Keypoint Detection: From Sports Broadcasts to the Factory Floor
Keypoint detection is best known through sports. Broadcasters and coaching staff already use it to track joint angles in real time — a golf swing, a sprinter's stride, the mechanics of a basketball free throw — turning what used to be a coach's eye into a number. It is a real, sizeable use of the technology: the global sports tech market passed $30 billion in 2025, and pose estimation is a meaningful part of the tooling behind it, active enough that FIFA is running a dedicated skeletal-tracking research challenge in 2026 (Springer, pose estimation in sports survey; BasicAI, computer vision in sports).
What makes the technique worth a longer look is that the underlying task does not know it is on a sports field. Keypoint detection tracks the position of specific points on a body — shoulders, elbows, hips, knees, ankles — frame by frame, regardless of what the person is doing or where. A knee bends the same way in a warehouse as it does on a running track. The model tracking a sprinter's stride and the model watching a worker's spine angle while they lift a box are doing the identical underlying job.
This article covers four places keypoint detection earns its keep: sports performance analysis, and three operational uses — manufacturing and logistics ergonomics, collaborative robot safety zones, and rehabilitation progress tracking — with the published evidence behind each, and the privacy architecture that makes camera-based movement tracking workable outside a lab.
Object detection tells you what is there. Keypoint detection tells you how it is moving.
A standard detection model draws a box around an object and names it: forklift, pallet, person. That answers "what is here." Keypoint detection answers a different question: what is this body doing, right now, in terms of joint angles and positions. It is the difference between knowing a person is present in a frame and knowing whether that person's spine is at 45 degrees while they are holding 20 kilograms.
This is a meaningful shift for any use case built around posture, movement quality, or the relative position of a person to something else in the scene, rather than simple presence or absence.
Sports: turning a coach's eye into a number
The sports use case is worth taking seriously on its own terms, not just as an entry point to the rest of this article. Coaching staff and broadcasters use keypoint tracking to measure things a trained eye used to estimate: joint angle at the top of a golf backswing, stride length and cadence through a sprint, elbow and knee position on a basketball free throw. What used to be a coach's subjective read of "your form was off" becomes a specific, repeatable angle measured the same way every time, and comparable across a season rather than a single session.
The research investment reflects that. Beyond commercial broadcast and coaching tools, the field is active enough to support a dedicated FIFA-run research challenge on 3D skeletal tracking from broadcast soccer footage in 2026, and a growing body of published surveys on pose estimation methods specifically for sports (Springer, 2025).
MAKRR does not sell into sports as a market — it sits outside the three operational wedges this site is otherwise built around — but the detection task is the same one described above: joints tracked frame by frame, with no sense of what sport, or what setting, they are in. That portability is the actual point of this article.
Manufacturing and logistics: catching unsafe lifting before it becomes a claim
Manual handling injuries are not a marginal cost. Manual lifting, pushing and pulling injuries cost US businesses an estimated $13.7 billion annually in direct costs, and musculoskeletal disorders broadly cost $18 billion a year in direct expenses, with NIOSH putting total costs — including indirect costs — as high as $50 billion or more once lost productivity, turnover and training are counted (Voxel AI; NSC Risk Spotlight).
Manual handling is cited as the cause of 20 to over 30 percent of all reported workplace injuries, and back-related cases account for nearly half of all musculoskeletal disorder incidents. Serious musculoskeletal disorder cases resulted in 18.5 million lost workdays in 2024, and forceful exertion — lifting, pushing, pulling, carrying — has ranked as the single largest contributor to workplace injury costs for six consecutive years in the Liberty Mutual Workplace Safety Index. Warehousing and transportation workers carry the highest exposure: 4.4 injuries per 100 workers, nearly double the national average (SHRM-Atlanta; sci-tech-today).
