How AI Is Changing Running Form Analysis in Sports Science
A friend films you jogging past a park bench. Your left foot seems to land wider than your right, or perhaps that’s the camera angle. Replaying the clip doesn’t settle the question. AI-assisted analysis can turn footage into movement measurements, though interpreting those measurements takes more than a video.

Key Takeaways
- AI-assisted running analysis estimates body position from video. It measures how you move, not why.
- Single-camera tools measure within one plane. Rotation and side-to-side motion need a multi-camera 3D setup.
- Research supports markerless capture for lower-limb forward-and-backward joint motion during running, with weaker results for pelvic tilt, hip bending and joint rotation.
- Comparing sessions fairly means keeping camera placement, treadmill speed and software version consistent.
- A reviewer still needs your training history to make sense of the numbers.
From Watching a Stride to Tracking Movement
Running gait analysis is a movement assessment: a structured look at how your body moves while you run. It can include visual review, marker-based tracking or AI-assisted analysis. Marker-based systems track reflective markers attached to the body. Markerless software estimates body position from video without those markers.
Commercial markerless systems estimate three-dimensional body position from multiple synchronized camera views using deep learning. Several cameras film the same stride at the same instant, and the software estimates where the joints are in space.
Choosing a suitable method. Kinovea, a tool for annotating and comparing clips, can synchronize two videos even when they were recorded at different frame rates. Its documentation states that it is a 2D program and that all measurements are made on a single plane. A side-view comparison can help assess motion in that plane, but measuring hip rotation requires a method suited to three-dimensional movement.
What AI Can Reveal About Your Run
A multi-camera assessment starts with camera placement and calibration, which establishes a shared reference for position and distance. Synchronized cameras record the runner, and software estimates body landmarks across the views to reconstruct movement in three dimensions.
Analysis of that movement data can examine joint angles, stride timing and differences between sides. A reviewer might compare knee bend at foot strike or check whether a difference persists across repeated strides. A difference between sides alone doesn’t establish a problem.
One commercial example. Theia Markerless recommends at least eight synchronized cameras to track a person in a standard capture space of roughly 5 m by 5 m. Its explainer on video gait analysis describes how this kind of setup turns footage into 3D movement data. That recommendation applies to this type of system; camera needs vary by method.
Turning data into an interpretation. The same vendor notes that its software doesn’t include built-in report generation or biomechanical analysis. Instead, it exports movement data in standard formats for dedicated analysis software, and a reviewer must interpret those outputs in the context of the runner’s training. Some systems process video locally on the user’s own computer, while others send it to a cloud service. As with other camera-based AI tools, processing location is worth checking, along with who can access and retain the recordings.
Where Human Judgement Still Matters
Accuracy varies by measurement. Walker and colleagues tested a commercial markerless system with fourteen participants on a treadmill, including running at 10 and 12 km/h. Their findings supported its use for lower-limb sagittal measurements, meaning forward-and-backward motion at the leg joints, but not for pelvic tilt. Those findings apply to the tested conditions, not to every running speed or movement.
A systematic review of the same system by Varcin and Boocock found reasonable pooled accuracy for lower-limb movement, with poorer accuracy for hip bending and straightening and for rotation at the leg joints.
Training context. UPMC’s running assessment, for example, asks about footwear, upcoming races and training program before the team reviews the runner’s video in Kinovea. A video alone cannot tell a reviewer that you recently increased your mileage or changed shoes.
Fair comparisons between sessions. Keep camera height, distance and treadmill speed consistent when comparing recordings. Note the software version too. Markerless vendors periodically release new tracking models, and one widely used system received an update in July 2026. Even when a vendor reports that a new model produces consistent results, recording the version helps ensure a change in output isn’t automatically attributed to a change in running form.
Movement and force data. Markerless systems use computer vision to output kinematic data, which describes body position and joint angles. Direct force measurements require separate equipment, such as force plates or an instrumented treadmill, synchronized with the cameras so movement and force data line up in one session. Force figures calculated from a model should be labelled as estimates.
Before relying on a result, ask:
- What is being measured, and in which plane?
- What evidence supports this measurement for the running task being assessed?
- Who interprets the result, and what do they know about my training?
- Where is the video processed and stored, and who can access it?
Conclusion
AI has made it far easier to turn running footage into measurable movement data, and multi-camera markerless setups now capture three-dimensional motion without markers or a darkened lab. The research is encouraging for forward-and-backward motion at the leg joints, and more cautious for the pelvis, hip bending and rotation. The most useful results come from matching the method to the question, keeping recording conditions consistent, and pairing the data with someone who understands your training.
FAQ
Can a phone video analyze running form?
A single phone camera can support slow-motion review and 2D angle measurements in tools such as Kinovea. Those measurements are limited to one plane, so rotation and side-to-side motion need a multi-camera 3D setup.
How accurate is markerless motion capture for running?
Studies support markerless capture for lower-limb sagittal-plane measurements during running. Results are weaker for pelvic tilt, hip bending and straightening, and joint rotation, and accuracy depends on the system and testing conditions.
Does markerless motion capture measure force?
No. It measures movement. Direct force data comes from force plates or an instrumented treadmill synchronized with the cameras, and any force values calculated from a model should be treated as estimates.
How many cameras does 3D running analysis need?
It depends on the system and capture space. One commercial markerless system recommends at least eight synchronized cameras for a standard capture space of roughly 5 m by 5 m.
People also read this:Â Old-School vs High-Tech: Choosing the Right Car Security

