Single Camera Kiosk: Program video and trained detector

Program video and trained detector

Minimum program setting: 1920×1080 at 30 fps. 1080p60 is selectable. Set the camera to matching output in camera setup; the kiosk configures OBS video settings after outputs are stopped. This does not prove the PC can encode/stream at 60 fps while tracking; verify dropped frames and sustained load on the actual hardware. Hardware encoding must be configured in OBS if used.

The included models/yolov8n.onnx is a pretrained COCO detector for people and the generic sports-ball class (including basketballs and soccer balls). Leave the model path blank to use it. The original official checkpoint and model provenance/license are included. It is not trained on your venue or camera. Built-in HOG/SVM is available as a people-only fallback.

Analysis uses a low-resolution RTSP substream of the same camera, independent of NDI program output. Enter its actual camera-configured URL; do not use the event publishing address. The detector evaluates a 640×640 letterboxed view at up to five analysis cycles/second, using two CPU threads. Camera frames are consumed independently to avoid inference backlog. Older than one-second observations are discarded. Slow inference stops movement rather than issuing delayed commands.