Pose Tracking

Pose tracking keeps body joints linked to the same person across frames so walks, lifts, and falls read as continuous motion, not scattered poses.

Two athletes hold identities for 90 frames with zero switches while a crossing without appearance matching causes two swaps

Why Does This Exist?

Single frame pose fires joints without names. Next frame fires again. A sprint, a lift, or a fall needs one trail per person to measure speed and form. Crossings swap names and corrupt every downstream count.

Pose tracking exists to hold identities. Input is video plus per frame poses. Output is tracklets with joint trails. The per frame base is 2D pose and 3D pose. The box tracker cousin is multi-object tracking.

Use it when motion over seconds matters. Use single frame pose when one still image answers.

Think of It Like This

Name tags at a square dance

Dancers swap partners and cross constantly. A caller who only shouts positions loses who is who after one swing. Sticky name tags plus step prediction keep partners straight.

Poses are positions. Appearance tags plus motion prediction are sticky tags. The analogy stops at sweat: tags fall off in crowds while re identification features fade under occlusion, so gaps still need bridging.

How It Actually Works

Detect poses per frame, then match across frames by joint overlap plus motion prediction plus appearance. Contract is frames plus poses in, tracklets with IDs out. Smooth joints with a short filter to cut $3$ pixel jitter to $1$ pixel. Count identity switches at crossings: zero switches over $90$ frames means clean trails. Two swaps at one crossing flag a matcher that ignored appearance.

Worked example

Two runners cross at frame $45$ of $90$ . Overlap only matching swaps IDs twice, giving $2$ switches and corrupted stride lengths. Adding appearance distance keeps IDs: runner A stays $0.25$ from self and $0.88$ from rival, so no swap. Knee jitter drops from $3.2$ pixels raw to $1.1$ smoothed, which steadies speed from $5.8$ plus or minus $1.4$ m per s to $5.8$ plus or minus $0.4$ .

Watch Out For

Crossings that trade jerseys

Same kit teams make appearance useless at the cross second. The symptom is paired swaps exactly at overlaps. Fix it with pose motion priors and by holding IDs through short gaps instead of forcing a match every frame.

Smoothing that erases the event

A long filter turns a $120$ ms trip into a wobble. The symptom is clean trails plus missed falls. Fix it with short windows for joints plus separate event heads on raw motion.

The Quick Version

Pose tracking links per frame joints into ID stable tracklets.
Matching uses overlap plus motion plus appearance.
The fixture holds $0$ switches over $90$ frames with appearance on.
Smoothing cuts $3.2$ pixel jitter to $1.1$ for honest speed.
Crossings and long filters are the first failures to test.