Stabilising a sprint video filmed with a pan

How to tell whether a video can be measured

1. 01Watch the background, not the athlete

   In a measurable video, the stands, the posts and the track lines stay put from frame to frame, and the athlete is the one crossing the picture. If the athlete stays mid-frame while the scenery slides past, the camera followed them: the video cannot be measured as it is.

2. 02Check the four markers never move

   The four corners of the measured section must be visible and in the same place on the picture from the start of the sequence to the end. A marker that shifts, even by a few pixels because someone knocked the tripod, skews every distance.

3. 03Spot the edits

   Published videos are often edited: two shots of the same start, filmed from nearby angles, stuck together. A cut between two very similar shots is easy to miss by eye, and the geometry changes all at once. A sequence to be measured must fit inside one continuous shot.

4. 04Know the real frame rate

   Slow motion exported at 24 or 30 frames per second tells the player a frame rate that is not the one it was shot at. Every speed and every contact time depends on it. A video whose real frame rate is unknown gives no reliable timing, however well it is stabilised.

Why a pan breaks the measurement

A video holds no metres. To measure, you click the four corners of a section of ground of known size on one frame of the video. The computation works out where the camera is and where it points relative to the ground, then uses that for every other frame.

That assumes the camera did not move. If it follows the athlete, the track slides across the picture: a given pixel no longer matches the same point on the ground from one frame to the next, the markers clicked on the first frame are no longer where you clicked them, and the athlete's motion on screen mixes their running with the camera's movement. A well-followed athlete looks almost still in the middle of the frame while running flat out.

What stabilising onto a fixed camera does

When a phone turns on the spot without moving, two consecutive frames are linked by a simple, exact transformation called a homography: a warp of the image plane that says where each pixel of one frame lands in the other. It holds for the whole scene, however far away things are, provided the centre of the lens stays in the same place.

It is estimated by finding background points shared by both frames, corners of the stands, posts, ground marks, and discarding the ones that really move, starting with the athlete. Chaining those transformations from one frame to the next brings the whole sequence back onto a single reference frame. Each frame is then reprojected onto a large canvas, as if a fixed camera with a very wide field of view had filmed everything. The scenery stops moving, the athlete crosses the picture, and you are back to the situation a calibration expects.

What we learned

We tried this on videos of starts filmed with a pan. The registration itself works well when the scenery has enough detail. The trouble lies elsewhere.

• The choice of reference frame decides everything. A flat projection stretches without bound as the angle to the reference approaches 90 degrees. Using the first frame as the reference blows the canvas up, using the frame in the middle of the sweep, in angle rather than in time, shrinks it a lot.
• The canvas is huge and mostly black. Each frame only fills the part the camera was looking at in that moment, and the region covered all the time is often empty. You cannot crop without losing the athlete.
• A single bad frame spoils everything after it, since the transformations chain. Motion blur from a fast pan is enough. Those frames have to be skipped by anchoring on a neighbour.
• Edits give themselves away through the collapse of shared points, not through how the picture looks: two shots of the same track, with the same kit, look too alike for ordinary scene detection to tell them apart.
• The chain drifts a little over the frames. To check that a video really has become fixed, register every output frame directly against the reference rather than trusting the chain that produced it.
• A track surface on its own offers few points to follow, and its regular patterns can fool the matching. It is mostly the stands, buildings and trees that carry the registration.
• A cylindrical projection gives a nicer wide shot to watch, but it bends straight lines. A picture where the lane lines curve cannot be measured.

The limits that remain

• A camera that travels, carried by someone walking or mounted on a bike riding alongside the track, produces parallax: the foreground and the background do not slide at the same speed. No single homography fixes that, so such a video stays out of reach of this method.
• Stabilising does not recover the frame rate. Slow motion whose real frame rate is unknown stays wrong in speed and contact time.
• Frames reprojected far from the reference are heavily stretched, so less sharp, and body detection suffers there.

What this means for you today

TrackGenius only analyses videos filmed with a fixed camera, and accepts neither panning nor handheld footage. This is not a cautious default, it is the condition for the metres, seconds and degrees to mean anything. Everything above shows how much has to be stacked up to get close to it another way, and how fragile the result remains.

In practice, put the phone on a tripod, side-on, and leave it alone. To see more of the run, step back rather than turn, or film another section on a second rep: the analysis covers a window of five seconds at most anyway, which a fixed section of eight to ten metres covers very well. The panned videos you already have are still useful for watching technique by eye, not for taking measurements from.

Frequently asked questions

In short

• A panned video cannot be measured as it is: a calibration made on one frame only holds for the others if the camera stays still.
• One homography per frame can bring a pan filmed from a fixed point back onto a virtual fixed camera, with the reference in the middle of the sweep, but the canvas is huge and largely black, and a travelling camera or slow motion at an unknown frame rate stay out of reach.
• TrackGenius accepts neither panning nor handheld footage: a tripod, a side-on view, four still ground markers and a window of five seconds at most.