Single Camera Kiosk: Training and tuning for your application

Training and tuning for your application

Open Train for your games after operator unlock. Start calibrated tracking, give the capture group a game-specific name, and capture frames explicitly. Drag a Player or Sports ball box around every relevant visible object; undo and correct boxes before saving reviewed labels. Include difficult examples: far-side balls, partial occlusion, motion blur, white uniforms/balls, different light/weather, and frames with no ball. Saving a reviewed frame with zero boxes makes an explicit negative image. The kiosk never records a training dataset in the background.

Capture multiple different games. Select one whole game group as validation and export the ZIP. Unreviewed frames are excluded. Whole-game separation helps avoid evaluating on near-identical adjacent frames from the training game. Export preserves all 80 COCO class indices: person=0 and sports ball=32; only these classes are annotated here. You do not need to label the other 78 classes.

On a separate Ubuntu training workstation, extract the ZIP and install the optional training dependencies in a dedicated venv:

python3 -m venv .training-venv
.training-venv/bin/pip install -r training/requirements.txt
.training-venv/bin/python training/train_model.py --data /absolute/path/dataset/dataset.yaml --device cpu --epochs 50 --batch 4

Use --device 0 when a supported GPU/PyTorch installation is available. Do not train on the N5095 while it is responsible for a live stream. The training script fine-tunes the included checkpoint, evaluates the held-out group, exports a compatible fixed 640-input ONNX file, and saves metrics/provenance next to the trained model. Fifty epochs is a starting setting, not a quality guarantee. Start with diverse, accurately labeled frames and evaluate ball precision/recall separately from player precision/recall.

Copy the exported ONNX to the kiosk's local models directory, stop AI tracking, enter its path in operator PTZ setup, and save. Retain the included pretrained model as rollback; blank model path restores it. Commission each replacement on an independent game before using it unattended. A model with low ball recall should retain the wide/group fallback rather than increasing zoom.

Tune player confidence, ball confidence, center deadband, smoothing, maximum movement speed, analysis cycles/second and lost-target delay through the operator form. Tune framing/control separately from detector training: thresholds do not retrain the model. Current observations/status are visible in the operator state; log frame detections by collecting labeled examples rather than assuming tracking success from camera movement.

Model licensing and provenance are recorded in models/MODEL_CARD.md and the included license file. Review those terms for your deployment/distribution.