Computer vision in tennis is sold the way most new technologies are sold. As an oracle. Drop the cameras in, the system reads the match, you get insight. The buyers — academy directors, federation tech leads, head coaches — sit through demos that conflate three different things: what the camera physically sees, what the model statistically extracts, and what a coach can do with the output. Each is a separate problem, and the gap between them is where most procurement decisions go wrong.
This piece is a primer for the people who sign the procurement document. It is not a technology survey. There are plenty of those, written by people who do not have to make the system work on a Tuesday afternoon at the side of a court. The goal here is narrower: name what the technology actually reads, name what it does not, and give you a vocabulary for the questions you should ask a vendor before signing.
What the camera physically sees
A broadcast-grade tennis match generates between four and twelve fixed video feeds, depending on the venue. The two principal feeds — baseline and elevated centre — capture the full court at roughly thirty to sixty frames per second. Modern systems add court-corner cameras for serve toss, net-cord cameras for low-trajectory shots, and in a few Grand Slam venues overhead drones or rail-mounted tracking rigs.
What the cameras see is geometry. Pixels on a sensor. The match becomes a stack of images, time-coded, with everything in frame: the players, the ball, the net, the court lines, the chair umpire, the ball kids, the crowd in the background, the wind moving the curtains. The camera does not know that the small yellow object near the baseline is a ball. To the sensor, it is a cluster of pixels with a certain brightness and a certain edge profile.
Everything that follows — every “AI insight” the vendor demo promises — depends on what the model can extract from those pixels.
What the model statistically extracts
The first job of any tennis computer-vision pipeline is to find the ball. This sounds trivial. It is not. The ball is small, fast, often occluded by the player, sometimes lost against background of similar colour. State-of-the-art trackers built on convolutional neural networks achieve detection accuracy in the high nineties on broadcast video under good lighting. On academy-grade video shot with two consumer cameras and uneven exposure, that number drops considerably. Anyone who has read the MIT Sloan Sports Analytics Conference papers on tracking quality knows the headline numbers are venue numbers, not the numbers you will see on your court.
Once the ball is reliably tracked across frames, the model can derive a series of measurements. Shot direction, with confidence. Shot depth, expressed in metres from the baseline. Bounce location, the most reliable measurement because the ball is briefly stationary. Contact point, derived by triangulating ball trajectory and player skeleton position. Rally length, easy. Average shot speed, reliable. Spin signature, partial — current systems estimate topspin and slice direction within useful bands, but absolute revolutions-per-minute remains noisy without dedicated radar.
These are good signals. A tactical analyst can build a great deal of insight on top of them. They are the foundation of the brief I describe in the next article, which deals with how the brief itself gets assembled.
What the model does not read
Here is the part vendors rarely volunteer. Computer vision does not read intent. The system can tell you the ball went to the backhand. It cannot tell you whether the player chose the backhand to set up the next shot, was forced there by court position, or was guessing. The system measures the outcome. The reason sits in the player’s head, accessible only to the coach who knows her.
The model does not read mental state. There is a strand of research on “stress detection from posture” — most of it is wishful. The signal-to-noise ratio is poor, the labels are inconsistent, and the operational use case is thin. A coach can usually read the player faster than the model can.
The model does not read pre-shot decision time. The interval between the opponent’s contact and the player’s first preparatory movement is a real signal, but extracting it cleanly requires a level of pose precision most systems do not deliver outside controlled environments.
The model does not read what would have happened. Counterfactuals — what if the player had hit cross-court instead of down the line — are a separate modelling exercise built on top of the measurements, not a property of the camera.
The procurement implication
If the four bullets above are clear, the conversation with a vendor becomes more useful. The right questions stop being “does your system read intent?” and start being more specific.
- What is the ball-detection accuracy on video shot in lighting conditions matching mine?
- Is spin-signature output qualitative (topspin / slice direction) or quantitative (revolutions per minute), and where does the calibration come from?
- What does your contact-point estimate degrade to under partial occlusion?
- What pose-estimation framework are you running and at what frame rate?
- How is bounce-location accuracy validated on my court geometry?
These are questions a vendor can answer concretely. If they cannot, you have the answer you needed.
Hawk-Eye and what the broadcast tier built
Hawk-Eye Innovations is the reference point for production-grade tennis tracking. The system that calls the lines at Wimbledon and the US Open is not the same system most academies will deploy. It runs on ten high-speed cameras, custom calibration, and a tightly controlled environment. Replicating it costs more than most academies’ annual technology budget combined.
What this means in practice: the broadcast tier sets the upper bound on what is technically possible, not the operational standard for the consulting tier below. Knowing that gap exists changes how you read a demo. A system that performs on a Tour stage will not necessarily perform on your indoor courts under fluorescent lighting. A system that performs in your facility may not match Tour-grade ball detection — and may not need to, if the use case is coaching feedback rather than line-calling.
What this means for your operation
The clearest framework for buying tennis CV is this. Decide first what decision you want the system to support. Coaching debrief? Talent identification? Opponent prep? Each decision needs a different subset of measurements at a different accuracy threshold. A debrief tool can tolerate noise that a line-calling tool cannot. A scouting tool needs cross-match comparability that a single-session debrief tool does not.
Match the system to the decision. Do not buy the system because the marketing video shows a heat map.
If you want to scope this against the operations you actually run, book a 60-minute audit. We will map what your decision needs to support, what data you already have, and what the cheapest credible path from one to the other looks like. The other half of the picture — how the extracted measurements turn into a usable tactical brief — sits inside the match analysis and performance cluster.