An augmented reality football stadium experience lets a fan in the ground point a phone at the pitch. The app recognises the pitch and each player, shows names and statistics, and gives live speed or a heatmap for a tapped player. FIFA offered this to fans in the stadium at the 2022 World Cup through the FIFA+ Stadium Experience, and Immersiv.io launched a similar system, ARISE, in Wolfsburg in 2019.
What fans see through augmented reality at a football stadium
FIFA described the 2022 FIFA+ Stadium Experience as a live augmented reality overlay of stats, heatmaps, insights, extra camera angles and VAR replays “as on TV”, available only to fans attending the match. In practice the fan taps “Start Scanning”, waits a few seconds for “Pitch recognition successful”, then taps a player to see live speed or a heatmap.
ARISE, Immersiv.io’s AR solution, first ran with Vodafone 5G at the Volkswagen Arena in Wolfsburg in September 2019 and offered two modes. In the interactive mode the fan selects a player on screen to track them and see speed, passes, shots or a heatmap. In the lean-back mode insights such as goals or players “on fire” appear automatically in 3D on the pitch.
Other AR uses in stadiums are less about the match: overlays tied to specific seats or zones, or a view that highlights the fan’s own seat and nearby amenities.
How does a phone or AR glasses locate the pitch and the players in real time?
A football pitch is itself a calibration target. Its markings have strictly regulated shapes and dimensions, which is why sports camera calibration works “in the wild” with no prior knowledge beyond detecting those markings correctly. Calibration (in sports also called field registration) estimates the camera’s intrinsic and extrinsic parameters, and a homography maps the flat playing surface to the visible area of the image. With that mapping, a point on the playing field can be located in the image.
The BroadTrack authors (WACV 2025) note that broadcast cameras often show radial distortion and model it with a distortion term applied in normalised coordinates. Broadcast lenses zoom up to 25 times, and the BroadTrack authors argue the projection centre moves along the optical axis as the camera zooms.
Immersiv.io says ARISE now uses computer vision and machine learning to recognise the field automatically, so the experience can launch easily in stadiums around the world. A second route is visual positioning: Google’s ARCore Geospatial API matches what the camera sees against a 3D point cloud built from Street View imagery and computes device position and orientation more accurately than GPS alone.
In ARISE, Sportec Solutions and ChyronHego collected up to 3.6 million positional data points and provided data in real time through the game. Our piece on creating AR experiences with computer vision covers the detection and tracking side.
How does in-stadium AR put live stats and heatmaps onto the pitch?
The match data is positional data supplied in real time by the tracking provider. Tracking and incoming results arrive at different times. The Rutgers MobiCom 2019 paper shows the effect of lateness: with delayed detection results, a rendered object trails behind the hand it should sit on. The fix is to decouple rendering from slower work instead of waiting for it.
Worked example: fitting an overlay into a 16.7 ms frame
The figures below come from a 2019 Rutgers WINLAB paper on edge-assisted object detection for mobile AR. Its test is a lab rig, not a football stadium: a Jetson TX2 as the AR device (the same Tegra TX2 SoC as the Magic Leap One), Wi-Fi to a router, and a PC with a Titan XP GPU as the edge server. The assumption we add is that a stadium app faces the same per-frame budget on the same class of hardware.
| Stage | Where it runs | Figure reported |
|---|---|---|
| Frame budget at 60 Hz | Device display | 16.7 ms (1000 ms ÷ 60) |
| Object detection on an HD frame | Mobile GPU (Tegra TX2) | More than 50 ms |
| Lightweight detection models | High-end phone GPU | More than 500 ms |
| Offload to edge, baseline | Wi-Fi 2.4 GHz / 5 GHz | 34.56 ms / 22.96 ms mean |
| Offload to edge, optimised | Wi-Fi 2.4 GHz / 5 GHz | 17.23 ms / 15.52 ms |
| Motion-vector tracking | Device (Jetson TX2) | 2.24 ms |
| Device resources used by tracking | Device | 15% CPU, 13% GPU |
The arithmetic tells the story. Detection on the device takes at least three frames, so it cannot run per frame. Even the optimised offload is close to a whole frame on its own, before rendering starts. The paper’s answer is to decouple: the device tracks and renders every frame from cached results, shifting old detections by motion vectors, and folds in fresh edge results when they arrive. With tracking at 2.24 ms, about 14.5 ms of the 16.7 ms frame is left for rendering (16.7 − 2.24).