The standard control for this today is a periodic ergonomics walk — a safety officer observing a line or a dock for an hour, a few times a month. That is the same coverage gap that shows up in every manually audited safety process: the moment nobody is walking the floor, nothing is being measured. A keypoint model watching the same cameras already mounted on a line or a dock can flag a spine angle beyond a safe threshold, a load lifted with the knees straight, or a reach outside a safe zone, continuously, and log it the same way every time — no observer effect, no sampling window.
The output is a posture flag, not a photograph. What gets logged is the event and the joint angles that triggered it, not a searchable video archive of the workforce.
Drones and robotics: the safety case for collaborative robot workspaces
Wherever a robot and a person share a workspace, there is a real, codified engineering requirement to track the human continuously. ISO/TS 15066, the safety standard for collaborative robots, defines speed and separation monitoring: the robot's safe operating speed is calculated in real time from the protective distance between it and a nearby worker, factoring in the robot's stopping distance, system latency and measurement uncertainty. Below that distance, the robot must issue a safety-rated controlled stop (ISO/TS 15066; PMC — Implementing SSM in Collaborative Robot Workcells).
Published implementations of speed and separation monitoring have used depth sensors like Kinect V2 to track worker position and movement continuously against exactly this standard (ScienceDirect). A keypoint model does the same underlying job — tracking where a person's body actually is, not just whether a person-shaped box is somewhere in the frame — on standard cameras rather than a dedicated sensor rig, and can run on an edge device at the cell rather than adding another system to integrate.
This is squarely a drones-and-robotics use case: the cross-cutting thread that runs through MAKRR's manufacturing, logistics and unmanned-systems work, wherever a machine needs to know not just that a person is nearby, but exactly how close, and moving which way.
Healthcare operations: an objective number instead of a clinician's estimate
Physical rehabilitation runs on range-of-motion measurement — a clinician positioning a goniometer against a joint and reading an angle, repeated at every visit to track recovery. It is a manual, hands-on measurement, and manual measurements carry rater-to-rater variability.
Recent published research has directly tested whether markerless, camera-based pose estimation can match that measurement. A 2026 single-camera markerless motion capture study found strong agreement with goniometer readings for shoulder range of motion, including in patients with adhesive capsulitis, with validity correlations of 0.85 for flexion and 0.95 to 0.97 for abduction (PMC, 2026). A separate 2025 study in Healthcare Technology Letters found computer-vision-based markerless pose estimation achieved reliability scores from 0.74 to 0.93 across hip and knee range-of-motion measurements (Wiley Online Library, 2025). Across this body of research, markerless systems are consistently reported as faster to set up and cheaper than marker-based motion capture, with a known trade-off of somewhat lower precision on the most demanding measurements.
The operational use case this points to is narrow and specific: tracking a movement metric — a joint angle, a range, a symmetry score — over successive visits, as a number a clinician's own record can reference, without a clinician manually angling a goniometer at every session. This is not a diagnostic tool, and it does not make treatment decisions. MAKRR is not a medical device and is not used for clinical decision-making; the output is an objective movement measurement that supports a clinician's own existing assessment, the same operational framing that applies across every non-clinical healthcare use case on this site. (For the wider set of non-clinical hospital use cases, see our article on computer vision in hospitals, which also covers gait and fall-risk monitoring.)
FAQs
-
Object detection draws a box around something and names it. Keypoint detection tracks the position of specific points on a body — joints — frame by frame, which makes it suited to questions about posture, movement quality and the relative position of a person to something else in the scene, not just presence or absence.
-
No. The published research and MAKRR's own deployments both use standard single cameras — CCTV, RTSP or USB — rather than a dedicated motion-capture rig or marker suit.
-
Not by design. A keypoint model needs joint positions, not identity, so systems built around it can be configured without any facial recognition or biometric identification.
-
Published research on markerless pose estimation for range-of-motion tracking has found strong agreement with manual goniometer measurement — reliability scores as high as 0.97 in some joints and movements — though marker-based systems still edge it out on the most demanding precision requirements.