Player Tracking in Amateur Football Footage

We are building a free analytics platform for an amateur soccer club as a side project. The goal is to track players across full games, output structured statistics per player, and feed that data into a data warehouse and dashboard for the club — giving us something richer than what the club's current camera subscription provides. This engagement is a proof of concept on a short clip before we scale to full 90-minute games and eventually multiple games across a full season. The footage We have two sample video clips from the same game. We have plenty more footage from different games but all from the same camera setup with no alternative angles available. The first clip is a screen recording of the panoramic wide-angle stationary view showing the full pitch. The second is a downloaded ball-tracking follow-cam where the camera moves with the ball — players outside the action are cut off but the ball and nearby players are clearly visible. These two clips are from the same match and are the format we will be working with going forward. What we want For a 10-minute passage of active play we want: Individual player tracking with stable persistent identities across the clip Per-player statistics output — position over time, distance covered, pitch zones Ball tracking and position output Event detection — goals, free kicks, corners, goal kicks Team assignment per player A clean JSON or CSV output of all of the above for ingestion into a data warehouse later The key challenge and the most important deliverable is individual player identity — we need to know that player A in frame 1 is the same player A in frame 500, with their statistics correctly merged across the full clip. Detection of players is largely working. What we need is stable, accurate, merged per-player tracking. Technical context A PoC repo is active built around the Roboflow sports codebase. A CLAUDE.md session file is available which documents the full project context. The current pipeline uses RF-DETR (Roboflow football-players v20 via ONNX) as the detector running fully locally on a Mac M5 Pro. The engineer is welcome to review what exists, improve it, replace parts of it, or take a different approach entirely — whatever gets the best result on this footage. Critical requirement — fully offline All inference must run locally with no per-frame API costs. We are looking at 90-minute games across a full season so any hosted API approach is not viable at scale. The current ONNX local inference pattern should be maintained. Success criteria Stable player identities across a 10-minute clip with merged statistics per player Ball tracking and event detection present in the output Team assignment per player visually verifiable Clean JSON/CSV output ready for data warehouse ingestion All inference runs locally with no per-frame API cost Code handed over with documentation on what was built and how to run it Budget: $400–600 fixed price If this 10-minute clip works correctly the next step is full game analysis and eventually multi-game season tracking. The engineer who solves this well is well placed for that ongoing work.