World Cup 2026 Offside Technology: AI, Computer Vision, and the Connected Ball

The 1966 Wembley Goal was a tragedy.

Not just for West Germany, which lost the World Cup final after one of the most controversial goals in football history, but for the idea that human vision alone could settle every decisive moment in sport. The ball hit the crossbar, bounced down near the goal line, and the officials awarded England a goal. Decades later, computer vision researchers revisited the incident with multiple-view geometry and concluded that the ball did not fully cross the line. We covered that story in detail in How Computer Vision Solved the Greatest Soccer Mystery of All Time.

That incident was about goal-line technology, not offside, but the underlying lesson is the same: at the highest level of football, a decision can depend on centimeters, camera geometry, timing, and human perception under pressure.

Offside is another rule that looks simple until you try to measure it at match speed. The decision depends on a precise instant, a moving defensive line, several moving attackers, body parts that count, body parts that do not count, and then a legal judgment about whether the attacker actually became involved in play.

That makes it a surprisingly rich computer vision problem.

At the 2026 FIFA World Cup, offside decisions are supported by an upgraded version of Semi-Automated Offside Technology (SAOT). The system combines calibrated stadium cameras, skeletal tracking, a 500Hz sensor package inside the adidas Trionda match ball, player-specific 3D models, and a centralized video review workflow.

That sensor package is not a vague “AI chip.” Public FIFA material describes a 500Hz motion sensor in the ball, while technical reporting on connected-ball systems describes inertial sensors such as an accelerometer and gyroscope, commonly packaged together as an Inertial Measurement Unit (IMU).

An accelerometer measures linear acceleration.
A gyroscope measures angular velocity, or how fast the ball is rotating around each axis.

Together, those signals help identify the exact instant of contact when a player kicks, heads, or deflects the ball.

The important point is that this is not a robot referee. It is a measurement system. The ball helps answer when the play happened. The cameras help answer where the players were. The player models help estimate which body part was farthest forward. Referees still decide whether that position becomes an actual offside offence.

In this post, we will unpack how the system works from a computer vision perspective and why it is harder than simply drawing a line on a broadcast frame.

Table of Contents

Why Offside Is Hard to Automate

The offside law has two parts that often get mixed together.

First, there is the geometric question: was the attacker in an offside position when a teammate played or touched the ball?

Second, there is the football judgment: did that attacker become involved in active play?

According to IFAB Law 11, it is not an offence merely to be in an offside position. A player is in such a position if any part of the head, body, or feet is in the opponents’ half and nearer to the opponents’ goal line than both the ball and the second-last opponent. Hands and arms are excluded because you cannot score with hands or arms (unless you are Maradona!). The player is penalized only if, at the moment the ball is played or touched by a teammate, the player then becomes involved in active play by touching the ball, challenging an opponent, obstructing line of vision, gaining an advantage from a rebound, or a similar action.

That means an automated system has to solve several subproblems:

• Detect the exact moment of the pass, touch, header, or deflection.
• Estimate the 3D pose of every relevant player at that instant.
• Identify the second-last opponent.
• Determine which parts of each body are legally relevant for offside.
• Compare attackers, defenders, and the ball in a calibrated pitch coordinate system.
• Decide when the measurement is clear enough to alert the officials.

Only some of those are pure measurement problems. Others are still interpretation problems.

That split is the key to understanding the technology. Computer vision can make the geometry faster and more reproducible. It cannot make every football judgment disappear.

What Changed for World Cup 2026

The core idea is not new. FIFA introduced semi-automated offside technology at the 2022 World Cup in Qatar. That system used 12 dedicated tracking cameras, tracked the ball and up to 29 data points on each player 50 times per second, and used a 500Hz inertial measurement unit inside the match ball to help detect the kick point.

For 2026, the system has been upgraded in several ways. Public FIFA material and recent technical reporting describe:

• More advanced optical tracking, with recent reporting describing 16 high-resolution cameras for the 2026 setup.
• The adidas Trionda match ball with a 500Hz sensor chip that sends real-time motion measurements to the VAR system.
• Player-specific 3D avatars, sometimes called digital twins, built from scans of each World Cup player.
• Advanced SAOT alerts for clear offside situations that can reach on-field officials through the audio communication system.
• Continued VAR and referee involvement for subjective or complex offside incidents.

The 2026 upgrade is therefore not just “more AI.” It is better sensor fusion: more camera coverage, more precise ball timing, more accurate player geometry, and a shorter communication path for clear positional offsides.

The System at a Glance

The easiest mental model is a four-layer stack.

