Single Camera Kiosk: Train locally from administration

Train locally from administration

The admin Training data tab now contains the full local training workflow. Drawing boxes supplies ground truth; it does not immediately change the model. Confidence thresholds, smoothing, deadband and PTZ speed are tuning, not model training.

1. In PTZ & AI tuning, configure the analysis stream, connect the camera, save the wide preset and start operator tracking calibration.
2. In Training data, capture representative frames from several games. Label every visible player and ball; explicitly save empty frames as negatives. Use a separate group for each game. More varied, accurately labeled examples improve coverage; a handful of frames only tests the workflow.
3. Stop the broadcast. Click Install local training tools once. This creates a separate environment and downloads pinned CPU training dependencies; Ubuntu python3-venv must be installed. Allow internet access and several GB of disk.
4. Enter a completely separate validation game group. Start with 10 epochs, batch size 1, image size 640; 320 is an optional faster workflow check with less small-ball detail. Click Start local training. The kiosk snapshots the reviewed dataset, fine-tunes the included model, evaluates the held-out game, then exports a 640-pixel ONNX model. CPU thread count is capped at two and the pretrained backbone is frozen for conservative fine-tuning. The UI remains available while the job runs.
5. Review the log and validation results, including player and ball precision/recall/mAP. Training time on the N5095 depends on the dataset and may be substantial. No accuracy guarantee is implied by a completed run.
6. Confirm review and click Activate trained model. Activation validates ONNX compatibility, stops current AI tracking, and saves the new detector path. Start tracking calibration again and inspect detections and framing on another game before broadcast use. Restore included model provides rollback.

Local training runs only when explicitly started. It does not need the admin page to stay open; after automatic admin lock, sign in again to view it. Cancelling kills the local training process. An interrupted job is recorded after restart and never silently resumes. Camera setup and customer stream starts are blocked while training is active. An automatic scheduled broadcast or production truck handoff cancels training before taking priority. A completed model never replaces the active detector automatically.