The classical biomechanics study in tennis runs on motion capture. Reflective markers, six to twelve cameras, a player swinging in a lab while a biomechanist captures sub-millimetre joint positions at four hundred Hertz. The output is precise, the cost is high, the operational use case is narrow. Almost no academy below the federation tier runs this work routinely. The data is too expensive to produce at the volume coaching needs.
In the past decade, a second-best option has matured. Pose estimation from video. The same convolutional architectures that read the ball trajectory in the computer vision primer read the player’s skeleton — the joint positions, the limb angles, the postural geometry — from broadcast video or academy footage. The output is not lab-grade. It is operationally useful, and at a cost an academy can actually pay.
This article describes what pose estimation reads in a tennis context, where it provides credible signal, and where the wearable or the motion-capture lab remains necessary. It connects to the training-load article covered earlier; load and biomechanics are independent measurement layers but inform the same recovery decisions.
What pose estimation actually outputs
A pose-estimation pipeline reads the player’s body in every frame and outputs a stick-figure: positions of the head, shoulders, elbows, wrists, hips, knees and ankles. Modern systems based on HRNet, MoveNet or VITPose achieve sub-pixel joint accuracy on broadcast footage under reasonable lighting. The output is a per-frame skeleton.
From the skeleton, derived measurements are computed.
Shoulder rotation angle in the serve, frame by frame, from the trophy position through racket-drop and contact. The signature of an efficient serve is a particular pattern of shoulder external rotation reaching a particular maximum. Deviations from a player’s own baseline are computable.
Knee flexion at the load phase of the serve. Reduced knee flexion in late sets is a fatigue signal that surfaces consistently in research catalogued at IEEE Xplore.
Hip-shoulder separation at contact in groundstrokes. The kinetic chain depends on this separation; its absence indicates either a technical issue or a fatigue compensation.
Serve toss variance, measured as ball position at the top of the toss across multiple serves in a match. Toss variance correlates with serve accuracy and increases under fatigue.
Left-right asymmetry in stride length, ground contact time and stance width across baseline movement. Asymmetry growth across a match flags compensation, either for an undisclosed injury or for accumulating fatigue.
The five measurements above, computed per match, plotted over a season, surface meaningful biomechanical signal without any wearable on the player.
Where it works well
Three conditions matter for reliable pose estimation in a tennis context.
Lighting. Bright, even, daylight or strong artificial lighting works. Mixed-source lighting in indoor academy facilities, with sunlight through skylights crossing fluorescent banks, degrades performance. Performance also degrades on dark surfaces (some clay surfaces, certain rubberised indoor courts) under low ambient lighting.
Camera angle. Elevated centre and baseline angles work well. Court-side angles (player-cam, ball-kid level) introduce perspective foreshortening that confuses the skeleton estimator on certain limb configurations. Drone angles do well above forty-five degrees from horizontal.
Player clothing. Modern systems are robust to clothing colour. Loose clothing (warm-up jackets, baggy shorts) degrades joint precision because the skeleton is estimated from clothing edges rather than body edges. The same player in tournament whites is read more accurately than the same player in a training tracksuit.
When the three conditions are met, pose estimation produces biomechanical measurements at sub-degree accuracy on the major joints. When they are not, the measurements remain useful for trend analysis (within-match comparison) but should not be used for absolute baseline comparison (against motion-capture lab numbers).
Where the wearable becomes necessary
The single biggest gap in video-derived biomechanics is internal load. The skeleton estimator reads what the body does in three-dimensional space. It does not read what is happening inside the joint — ligament strain, tendon load, muscle activation pattern. For these measurements, the wearable (inertial measurement unit, electromyography sensor, ground-reaction force plate) remains necessary.
The practical implication: pose estimation reads enough to flag a problem worth investigating. The wearable, or the lab visit, reads enough to characterise the problem. The two layers are complementary, not interchangeable. The error vendors make — and the operations make in selecting vendors — is treating them as the same layer.
The Tennis Industry Magazine technology survey covered this distinction usefully in recent issues. Pose-based screening, wearable-based diagnosis. The order of operations matters.
The serve-toss case study
Of the measurements above, toss variance is the cleanest to operationalise. The toss is a deliberate, repeated motion. The ball position at the top of the toss should be tightly clustered for a well-prepared player. Variance is a measurable signal.
A player whose toss variance increases significantly across a match, particularly in the third set, is signalling fatigue, mental drift, or an upper-body niggle. The coach can corroborate with the player. The intervention is sometimes technical (reset the toss routine), sometimes physical (a recovery decision), sometimes mental (a pre-serve reset).
The model surfaces the signal. The coach makes the call. The interaction is small and tractable, which makes it a good first project for an academy adding biomechanics tooling to its workflow. The cost is low, the operational disruption is low, the gain is visible quickly.
What the model misses
Three things, in order of magnitude.
Acute injury risk. Pose estimation reads degraded biomechanics; it does not predict the moment the ligament goes. Acute injury prediction remains, in 2026, an unsolved problem. Anyone selling you the prediction is selling marketing language. The framing that survives scrutiny is “early signal of degraded biomechanics, prompting investigation”. Anything stronger is not credible.
Subtle technique errors. A defensible biomechanics model surfaces large-magnitude deviations. The 0.5-degree error in racket head angle that a senior technical coach reads from the side of the court is below the model’s noise floor. The model and the coach see different things. The coach sees the small adjustment; the model sees the seasonal drift.
Mental and tactical state. Covered already in the computer vision primer. The skeleton does not read intent.
Operational integration
A pose-estimation layer integrates well with the load-modelling layer. The biomechanical flags become an additional input to the weekly recovery shortlist. Where the load model says “this player’s recovery curve has dipped”, the biomechanical model can sometimes say “and her serve toss variance has increased and her knee flexion at the load phase has decreased by 4 degrees over the past three matches.” The combination is stronger than either alone.
The same operational discipline applies. The head of physical preparation reviews the combined output weekly. The head coach sees the flagged shortlist. The conversation with the player follows. The data is operational documentation, not parental communication.
The work fits inside the coaching analytics and player development cluster.
Vendor considerations
The pose-estimation market is bifurcated. Tier-one systems are built on top of broadcast infrastructure for Tour-level use and are not priced for the academy tier. Tier-two systems are built on top of GPU-accelerated pipelines that run on academy hardware and are affordable but require a moderate technical integration.
The questions to ask:
- What pose estimation framework do you run, and at what frame rate?
- Are joint outputs in 2D pixel coordinates or in 3D world coordinates (and if 3D, calibrated against what)?
- What is your accuracy band under suboptimal lighting?
- How do you handle multi-player frames (doubles, warmup with a partner, hitting partner present)?
- What is the integration shape with our existing video infrastructure?
The same vendor evaluation discipline established in the computer vision primer applies. Match the system to the decision; do not buy on demo theatre.
What this means for your operation
Three implications.
First, biomechanics screening from video is now affordable enough to be a routine layer in any high-performance academy operation. The investment is roughly comparable to one season of additional travel for one junior player. The decision is not financial; it is whether the academy is staffed to act on the output.
Second, the layer is screening, not diagnosis. Investment in pose-estimation should be paired with a clear referral path — to a physiotherapist, to a biomechanist, to a wearable diagnostic — for players the screening flags. The screening without the referral is unfinished work.
Third, the operational integration with load modelling is where the value compounds. Either layer alone produces useful signal. Together, they produce the kind of shortlist that lets a head of physical preparation act early instead of late.
To scope a pose-estimation pipeline against your specific video infrastructure and your specific coaching workflow, book a 60-minute call. Bring a sample of academy footage and we will run a feasibility read on the data quality.