Under the surface, offside technology is a practical combination of camera calibration, pose estimation, multi-view geometry, tracking, sensor synchronization, and decision support.

Layer 1: Optical Tracking Cameras

A single broadcast camera is not enough for reliable offside measurement. Perspective distortion can make one player appear farther forward than another even when the pitch-coordinate geometry says otherwise. Player occlusion, field camber, and camera angle can also introduce error.

FIFA’s offside technology documentation explicitly calls out field topography, camera-angle distortion, player occlusion, and body-part tracking as sources of error in virtual offside line systems.

Modern SAOT addresses this with calibrated multi-camera tracking.

Camera calibration means estimating how each camera maps 3D points on the pitch to 2D pixels in the image. In camera-calibration language, this has two parts:

• Intrinsic parameters: focal length, principal point, and lens distortion. These describe the camera itself.
• Extrinsic parameters: the camera’s 3D position and orientation relative to the pitch. These describe where the camera is mounted and where it is looking.

Once each camera is calibrated, the system can reason in a shared field coordinate system instead of in raw image pixels.

Each camera observes the pitch from a known position and orientation. If two or more cameras see the same body point, the system can triangulate its 3D position. Triangulation means back-projecting rays from multiple camera views and finding the 3D point that best explains the observed 2D locations.

In practice, the tracking system is continuously estimating many points per player, not just a center of mass.

Hawk-Eye’s SkeleTRACK documentation describes tracking 29 key points on an athlete’s body. For offside, that matters because the system needs the body parts that can legally score: head, torso, legs, and feet. A single centroid is not enough. A shoulder, knee, or boot can define the decision.

The vision tasks behind this layer are familiar, but the production version is much more demanding than a single-camera demo:

• Detect players in every camera view.
• Associate each detection with the correct player, a step called data association.
• Estimate 2D body keypoints in each view.
• Match the same body point across camera views.
• Triangulate 3D coordinates.
• Track positions over time to reduce jitter and handle occlusion, often using filtering ideas similar to a Kalman filter.

This is much closer to markerless motion capture than to ordinary object detection.

A few years back, our company bigvision.ai helped build and deploy a 12-camera computer vision system for tracking players and the soccer ball at a European football club. We cannot disclose details of that system, but the experience reinforced the same engineering lesson: multi-camera sports tracking is not just model inference. Calibration, synchronization, player identity tracking, ball localization, and operational reliability dominate the hard parts.

Layer 2: The Connected Ball

Offside is judged at the moment the ball is played or touched by a teammate. That moment is harder to identify than it sounds.

A standard video stream at 50 or 60 frames per second gives one image roughly every 16 to 20 milliseconds. A foot can contact and release a ball between frames. In a tight offside, choosing one frame earlier or later can move a sprinting player by several centimeters.

That is why the ball matters.

The 2026 official match ball, adidas Trionda, includes connected ball technology. FIFA’s launch material says the ball includes a 500Hz motion sensor chip that sends precise measurements to the VAR system in real time, helping match officials make decisions including offside incidents.

At 500Hz, the system receives one sample every 2 milliseconds. A kick, header, deflection, or touch creates a characteristic change in acceleration and rotation. That signal helps identify the candidate kick point.

The ball does not decide offside. Its main job is timestamping.

Once the system estimates the touch time, the optical tracking layer can ask: where were all relevant body parts at that exact instant?

In engineering terms, this is sensor synchronization:

ball sensor event time -> synchronized tracking frame -> offside geometry

That timestamp is the bridge between ball physics and player geometry.

What the Ball Sensor Actually Measures

It is tempting to say the ball sends “motion data,” but that phrase hides the interesting engineering.

A connected match ball typically uses an inertial measurement unit, or IMU. An IMU is a small electronics package that measures motion without needing an external camera. In many systems, it contains at least:

• Accelerometer: measures linear acceleration along three axes, usually written as ax, ay, and az. In a football, this can reveal sudden impacts because a kick creates a large acceleration spike.
• Gyroscope: measures angular velocity around three axes, usually written as wx, wy, and wz. This tells the system how fast the ball is spinning and around which axes.
• Clock and radio interface: timestamp the measurements and transmit them to the review system. Some reported connected-ball systems also use ultra-wideband, or UWB, a radio technology that can support high-precision ranging and localization by measuring signal timing.

The exact production signal-processing pipeline is proprietary, but a useful mental model is:

accelerometer spike + gyroscope change + synchronized timestamp = candidate ball touch

An accelerometer alone can detect a large impact, but it may not fully describe the event. The gyroscope adds rotational information. For example, a glancing touch may produce a different acceleration and spin pattern from a clean kick through the center of the ball.