The same authors point out that richer mixed-reality graphics will likely need motion-to-photon latency under 20 ms, the threshold associated with avoiding motion sickness in VR. A stadium label is simpler than that, but the principle carries: the frame deadline belongs to rendering, and everything slower has to work asynchronously around it. Latency budgets for broadcast AR graphics versus post-production overlays are covered in our piece on XR game development and broadcast-adjacent pipelines.
What compute and GPU hardware does real-time stadium AR need?
The table sets each stadium AR workload beside the compute the sources report for it.
| Workload | Where it sensibly runs | Evidence |
|---|---|---|
| Per-frame pose tracking and rendering | Fan’s phone or glasses | 2.24 ms tracking, 13% of the GPU on a Tegra TX2 |
| Heavy object or keypoint detection, if used | Edge GPU server | Titan XP with inference built on NVIDIA TensorRT |
| Broadcast camera calibration and tracking | Server GPUs | BroadTrack: 16 fps on HD frames on two RTX 4090s, unoptimised |
| 3D ball localisation from one camera | CPU | 50+ fps on one Xeon E5-2660 core, results in under two seconds |
| Player positional data | Tracking provider’s system | Up to 3.6 million data points per match feed in ARISE |
Some things to read carefully in that table. The BroadTrack authors expect their code to reach real time with proper optimisation, but 16 fps is what they measured. The ball localisation result is from Russian Premier League footage at 6K resolution, and its sub-two-second latency is aimed at live broadcasting. A review of AI in sports vision covers both GPU-based workstations and GPU-based embedded platforms for real-time detection and tracking at the edge.
If the app takes player positions from a feed, the device mostly needs a camera pipeline and a modest GPU for drawing, and the expensive inference sits upstream with the tracking provider or an edge server.
For comparing GPU candidates, LynxBenchAI measures throughput, latency, memory and energy across accelerators on one scale.
Qualcomm’s 2022 announcement of the Snapdragon AR2 Gen 1 describes a dedicated engine for motion tracking and localisation, an AI accelerator for latency-sensitive input, a reprojection engine and support for up to nine cameras. Those are Qualcomm’s announced specifications, not measurements. The closest thing to a measurement is the Rutgers tracking figure on the same SoC as an earlier headset.
The stadium network decides whether the overlay arrives
Modern stadiums present uniquely challenging wireless environments for a mobile app. A 2025 review of large-venue Wi-Fi in the World Journal of Advanced Engineering and Technology Sciences gives planning targets of 8 to 10 Mbps per client and under 50 ms latency for real-time venue applications. Those are planning targets, not measurements of an AR app. The same review notes that deployments typically target 60 to 80 clients per access point, and that a modern access point can push up to 1.3 Gbps. If every client on one access point wanted 8 to 10 Mbps at once, the total would be 480 to 800 Mbps (60 × 8 to 80 × 10): inside one access point’s ceiling.
A 2026 study at Notre Dame Stadium, with 77,622 spectators and about 900 outdoor Wi-Fi access points alongside a multi-carrier 4G/5G distributed antenna system, found that game-day crowds strain cellular uplink and latency badly. Under 5G Non-Standalone anchor congestion, image upload failure rates reached 46%. The dense Wi-Fi network gave the lowest game-day page-load failure rate (3.9%) and held image upload latency to 2.1 times the empty-stadium baseline.
