CASE STUDY · CLIENT PROJECT · URBAN SOCCER PARK, USA
AI-Powered Soccer Camera System
An end-to-end physical AI system that understands live soccer, tracks players and the ball, and directs its own virtual camera — with all the vision running on edge hardware at the pitch.
Client: Urban Soccer Park ↗, a US company that builds modular 5-a-side mini-pitches for schools, cities and clubs nationwide. The AI camera is being piloted at one of their venues.
- CLIENT
- Urban Soccer Park ↗ (USA)
- MY ROLE
- Head of AI & Computer Vision
- TYPE
- Physical edge AI system
- STATUS
- Live pilot at the venue
WHAT THIS ENABLES AT SCALE
A venue selling match recordings.
The business model the pilot is proving out — an illustration, not pilot revenue.
ASSUMPTIONS
Pitches
3Matches per pitch per day
8Matches that buy a recording
25%Price per recording
$8
RECORDINGS SOLD
~6 / day
from ~24 matches a day
NEW REVENUE
~$1,450 / mo
about $17,000 a year per venue
Filming 24 matches a day by hand isn’t realistic — this revenue only exists because the camera runs on its own.
THE CHALLENGE
How can a physical camera understand a live soccer game — and decide where to look?
Filming amateur matches well normally needs a camera operator. A fixed wide shot is cheap but hard to watch: the ball is a few pixels and the action gets lost.
The goal was a camera that perceives the game the way an operator does — who's on the ball, where the play is heading — and produces a watchable view on its own, in real time, at the pitch.
Project reel
- 0:00 Physical camera
- 0:05 Installed at field
- 0:10 Raw video
- 0:20 AI detections
- 0:30 Ball & player tracking
- 0:40 Virtual camera
- 0:50 Final output
From pitch to picture, locally.
Every frame is captured, understood and reframed on the device — no cloud round-trip in the loop.
01
Soccer fieldLive game, wide coverage
02 · CAPTURE
Camera systemArducam IMX708 over MIPI CSI
03 · EDGE
Edge deviceRaspberry Pi 5 · Jetson path
04 · PERCEPTION
Computer visionPlayer & ball detection, tracking, scene understanding
05 · DECISION
Camera intelligenceTarget selection, smoothing, virtual camera
06 · OUTPUT
Video outputRTSP stream · recordings
DESIGNED FOR EDGE DEPLOYMENT
Not a cloud app. That's the point.
Unlike cloud video analytics, core vision runs locally next to the camera. That keeps latency low and avoids streaming high-resolution video to a remote server just to decide where to point.
Decisions made at the source
No raw-feed upload needed
Keeps running if the link drops
A real physical system.
Cameras, compute, wiring and mounting — designed and tested on the actual pitch.
Camera: Arducam IMX708
Interface: MIPI CSI
Compute: Raspberry Pi 5
Next-gen path: NVIDIA Jetson
From prototype to field system
BUILT
PrototypeRaspberry Pi 5 + Arducam IMX708
BUILT
Vision pipeline720p real-time inference
LIVE PILOT
Intelligent cameraBall/player tracking + virtual camera, at the venue
NEXT
Multi-cameraWide field coverage
NEXT
Next-gen hardwareJetson AGX Orin, multi-camera
Understanding the game, frame by frame.
01
Raw frameWide, small players, tiny ball.
02
DetectionPlayers and ball found every frame.
03
TrackingStable IDs and trajectories through occlusion.
04
Camera framingChoose what matters, frame it smoothly.
THE INTELLIGENT CAMERA
Detection is the input. Deciding where to look is the product.
The perception layer tells the system where every player and the ball are. Camera intelligence turns that into a decision: which region of the field matters right now, and how to move there without jitter.
01
Game stateBall position, nearby players, direction of play
02
Target regionWhere the action is — and where it's heading
03
SmoothingKalman filtering and PID control for natural motion
04
Virtual cameraA cropped, broadcast-style view of the wide feed
The hard parts.
01
Real-time inference
Process live video on edge hardware with latency low enough to keep up with play.
02
A small, fast ball
Tiny, motion-blurred, often occluded, and easy to confuse with other objects.
03
Player occlusion
Players overlap constantly; tracking must keep identities stable through it.
04
Wide field coverage
Cover enough of the pitch while keeping enough pixels on target for inference.
05
Natural camera movement
Follow the play without twitchy or seasick motion.
By the numbers.
SYSTEM SPEC
~150°
Field coverage
ACHIEVED
720p
Real-time inference on edge
TARGET
22+
Players + ball tracked
TARGET
92%
Ball continuity
TARGET
88→95%
Player detection improvement
Green = measured on the system. Grey = project objective.
Hardware to finished product.
Camera modules, MIPI CSI, edge compute, enclosure & mounting
Detection, tracking, ball and player perception
RTSP, GStreamer, GPU inference, low latency
Virtual camera, framing, game-aware movement
An automated sports video system
My role
- AI and system architecture
- Computer vision: detection and tracking
- Real-time video pipeline
- Edge deployment and optimisation
- Camera intelligence and virtual camera
Technology
Client project — selected technical details shared with permission and simplified for confidentiality.