There is also a useful derived quantity called jerk, which is the rate of change of acceleration. If acceleration changes very sharply, jerk is high. Touch detection systems often look for abrupt changes in the inertial signal rather than only the raw acceleration value.

A simplified IMU sample might look like this:

time_ms, ax, ay, az, wx, wy, wz
124500.000, 0.3, -0.1, 9.7, 0.02, 0.01, 0.05
124502.000, 38.5, 14.2, -6.8, 8.10, 2.40, 5.70  <- likely contact
124504.000, 12.4,  5.9, -1.1, 7.80, 2.35, 5.55

The units depend on the sensor system, but conceptually the accelerometer reports translational acceleration and the gyroscope reports spin rate. The important part for offside is not the exact force of the kick. It is the timestamp of the first relevant contact, because the player positions must be evaluated at that instant.

Layer 3: Player Digital Twins

The 2026 system also adds more realistic player-specific 3D models.

Earlier offside animations commonly used generic human figures. They were useful for explaining a decision, but they did not represent every player’s exact proportions. In close calls, that distinction can matter. Two players can have similar heights but different leg length, shoulder width, boot size, and running posture.

Reports around the 2026 system describe each World Cup player being scanned to create a physically accurate 3D model. These models are used to personalize the visualization and help the tracking system represent the player’s body more realistically.

From a vision perspective, this is an important idea: skeletal keypoints are sparse. A digital twin helps turn sparse keypoints into a more complete estimate of the surface and extremities of a specific body.

For example:

• A hip, knee, ankle, and foot keypoint can describe leg pose.
• A player-specific model helps estimate where the boot actually extends.
• Shoulder and torso dimensions can improve the boundary between playable body and arm.
• The same skeleton can be rendered in a way fans can identify on broadcast replays.

This does not make the 3D animation raw evidence by itself. The animation is still a reconstruction from sensor estimates. But player-specific geometry can make both measurement and explanation better.

Layer 4: VAR and Human Judgment

The word “automated” can be misleading. FIFA describes SAOT as a support tool for video match officials and on-field officials. Its purpose is to make offside decisions faster, more reproducible, and more accurate.

The important 2026 upgrade is Advanced SAOT. FIFA’s own explainer says the advanced system additionally sends an automated audio alert to on-field match officials for clear offside situations.

That is different from saying the software has final authority.

In straightforward cases, the system can detect that an attacker who receives the ball was clearly beyond the offside line when the teammate played it. In those cases, the assistant referee can receive information sooner.

But many situations still require human review:

• Did an offside-positioned player interfere with the goalkeeper’s line of vision?
• Did the attacker challenge a defender for the ball?
• Was a defender’s touch a deliberate play or only a deflection?
• Was there a rebound from the goalkeeper, post, crossbar, or match official?
• Did a player in an offside position actually become involved in active play?

Those questions are legal and contextual, not just geometric.

How an Offside Check Works

The exact production software is proprietary, but the functional sequence is clear enough from FIFA’s public explainers and technical reporting.

1. A teammate plays the ball

The ball’s IMU produces high-frequency accelerometer and gyroscope readings. A sharp acceleration spike, a sudden change in angular velocity, or a high-jerk event marks a likely touch, kick, header, or deflection.

2. The system retrieves the synchronized player state

The tracking platform aligns the IMU event timestamp with the camera-derived player positions. Because the cameras and ball sensor are synchronized to a common clock, the system can estimate the 3D pose of every player at the relevant time. If the exact touch falls between two camera frames, the system can interpolate the player positions between adjacent tracking samples.

3. The offside line is computed

The line is normally defined by the second-last opponent’s nearest legal body part to the goal line. If the ball is nearer to the goal line than the second-last opponent, the ball defines the line instead.

4. Attacker body parts are compared

The system compares each attacker’s legal body parts against the line. Hands and arms are excluded under Law 11, so the measurement must distinguish between playable body regions such as the head, torso, legs, and feet, and non-playable regions such as the arms.

5. The system waits for involvement

Being in an offside position is not automatically an offence. In the clearest positional cases, the relevant trigger is the attacker receiving or touching the ball.

6. Officials are alerted or VAR reviews the incident

Clear cases can produce an audio alert for match officials. More complex cases remain with the VAR and referee.

7. A 3D visualization is generated

Once the decision is confirmed, the same calibrated 3D coordinates can be rendered into a 3D animation for broadcast and stadium screens. The animation is a visualization of the tracked geometry, not a separate camera recording.

