance with independent, task-specific research when assessing measurement performance.

Hardware and Environment Checklist

Software is only part of the system. Review the following requirements before choosing a multi-camera setup:

• Cameras: Match frame rate to movement speed, resolution to capture distance, and lenses to the area you need to cover.
• Coverage: Keep body segments visible from several angles across the full movement. Theia3D’s capture guidance can help plan camera coverage for the intended space and tasks.
• Lighting: Provide even illumination and check for flicker. Infrared marker tracking and color-video systems have different lighting requirements.
• Computer: Confirm the supported workstation and graphics processor specifications, along with expected processing times.
• Storage: Multiple video streams create large files. Plan capacity, transfer speed, backup, and retention before the first session.
• Synchronization: Use a supported method for coordinating capture and a documented way to check timing afterward.
• Capture space: Allow room for movement, secure camera mounts, and calibration throughout the usable area.

Treadmill vs overground: A treadmill concentrates the capture area and allows many strides per trial, which can simplify coverage. Overground capture measures walking along a walkway but needs enough space for the intended task. Choose the condition that matches your question rather than treating the two as interchangeable. Force plate and muscle-activity measurements also need compatible hardware and synchronized data.

Data Portability and Analysis Workflows

Even useful measurements can create extra work if they are difficult to export. Ask for a sample dataset early and open it in the tools your team already uses.

C3D is widely used in biomechanics, FBX serves animation and visualization workflows, and JSON works well with custom scripts. Theia3D supports these formats and a direct handoff to Visual3D. For any system, confirm the contents of each export: a supported file extension alone does not guarantee that the required variables, units, or model definitions are included.

A usable export should come with:

• A consistent file format and a clear list of included variables and units.
• A documented coordinate system, including the vertical axis, direction of travel, and origin.
• Filtering details, including whether the data is raw or smoothed.
• A naming convention covering participant ID, session date, task, and trial number.
• Calibration and synchronization records saved with the trial data.

Validation and Repeatability: What to Look For

An accuracy figure is useful only when you know how it was measured. Look for the tested population, task, speed, joints, movement planes, and comparison method.

Pay attention to systematic differences as well as average error. Two systems may produce similarly shaped angle curves while reporting different peak values. Marker-based comparisons are useful, but the reference method also has measurement limits.

Repeatability matters for tracking change. If repeated sessions produce large differences without a meaningful change in the participant, small week-to-week changes may be hard to interpret. Include repeated trials and separate-day testing in your pilot, then compare the measurement variability with the size of change that matters to your work.

Ask each vendor:

1. What are the joint-specific errors by movement plane and task, and which participants and comparison methods were used?
2. How does the system flag blocked camera views, missing frames, or poor reconstruction?
3. Which published evaluations were conducted by researchers independent of the company?

Shortlist by Use Case

Different methods suit different measurement needs. Use these categories to narrow a shortlist rather than rank every system on one scale.

2D video tools, such as Kinovea and Dartfish. Useful for education, coaching feedback, and selected timing or visible-plane measurements. Costs range from free software to paid subscriptions. These tools cannot support full 3D conclusions from a single view.

Marker-based systems, such as Vicon Nexus, Qualisys QTM, and OptiTrack Motive. Suitable for labs that need established measurement protocols and hardware integrations. Plan for trained operators, participant preparation, and checks for marker visibility and placement consistency.

Multi-camera AI markerless systems, including Theia3D. Reduce marker-placement work and support capture in ordinary clothing. The preparation shifts toward camera coverage, calibration, synchronization, and checking that the participant remains clearly visible. Test processing time and exports as part of the evaluation.

Wearable inertial systems, such as Xsens MVN. Body-worn sensors support capture without a fixed camera installation. Check sensor placement, drift over longer recordings, and sensitivity to the environment. Position estimates and correction methods vary, so confirm that the outputs meet your needs.

Smartphone-based markerless tools, such as OpenCap. Combine video, pose estimation, and body modeling to estimate 3D movement and, in some workflows, dynamics. Force-related outputs are model-based estimates rather than direct force measurements. Review validation for your population and tasks.

Budget, Licensing, and Privacy

Compare total system cost, not just the software price. Free 2D tools can cover basic needs, while paid tools may use subscriptions. Many 3D systems require a quote. Request an itemized budget covering cameras, synchronization hardware, mounting, lighting, workstation, storage, training, support, and license renewal.

Data handling deserves equal attention. Participant video can be identifiable, and processing location may matter to ethics boards and institutional IT teams. Theia3D describes local, on-premises processing without transmitting participant videos or results to the vendor or third-party clouds. Confirm the arrangements for your installation, including remote support, access controls, backups, and retention. For other systems, establish whether cloud processing is optional or required.

Licensing also affects capacity. Check whether capture and processing can run at the same time, how many workstations are permitted, and whether additional staff need separate licenses. A lower subscription price may not help if processing delays limit daily sessions.

Integrations to Plan From Day One

If you need kinetics, test the complete route from force measurement to analysis. Some capture platforms connect directly to force plates and electromyography (EMG) equipment, which records muscle electrical activity. Other workflows combine exported motion and force data in a separate biomechanics application. Confirm support for your exact hardware rather than relying on a general compatibility claim.

Time alignment is critical. Motion and force measurements need a shared timing reference or a validated alignment method so calculations use data from the same instant. Sampling rates may differ, so check how the analysis software handles them. A successful pilot should produce the final variables you need, not merely import both files.

Run a Three-Week Pilot

A short, structured pilot can reveal whether the system fits your tasks and workflow. One workable plan is:

• Week 1: Define and prepare. List the required variables by joint, plane, and task. Set acceptance criteria for measurement variability and processing time. Set up the cameras, calibrate the full capture area, and verify synchronization. Record the settings.
• Week 2: Capture. Record a small group performing the intended tasks at realistic speeds, with repeated trials. Include separate-day testing where possible. Test a challenging condition, such as fast movement or partial occlusion, within the system’s intended use.
• Week 3: Analyze and decide. Export data into your normal analysis tools. Assess repeatability for each required variable and agreement with a reference method if one is available. Record failed trials, manual cleanup, processing time, and unresolved integration issues.

Keep the conclusions within the pilot’s scope. Motion-only recordings cannot validate direct force measurement. If you plan to use model-estimated loading, evaluate that output and its assumptions separately.

Common Pitfalls and Easy Wins

Common problems to check for include:

• Calibration that covers the center of the capture area but misses its edges.
• Small timing offsets between cameras that go unchecked.
• Drawing 3D conclusions from a single side-view recording.
• Combining differently filtered data within one comparison.

A few simple habits help prevent these problems:

• Save calibration files and synchronization records with each session.
• Check synchronization after capture, even when hardware synchronization is used.
• Record the actual task at the intended speed, not a slowed-down substitute.
• Standardize file names, units, and export settings before collecting participant data.

Match Outputs to Decisions

Choosing motion analysis software starts with the measurement question. Work backward from the required outputs, acceptable error, daily workload, and essential integrations. Marker-based systems offer established laboratory workflows. Theia3D supports multi-camera 3D capture without participant marker placement, with camera coverage, calibration, and synchronization central to the setup. Wearable and smartphone-based methods offer other ways to capture movement, each with distinct measurement limits.

The most useful evaluation is an end-to-end test: capture your tasks, produce your final outputs, and check whether the results are reliable enough for the decisions you intend to make.