Start with the decision, not the model
Real-time offside detection is not simply a matter of placing a camera beside a football pitch and running an object detector. The system must answer a precise question at a precise instant: when the attacking player’s teammate plays or touches the ball, is any relevant part of that player ahead of the ball and the second-last defender? It must then present useful evidence to the referee without pretending to replace the referee’s judgement.
For amateur Indian football, the strongest 2026 approach is a referee-assistance system. It can flag likely offside events, save short video clips, and display calibrated lines for review. This is more practical than claiming fully automated decisions, particularly on uneven grounds, during monsoon weather, or where camera coverage and internet connectivity are limited.
Teams starting from scratch should first review the fundamentals in how to build computer vision models on GitHub, especially dataset versioning, reproducible training, and deployment documentation.
Define the rule and operating conditions
Write the football logic before selecting hardware. Under the Laws of the Game, offside position is assessed at the moment the ball is played or touched by a teammate. A player is not penalised merely for being in an offside position; they must become involved in active play. A practical prototype can initially detect offside position at the pass moment, leaving active-play interpretation to the assistant referee.
Record the assumptions your product will make:
- Which competitions and age groups will use it?
- Will it operate from one touchline camera or several views?
- Is the result an alert, a replay, or an official recommendation?
- What latency is acceptable—one second for review, or near-live signalling?
- How will the system handle throw-ins, goal kicks, corners, and deliberate defensive plays?
These decisions prevent teams from training a model for a narrower problem and marketing it as a complete officiating solution.
Choose an affordable camera architecture
A single elevated camera aligned with the halfway line can support a useful prototype, but it will struggle with player occlusion and depth. A field-ready system should use at least:
- Main tactical camera: A fixed 1080p or 4K camera on the halfway-line side, high enough to see defensive lines.
- Ball or wide camera: A second camera covering the attacking third and helping identify the pass moment.
- Calibration references: Visible pitch lines, goalposts, corner flags, and known field dimensions.
- Local compute: A laptop with a modern GPU, edge device, or compact workstation so the core pipeline continues during poor connectivity.
For Indian amateur venues, local processing is usually preferable to streaming every frame to the cloud. It reduces latency and protects match footage when mobile networks are congested. Use cloud storage only for selected clips, model training, dashboards, or post-match review.
Camera placement matters more than headline resolution. Mount cameras securely, avoid direct glare, and test against evening floodlights, shadows, dust, rain, and crowded backgrounds. A stable tripod or fixed pole is essential; even small movement can invalidate pitch calibration.
Build the vision pipeline
A robust prototype normally contains five stages:
1. Frame capture and synchronisation: Timestamp every stream and measure clock drift. If cameras are not synchronised, the apparent positions of players and ball can be wrong at the critical touch.
2. Pitch detection: Identify field lines and map image coordinates to a top-down pitch representation using homography. Recalibrate whenever the camera moves.
3. Detection and tracking: Detect players, referees, and the ball, then maintain identities across frames with a multi-object tracker.
4. Event detection: Identify the exact pass or touch using ball trajectory, foot proximity, and short temporal context rather than a single frame.
5. Offside calculation: Estimate the relevant body point—normally the head, torso, or feet depending on the rule interpretation—compare it with the ball and second-last opponent, and produce a confidence score.
Do not rely on a generic person detector alone. Jerseys can be similar, players overlap, and the ball may be only a few pixels wide. Annotate representative Indian match footage across grounds, kits, camera heights, lighting conditions, and weather. Include difficult negatives such as onside runs, delayed passes, rebounds, and defenders partially hidden behind attackers.
A useful development path is to begin with recorded video and frame-by-frame review. Once tracking is stable, add live inference. Teams evaluating video-language systems may also learn from evaluating OpenRouter vision models for video understanding, but general-purpose multimodal models should support analysis—not replace a low-latency, deterministic tracking pipeline.
Engineer for uncertainty and latency
Every output should include confidence and a reason. For example: “Likely offside position, confidence 0.86; pass frame uncertain by 80 ms.” This is more useful than a binary red or green signal.
Track these metrics separately:
- Pass-frame error: milliseconds between the true touch and the detected touch.
- Position error: distance between estimated and manually labelled player reference points.
- Event precision and recall: false alerts and missed likely offside events.
- End-to-end latency: time from ball touch to referee notification or replay availability.
- Coverage rate: percentage of match time with usable camera views and stable tracking.
Avoid unsupported claims such as “95% accurate” unless the test set, competition, camera setup, and error definition are published. A system can have high player-detection accuracy while still making poor offside calls because a 100-millisecond timing error changes the result.
Use lightweight models, frame skipping, region-of-interest processing, and hardware acceleration where necessary. A practical system can process every frame around a suspected pass while using a lower rate between events. This reduces compute cost without sacrificing the important moment. Deployment teams can also study highly performant runtime for AI applications when optimising inference and video pipelines.
Design the referee workflow
The output should assist, not distract. Give the referee or video operator:
- A discreet likely-offside alert rather than an automatic whistle instruction.
- A synchronised replay covering several seconds before and after the pass.
- A top-down overlay showing the calibrated offside line.
- A confidence indicator and a visible “insufficient view” state.
- Manual controls to confirm, reject, bookmark, or correct the event.
If the camera loses the ball, players are heavily occluded, or calibration is stale, the correct output is no reliable recommendation. Explainability and the ability to override the model are essential for player trust and competition governance.
Pilot safely in Indian amateur leagues
Start with one venue and a small number of matches. Obtain consent from organisers, referees, players, and guardians where minors are involved. Establish who owns the recordings, how long footage is retained, who can access it, and whether clips can be used for training. Blur faces for public datasets unless clear permission exists.
Before live use, run the system in shadow mode: generate recommendations without showing them to officials, then compare them with manually reviewed decisions. After calibration, use it only for post-match review or referee assistance. Publish an evaluation report that includes lighting, camera failures, false positives, missed events, and examples where the system abstained.
A student or early-stage team can keep the first pilot manageable with open-source detection and tracking tools, a fixed camera, local storage, and a manually reviewed dataset. Students seeking a broader product direction can also explore best machine learning projects for computer science students for adjacent ideas in data collection, evaluation, and deployment.
Budget and rollout plan
Costs vary widely, but the main expenses are predictable: cameras and mounts, local compute, power backup, storage, field installation, annotation, engineering time, and match-day operations. Treat connectivity as optional infrastructure rather than a single point of failure. A modest pilot is more valuable than an expensive multi-camera system that has not been validated.
A sensible rollout is:
- Weeks 1–3: Define the rule scope, consent process, camera geometry, and annotation guide.
- Weeks 4–8: Collect and label representative footage; build pitch calibration and tracking.
- Weeks 9–12: Detect pass events, calculate offside position, and benchmark latency.
- Following matches: Run shadow tests, review errors with referees, and improve placement and abstention logic.
- Scale-up: Add cameras, better synchronisation, and competition dashboards only after the single-venue system is dependable.
Final recommendation
The most credible product is not an “AI referee.” It is a low-cost, transparent video-assistance tool that works offline, shows its evidence, reports uncertainty, and fits the habits of Indian amateur officials. Focus first on reliable camera calibration, pass timing, player tracking, and honest evaluation. Once those foundations work, richer analytics—such as tactical heat maps, player development reports, or searchable match clips—can follow without compromising the core officiating use case.