OmniPT: Unleashing the Potential of Large Vision Language Models for Pedestrian Tracking and Understanding