Apply for AI Grants India

What possession analysis should measure

Ball possession is more than a percentage shown after a match. For a school team, useful possession analysis should answer practical questions: Which team controlled the ball? Where did possession change? How long did sequences last? Did players retain the ball under pressure, or circulate it safely without creating chances?

A computer-vision system estimates these answers from match video. It detects the ball and players, follows them across frames, identifies team affiliation, and assigns each moment to the team most likely to control the ball. The output is an estimate—not an official record—so the system should expose uncertainty and allow a coach or student analyst to correct errors.

For schools building a prototype, start with one measurable target: team possession percentage and possession sequences. More advanced metrics, such as field zones, progressive passes, turnovers, and pressure, should come later. Students who want to build the model can use the workflow in this guide alongside resources on building computer vision models on GitHub.

Plan the project around Indian school conditions

A useful system does not require a professional stadium, a multi-camera broadcast feed, or an expensive cloud platform. Many Indian schools play on grounds with uneven lighting, dust, partial markings, crowded touchlines, and limited internet access. Design for those constraints from the beginning.

Use a fixed camera on a stable tripod, preferably high on the halfway line. A phone can work for an initial pilot if it records at 1080p and remains wide enough to cover most of the pitch. Avoid frequent panning and zooming: a stable viewpoint makes calibration and tracking considerably easier. Record the fixture, age group, date, weather, pitch type, and camera position in a simple match log.

Before recording, obtain written permission from the school and the relevant guardians or players. Explain what is being captured, why it is needed, how long it will be stored, and who can access it. Do not publish identifiable footage or player-level rankings without appropriate consent. Store raw videos securely, restrict access, and define a deletion date.

Build a small, representative dataset

Do not begin by recording dozens of matches without a labelling plan. Capture a few games across different conditions: morning and afternoon light, different kits, occlusions, long balls, crowded penalty areas, and moments when the ball leaves the frame. A dataset that reflects real school football is more valuable than a large collection of easy clips.

Create labels for:

• The ball, including difficult cases where it is blurred or partly hidden.
• Players, referees, and goalkeepers as separate classes if useful.
• Team identity, based on shirt colour or manually assigned track identity.
• Visible pitch lines and corners for field calibration.
• Possession events, such as control, pass, tackle, interception, out-of-play, and uncertainty.

For a first version, sample frames rather than labelling every frame. Review short clips manually to check whether the detector and tracker behave sensibly between labelled frames. Keep training, validation, and test footage separated by match, not just by frame. Randomly splitting neighbouring frames can make accuracy look much better than it is because nearly identical images appear in every set.

A practical computer-vision pipeline

The workflow can be built with Python, OpenCV, a modern object detector, and a multi-object tracker. The exact model matters less than consistent data and evaluation.

1. Detect objects. Train or adapt a detector for the ball, players, referees, and relevant field landmarks. The ball is usually the hardest target because it is small, fast, and frequently occluded.
2. Track identities. Use a tracker such as ByteTrack, BoT-SORT, or a Kalman-filter-based approach to maintain player paths between detections. Expect identity switches when players cross or disappear behind others.
3. Calibrate the pitch. Map image coordinates to pitch coordinates using visible lines or manually selected landmarks. A homography helps compare locations across the field and supports zone-based analysis.
4. Assign teams. Use shirt-colour features only as an initial signal. Similar kits, shadows, and goalkeepers can confuse an automated classifier. Manual correction is often worthwhile in school footage.
5. Estimate ball control. Combine ball location, nearest-player distance, player movement, and short temporal context. A player being closest to the ball does not automatically mean that their team has control.
6. Label uncertainty. Mark frames as contested, occluded, out of play, or ball-not-visible instead of forcing a team assignment.

Possession should normally be assigned in short windows rather than independently frame by frame. Smooth predictions over time and require a minimum duration before declaring a change. This reduces noisy switches caused by a single missed detection.

Calculate possession transparently

A simple calculation is:

Team possession percentage = controlled-ball time for the team ÷ total in-play controlled-ball time × 100

Exclude stoppages, half-time, goal celebrations, and periods when the ball is clearly out of play. Decide in advance how to treat contested balls, rebounds, aerial challenges, and goalkeeper distribution. Publish those rules with the result.

Track more than the headline percentage:

• Number and median duration of possession sequences.
• Possession by pitch zone: defensive third, middle third, and attacking third.
• Turnovers after receiving under pressure.
• Time from regain to a forward action.
• Share of sequences ending in a shot, foul, or entry into the penalty area.
• Confidence or percentage of frames reviewed manually.

Compare the automated estimate with a human-coded sample. Have an analyst label selected five- or ten-minute intervals, then report the difference. If the system says 58% and the manual review suggests 52%, that gap is important. It may reveal ball-detection failures, incorrect team assignments, or a definition of possession that is too generous.

Turn video into coaching decisions

A dashboard should support discussion, not replace coaching judgement. Show a timeline of possession changes, a pitch map of regains and losses, and clips linked to important events. A coach might discover that a team had 55% possession but lost the ball repeatedly in the middle third, or that its longest sequences started after defensive recoveries.

Use age-appropriate feedback. Young players benefit from concrete clips and questions: “What passing option was available before the turnover?” or “How did the team create a third-player option here?” Avoid public leaderboards that label children as poor performers based on uncertain automated data.

For schools that want to make the project part of STEM learning, let students test detection accuracy, build annotation tools, and explain the limits of predictions. This fits naturally with broader machine-learning projects for computer science students and can become a structured classroom exercise rather than a black-box sports product.

Costs, deployment, and maintenance

A pilot can run locally on a capable laptop using recorded video. Processing lower-resolution footage or selected clips reduces cost, while a GPU can speed up detection. Cloud processing may be convenient, but schools should account for upload time, recurring charges, and data governance. Keep an offline workflow available for grounds with unreliable connectivity.

Start with a baseline that coaches can trust: one camera, one age group, a small set of metrics, and manual review. Add automated reports only after measuring errors. Recalibrate when the camera moves, kits change, or the pitch layout differs. Models trained on one school’s ground may not transfer cleanly to another.

Common failure modes

• The ball disappears: use temporal prediction, higher shutter speed where possible, and manual review of uncertain clips.
• Players merge in crowded areas: combine detection with tracking and field position; do not rely on one frame.
• Team colours are confused: include varied lighting in training data and permit manual team assignment.
• Possession switches too often: smooth predictions and define a minimum control duration.
• The percentage looks precise but is not: report confidence, excluded time, and human-review agreement.
• The system becomes a surveillance tool: limit access, obtain consent, and avoid unnecessary biometric identification.

A sensible 2026 implementation checklist

• Define possession and out-of-play rules before collecting data.
• Record stable, wide-angle footage from a permitted location.
• Label representative clips from real school conditions.
• Split evaluation data by match and test on unseen games.
• Validate results against human-coded intervals.
• Present uncertainty alongside every headline statistic.
• Use insights for learning and training, not punishment or selection alone.
• Document storage, access, retention, and deletion practices.

The best school football analytics project is not the one with the most sophisticated model. It is the one that produces understandable evidence, respects young players, and helps coaches make better decisions with the resources they actually have.