How Ideas2IT Built a Custom Sports Tracking Platform for Live Broadcast Event Across US Clubhouses

Edge architecture and MVP: object tracking, camera switching, and a working prototype by month 3

The foundational constraint was latency. Real-time object tracking at 30-120 FPS with a live composited feed cannot route through the cloud: every processing decision had to run on an NVIDIA CUDA edge server on site.

Ideas2IT built two tracking architectures. In the first, a Camera+LiDAR fusion pipeline synchronised PTZ RGB frames and LiDAR point clouds via a common system clock, projecting bounding box UV coordinates into the LiDAR frustum to fuse a full 3D ball position. In the second, camera and LiDAR were integrated at the hardware input layer, combining RGB frames and point clouds before object detection, producing a 3D trajectory output in a single pipeline.

A Kalman Filter smoothed the predicted trajectory path across both architectures. PTZ encoder feedback drove pan-tilt-zoom commands from the object position in real time. Player check-in used QR code scan or identity card at the tee; a draw-based tee order system assigned the sequence for the first round, with leaderboard ranking applied after each shot. When a player stepped to the tee, the system triggered via admin confirmation or automated computer vision identification, and Twilio SMS notified their registered contacts immediately.

Deliverables

• Camera+LiDAR software fusion tracking pipeline
• Camera-LiDAR integrated hardware tracking pipeline
• PTZ encoder feedback loop
• Admin and CV tee-trigger system
• Twilio SMS spectator notification engine
• Electron venue display app
• Live streaming integration

Trajectory overlay, distance engine, and highlights: composited display and leaderboard for the full venue experience

The distance-to-target calculation was the hardest engineering problem in this phase. Camera pixels must map to real-world coordinates on the playing surface, but the target position changes.

Ideas2IT built the distance engine on OpenCV Homography with NumPy, with intrinsic camera calibration handled via a configuration layer. The admin's tablet interface let them tap the new target position on a surface map, triggering automatic recalibration of the entire coordinate system without engineering intervention.

On the display side, an OpenGL/Vulkan renderer composited the trajectory line, distance label, and leaderboard graphics directly onto the live feed at full framerate with sub-second latency. The daily and all-time leaderboard tracked across sessions. An automated highlights generation layer in the AI stack clipped and packaged each shot for replay. Between live shots, an idle content rotation layer managed advertising and branded content display, giving venues a direct revenue surface.

Deliverables:

• OpenCV Homography pixel-to-distance engine
• OpenGL/Vulkan trajectory overlay compositor
• CareFull-framerate composi
• Idle content and advertising rotation
• Full admin interface for operations

Production readiness and multi-venue deployment: monitoring, OTA updates, and a toolkit for scale

Scaling from one venue to many required a different class of infrastructure problem than building the tracking layer itself. Each venue runs an independent edge server; Ideas2IT built auto-recovery into the edge OS so display outages self-healed without on-site engineering intervention.

An MLOps layer handled model versioning and retraining as new shot data accumulated across venues, keeping detection accuracy improving without manual pipeline management. A cloud backend on AWS with Docker and CI/CD handled centralised monitoring, remote diagnostics, and OTA updates pushed to all venue edge servers simultaneously.

Amazon ElastiCache reduced query latency on the leaderboard and session data layer. PostgreSQL at the edge synced shot and leaderboard data to the cloud layer for cross-venue aggregation. The deployment toolkit packaged the full venue setup so the client could onboard a new venue without bespoke engineering work at each site.

‍

Deliverables:

• Production-optimised object tracking system
• MLOps and model retraining pipeline
• Edge server auto-recovery setup
• AWS cloud backend for centralised management
• Multi-venue deployment toolkit