The Computer Vision Pipeline

If we ignore the legal part for a moment, the vision system has a clean engineering shape.

Camera Calibration

Every camera must know how its pixels map to the pitch. In computer-vision terms, this includes intrinsic parameters, extrinsic parameters, distortion correction, and a consistent world coordinate frame.

Without calibration, a line drawn on a broadcast image is only a graphic. With calibration, the line corresponds to a real plane on the field.

In a simplified setup, this is the difference between drawing a line as a visual overlay and projecting a real 3D pitch line into the image with a camera matrix and pose estimate. The first is just a graphic. The second is geometry.

Multi-View Detection and Pose Estimation

Each player is visible from several angles. The system detects the player and estimates body keypoints in each view. These 2D keypoints are then associated across cameras and lifted into 3D.

This is where modern pose estimation and tracking matter. Offside is not asking “where is the person?” It is asking “where is the farthest legal body part of this person at a specific instant?”

This requires more than bounding boxes. A bounding box around a player includes empty background and usually includes the arms, which are not relevant for offside. Keypoints are more precise: ankle, knee, hip, shoulder, head, and foot landmarks give the system a way to reason about playable body parts.

Temporal Tracking

Raw per-frame keypoints are noisy. Players occlude each other. The ball moves fast. A production system needs temporal filtering and identity tracking so that player A does not become player B when they cross paths.

This is where tracklets and filters become important. A tracklet is a short sequence of observations believed to belong to the same object or player. A filter, such as a Kalman-filter-style estimator, combines the current measurement with the previous state to reduce jitter and predict where the player is likely to be in the next frame.

Ball Event Detection

The connected ball supplies high-frequency IMU samples. The system detects candidate touch events from acceleration spikes, angular-velocity changes, and possibly derived features such as jerk. It then synchronizes that event with camera data.

If the ball also provides UWB-style radio measurements, those can support ball localization or validation. The key point is that the ball sensor gives the review system an independent high-rate signal that is not limited by video frame rate.

Geometric Decision Support

Once all coordinates are in the same frame, the geometric test is conceptually simple:

attacker_legal_body_part_x > offside_line_x

In real systems, the details are harder: pitch orientation, attacking direction, calibration uncertainty, visibility, body-part eligibility, and synchronization all matter. A good system also needs an uncertainty model. For example, if the estimated attacker position has a few centimeters of uncertainty, the software should treat that very differently from a case where the attacker is clearly beyond the defender by a large margin.

Visualization

Finally, the system renders the offside line, the involved players, and the key instant as a 3D animation. This is not the measurement itself. It is an explanation layer built from the measured coordinates.

A Simplified Version

Here is a deliberately simplified sketch of how you might think about the pipeline if you were prototyping a toy version with pose estimation models.

This is not production SAOT. It is a conceptual bridge for computer vision engineers.

# Conceptual pseudo-code, not production SAOT.

# 1. Calibrate cameras into the same pitch coordinate system.
cameras = load_calibrated_cameras("stadium_calibration.json")

# 2. Keep a short rolling buffer of tracked 3D player poses.
tracking_buffer = PoseTrackingBuffer(seconds=3.0)

# 3. The ball stream contains IMU samples:
#    accelerometer: ax, ay, az
#    gyroscope: wx, wy, wz
ball_event_detector = BallTouchDetector(
    acceleration_threshold=30.0,
    angular_velocity_threshold=6.0,
    jerk_threshold=120.0,
)

for timestamp, camera_frames, imu_samples in synchronized_stream:
    detections_2d = []

    for camera, frame in zip(cameras, camera_frames):
        players = detect_players(frame)
        keypoints = estimate_pose_keypoints(frame, players)
        detections_2d.append((camera, keypoints))

    # Associate keypoints across views and triangulate them into 3D.
    poses_3d = triangulate_player_poses(detections_2d)
    tracking_buffer.add(timestamp, poses_3d)

    # Detect kick/header/deflection candidates from inertial signals.
    touch_events = ball_event_detector.update(imu_samples)

    for event in touch_events:
        # event.time is usually more precise than a video frame timestamp.
        # It may fall between two camera frames.

