Camera Angle Changes What AI Sees

A body moves in three dimensions. An image has two. Every camera resolves that by discarding one direction — the one along its line of sight — and keeping the two across it. Movement along the line of sight does not vanish entirely; it appears as a change in apparent size. But a change in size is a much weaker signal than a change in position, and it is entangled with distance.

So the question for any camera is not "can it see the person" but "which plane of the person's movement is it looking across". Gait analysts name two planes: the sagittal plane, the side-on slice in which the legs swing forward and back, and the coronal plane, the front-on slice in which the body sways and the feet are placed left and right. A camera sees one of them well, the other badly, or both partially.

Beside the walker, the camera looks across the sagittal plane. This is the textbook view, and for good reason: the knee's flexion and extension, the stride length, the arc the foot describes and the exact instant the heel meets the ground are all laid out across the image. Temporal measures read cleanly from here, because the events that define them — heel strike, toe-off — are unambiguous.

What the side view cannot see is anything that happens across the walk. Stride width — how far apart the feet are placed — is along its line of sight and collapses to nothing. Lateral sway is a small change in size. And the two legs overlap for part of every cycle, which is exactly the condition under which a pose estimator swaps left and right, as the first story in AI Under Stress showed.

Turn the camera to face the walker and the picture inverts. Now the coronal plane is across the image: stride width, the side-to-side sway of the trunk, whether the feet land in line or splayed, whether one shoulder drops. These are the signals that describe balance and support, and the side view could not give them.

The cost is everything along the walk. Stride length is foreshortened to a change in scale. Knee flexion — a rotation in the plane the camera is looking along — is nearly invisible. Heel strike is harder to place in time, because the event that marks it is a vertical motion seen at a shallow angle, and the person's apparent size grows every frame as they approach, so nothing can be compared frame to frame without correcting for it.

Almost no deployed camera is a side view or a front view. Cameras are mounted high in corners, looking down and across a space, and the people in that space walk in whichever direction they were going. The typical view is oblique: somewhere around forty-five degrees to the walk, from above.

An oblique view sees both planes — partially. Knee flexion is visible but compressed; stride width is visible but mixed with stride length; the foot's arc is a slanted ellipse. Every one of these can be recovered in principle, if the geometry is known. In practice the geometry is often not known precisely, and it changes as the person walks: a stride at the near end of the corridor covers many more pixels than the same stride at the far end.

Angle is the largest of the camera's choices but not the only one. Three quieter ones set the floor under everything a pose estimator can do.

• Height. A camera above head height looks down on the walker, so vertical motion — the foot lifting and landing — is seen at an angle and the ground plane is tilted in the image. Heel strike timing gets harder to place; body path gets easier to read.
• Distance. Further away, the person occupies fewer pixels. A shin that is thirty pixels long can carry a knee that is right to within a pixel or two; a shin that is eight pixels long cannot. Precision is bounded by pixels per limb before any model is involved.
• Lens. A wide lens covers the room and bends straight lines near its edges; a long lens sees a narrow slice with less distortion. The same walk at the edge of a wide frame is not the same shape as at its centre.

None of these are properties of the person and all of them change the measurement. A recording made from a different mount, or after a camera was moved by a few centimetres, is not directly comparable to the one before — which is why a longitudinal reading has to record its capture geometry alongside its values.

An engineering team cannot move the walls. What it can do is treat the view as a first-class input rather than a nuisance.

• Declare what is unavailable. If the geometry does not support knee flexion, do not compute it. A withheld measure is honest; a foreshortened one is a wrong number with a confident label.
• Prefer signals that survive the view. Rhythm, cadence and step-to-step variability are properties of timing, and timing survives almost any angle. Lean on them where the geometry is poor.
• Estimate the geometry, and say how well. A known camera height and angle turns an oblique view back into measurements; an unknown one should lower the confidence of every spatial value, explicitly.
• Use a second sensor for what the view cannot see. A wearable measures heel strike in the body's frame, not the camera's. The fifth Foundation's caution applies: fusion adds coverage only when the streams are honest about their own quality.
• Compare like with like. Two recordings from different mounts are two instruments. A trend is only a trend if the capture geometry was the same, or was corrected for.

The view-invariance literature in gait recognition spent a decade on exactly this problem, for identity; the covariate surveys in the evidence below are where that work is collected. Movement measurement inherits both the problem and the tools, and adds one requirement the identity work did not have: the output is a number a person will act on, so the geometry it came from has to be part of the answer.