Can computers recognize when chickens are playing?

Monitoring animal welfare usually focuses on distress signals, missing positive indicators like play. This work builds PlayClass, a video pipeline that tracks individual birds in poultry pens and classifies play behavior using foundation models. V-JEPA, a self-supervised video model, combined with hand-crafted motion features achieved 77% F1—outperforming simpler baselines despite occlusions and kinematic ambiguity between play and non-play movements.