        # 4. Interpolate player poses at the exact kick/touch time.
        poses_at_touch = tracking_buffer.interpolate(event.time)

        # 5. Compute the offside line from the ball and second-last opponent.
        offside_line = compute_offside_line(
            poses_at_touch,
            ball_position=event.ball_position,
            attacking_direction="left_to_right",
        )

        # 6. Compare attackers' legal body points to the line.
        candidates = find_positional_offsides(
            poses_at_touch,
            offside_line,
            attacking_team="A",
        )

        # 7. Send evidence to officials, not an unquestionable final verdict.
        if candidates:
            send_to_review_system(event.time, candidates, offside_line)

The core lesson is that the offside problem is not solved by one model. It is solved by a pipeline:

calibrated cameras + pose tracking + IMU timestamp + sensor fusion + geometry + human review

That is a common pattern in real-world computer vision. The model is only one part of the system. Calibration, synchronization, data association, validation, and the user workflow are just as important.

Why It Is Still Semi-Automated

The system can measure:

• the likely moment of ball contact from accelerometer and gyroscope signals
• 3D player positions
• body-part coordinates estimated from skeletal keypoints and player models
• the offside line
• whether an attacker is geometrically beyond that line

It cannot fully decide every offside offence.

Consider a player standing beyond the line who never touches the ball. Was that player blocking the goalkeeper’s view? Did the player challenge for the ball? Did the player’s movement affect a defender? Did a defender deliberately play the ball before it reached the attacker?

These are not just coordinate comparisons. They involve the Laws of the Game, context, and referee interpretation.

That is why “semi-automated” is the right phrase. The system automates measurement, not judgment.

Limitations and Failure Modes

Advanced SAOT is powerful, but it does not eliminate every problem.

Calibration Error

The entire system depends on camera calibration and pitch geometry. Small calibration errors can matter when the margin is tiny.

Occlusion

Players often block each other. A defender’s foot, an attacker’s shoulder, or the ball itself may be partly hidden in some views. Multi-camera coverage helps, but it does not make occlusion impossible.

Synchronization

IMU samples, camera frames, VAR video, and review interfaces must agree on time. A few milliseconds can matter when players are sprinting. This is a clock-synchronization problem, not just a computer-vision problem.

Sensor Noise and Bias

Accelerometers and gyroscopes are not perfect. They have noise, bias, saturation limits, and calibration drift. A production system must distinguish a true touch from normal ball flight, vibration, bouncing, or sensor noise.

Identity Tracking

In crowded penalty areas, tracking the correct player across frames and cameras is difficult. Player-specific models can help, but the association problem remains real.

The hard case is not one isolated attacker running in open space. The hard case is a crowded box where players overlap, switch directions, and partially hide each other’s limbs. In that case, a tracking system must maintain identity and body-part estimates through occlusion.

The Law Is Still the Law

Technology can show that a toe is beyond a line. It cannot settle the philosophical question of whether football should punish extremely marginal offsides in the same way as obvious ones. That debate belongs to rule-makers, not cameras.

Visualizations Can Look Too Authoritative

A clean 3D animation can make the decision feel obvious. But the animation is a model generated from tracking data. It should be understood as an explanation of the system’s estimate, not as independent raw footage.

World Cup offside technology is a useful case study because the stakes are high and the requirements are unforgiving. But the underlying ideas show up everywhere in applied computer vision.

The same building blocks appear in:

• autonomous driving, where multiple sensors must agree on object position and time
• robotics, where object pose and contact timing determine actions
• motion capture, where sparse keypoints must become stable 3D trajectories
• sports analytics, where player tracking becomes tactics and performance data
• AR and broadcast graphics, where virtual objects must align with the real world

The lesson is not simply that AI can help referees. The lesson is that real deployments require systems thinking.

A research model can estimate pose. A production system must also calibrate cameras, synchronize sensors, handle occlusion, quantify uncertainty, integrate with human operators, and explain its output to users.

That is exactly the gap between a computer vision demo and a system that can influence a World Cup goal.

Conclusion

World Cup 2026 offside technology is best understood as automated measurement with human adjudication.

The connected ball helps provide the kick point through IMU-style accelerometer and gyroscope measurements. The cameras provide multi-view player tracking. Skeletal tracking and digital twins estimate legal body-part positions. Sensor-fusion software aligns all of those signals in time and space, then computes the offside geometry. Clear cases can reach officials quickly, while complex cases still go through VAR and the referee.

For computer vision engineers, the system is a fascinating real-world example of multi-camera pose estimation, sensor fusion, 3D reconstruction, and decision support. It also shows why the last mile matters. Accuracy is not only about the model. It is about calibration, timing, workflow, and communicating uncertainty.

That is the real story behind the offside line. It is not a line drawn on a screen. It is a full-stack vision system built to answer one of the fastest and most controversial questions in sport.

References