How AI Analyzes Your Running Form, Step by Step

The short version: four steps from video to feedback

AI running-form analysis turns an ordinary phone video into coaching feedback in four steps. First it finds your body in each frame. Second it measures the angles between your joints. Third it tracks how those angles change over time to break your run into phases. Fourth it compares your patterns against what good running looks like and flags the differences. No sensors, no suit, no markers stuck to your skin, just video and software. Understanding those four steps tells you both why the analysis works and why how you film changes the result.

Step 1: pose estimation finds your joints

The foundation is pose estimation, a type of deep-learning model trained on huge labeled image datasets to locate human joints, the ankles, knees, hips, shoulders, elbows, in a photo or video frame. Tools you may have heard of, such as OpenPose, MediaPipe, MoveNet, and YOLO-pose, all do this job: they output a set of keypoints, essentially a stick-figure skeleton laid over your body, in every single frame. Because it works straight from normal video, this approach is called markerless motion capture, in contrast to the lab method of taping reflective markers to an athlete and filming with specialized cameras.

Step 2: joint-angle detection

Once the model has your joints in each frame, it does geometry. Three keypoints, say hip, knee, and ankle, define an angle, so the software can compute your knee flexion, hip extension, trunk lean, and arm angle at every instant of the stride. This is where the useful numbers come from: how much you flex at the knee on contact, whether your trunk leans the right amount in the drive phase, how high your knee drives. Research shows two-dimensional joint angles from pose estimation land within roughly 3 to 11 degrees of lab marker systems, accurate enough to see real form patterns, though not perfect to the degree.

One detail explains a lot about what follows. A single camera sees the world flat, in two dimensions. So it measures angles cleanly only in the plane it is pointed at. That is exactly why coaches and apps insist you film side-on: the side view is the plane where stride mechanics live, so that is the plane the model can read accurately.

Step 3: phase segmentation

A stride is not one thing; it is a repeating cycle of phases, ground contact and flight, or in a sprint, the drive, transition, and max-velocity phases. By tracking how your joint angles change frame to frame, the software can find the moments that divide those phases: when your foot strikes, when it leaves the ground, when your knee reaches its highest point. This is phase segmentation, and it is what lets the analysis say something specific like your knee drive drops off in the transition rather than a vague your form needs work. Peer-reviewed work confirms the structure the model is detecting: sprint kinematics change consistently through acceleration and transition, then plateau at maximum velocity.

Step 4: comparison and feedback

The final step is judgment. The model compares your measured angles and phase patterns against reference patterns of efficient running and flags where you differ in a way that costs you speed or risks injury. This is the part that turns numbers into a coaching cue. A good tool does not just dump data; it translates a measured fault, say a heel that floats up at toe-off, into a plain-language correction you can train. The numbers are the evidence; the cue is the point.

What AI form analysis can and cannot see

The honest boundaries, so you trust it where it is strong.

What AI cannot reliably see

Being honest about limits is what makes the tool trustworthy. A single-camera, two-dimensional system has real blind spots. It cannot accurately measure rotation about a limb's long axis, like hip internal rotation, where even good systems can be off by 7 to 13 degrees. It has no depth, so motion toward or away from the camera is distorted. And it struggles with occlusion: when one leg crosses in front of the other, the model can drop a joint or swap your left and right limbs, which is a common problem in sprint footage. The research is candid that these algorithms were not originally designed for biomechanics, and that loose clothing, bad lighting, and poor angles all degrade the result.

Why your filming decides the accuracy

Every limitation above points to the same conclusion: how you film controls how good the analysis is. Shoot side-on, so mechanics are in the camera's plane. Shoot in slow motion, 120 or 240 frames per second, so fast movement is not a blur for the model to misread. Use good light and fitted clothing so the joints are easy to find. Keep the whole movement in frame so nothing the model needs is cropped out. Do that and the four-step pipeline has clean input to work with. The full setup is in how to film yourself for track and field.

See it on your own stride

The clearest way to understand any of this is to watch it run on your own footage. Film one clean rep side-on and upload it to Track & Field AI. It maps your joints, measures your angles, segments your phases, and hands back the fault that is costing you, in coaching language, not raw data. Then put it to work with the full self-coaching loop in how to analyze your sprint without a coach.