Stadium operators are answering with private 5G or hybrid Wi-Fi/5G systems built for low latency and high capacity. ARISE itself launched alongside Vodafone’s 5G rollout in the Wolfsburg stadium. The engineering lesson for the app is about direction. An architecture that streams camera frames up to an edge server, as the Rutgers design does, depends on the uplink that game-day crowds strain. An architecture that keeps pose on the device and only pulls a compact data feed down relies mainly on the downlink, which is more mature than the fragile uplink.
Is stadium AR the same thing as VAR or Hawk-Eye?
No. Fan AR, VAR, semi-automated offside and goal-line technology answer different questions for different people.
| Fan AR (FIFA+, ARISE) | VAR | Semi-automated offside | Goal-line technology (Hawk-Eye) | |
|---|---|---|---|---|
| Purpose | Enhance the fan’s in-stadium experience | Help the referee fix clear and obvious errors | Alert officials to offside | Decide whether the whole ball crossed the line |
| Sensing | Fan’s phone camera plus a positional data feed | Broadcast footage with replay control | 12 roof-mounted tracking cameras, up to 29 points per player, 50 times a second, plus a ball sensor at 500 Hz | Seven cameras per goal |
| Who sees the output | One fan, inside the stadium | The referee, who makes the final decision | Video match officials first, then stadium screens and broadcasters | The referee’s watch within a second |
Under the IFAB protocol, VAR may assist only on goal or no goal, penalty or no penalty, direct red cards and mistaken identity, and the referee alone makes the final decision. Semi-automated offside adds AI on top of limb and ball tracking, but the automated alert is still validated by people before the referee hears it. Hawk-Eye’s SAOT, powered by its SkeleTRACK system, is reported by the company as used across more than 900 matches.
The systems meet at the output. FIFA+ included VAR replays in its AR layer, and the 3D offside animations built from the decision data were shown on stadium screens. Goal-line technology, by contrast, has to meet FIFA’s margin-of-error requirement of ±3 cm, which Hawk-Eye betters.
Augmented reality in football training and match analysis
For coaching, a formative study points away from headsets and towards phones and tablets. A 2024 formative study on tactical communication in outdoor team sports found that coaches and players considered bulky devices such as Apple Vision Pro, Oculus Quest or HoloLens impractical during play, citing setup time and discomfort when sweating. In its comparison, 12 players preferred mobile AR and 5 preferred traditional tactic boards, with little interest in MR or VR.
The requirements the coaches set are a useful spec for anyone building this: rapid communication, minimal disruption to play, reduced cognitive load, clear visuals for every player and clear audio. The authors’ answer is one device per player, so each sees the tactics from their own position on the field rather than crowding round one screen. Phones can be checked during breaks, and assistant coaches can preset each player’s view.
In match analysis, AR appears as augmenting inside video. A survey of video-based soccer analysis describes in-video augmenting as increasingly important for understanding a team’s capabilities and limits, and lists systems from BBC R&D and Ericsson’s Piero to Viz Libero. Field registration also underpins 3D player tracking on the field, the tracking engineering behind our AI engineering for sports analytics work. Our article on augmented reality in sports training goes further into the training side.
At Notre Dame, it was the cellular uplink that was strained under game-day crowd load.
FAQ
What does VAR stand for in football?
VAR stands for Video Assistant Referee: a match official with independent access to match footage who can assist the referee only on clear and obvious errors or serious missed incidents.
Is AI being used in football?
Yes. FIFA’s semi-automated offside technology applies artificial intelligence to limb and ball tracking data to generate offside alerts for video officials, and fan AR apps such as ARISE use machine learning and computer vision to recognise the field.
Can fans watching at home use stadium AR?
Not with the FIFA+ Stadium Experience, which was only open to fans inside the stadium. Viewers at home get AR as broadcast graphics instead, covered in our piece on augmented reality entertainment.
