BroilerTrack: Automatic multi-camera multi-broiler tracking

Efficient and continuous tracking of individual broilers is critical for improving poultry management, welfare, and breeding decisions in commercial settings. However, standard Multi-Object Tracking (MOT) techniques face significant challenges in poultry environments due to occlusions, high object similarity, and dense flocks.In this work, we introduce BroilerTrack, a novel multi-camera multi-broiler tracking framework tailored for the poultry industry. Unlike traditional approaches that rely heavily on appearance features, BroilerTrack employs a position-based tracking strategy in a unified coordinate system (unified plane), thereby circumventing identity ambiguity caused by the homogeneous appearance of broilers. Our proposed BroilerTrack system comprises three key modules: Top-view Aggregation, Side-view Distribution, and Identification Assignment, enabling robust identification (ID) consistency across multiple calibrated views. Furthermore, we present a new Multi-View Broiler dataset collected under commercial-like conditions, featuring synchronized footage from six strategically placed cameras (two top-view and four side-view). Notably, our method requires no unified-plane annotations during training and achieves superior performance over state-of-the-art Multi-camera MOT methods on both detection and association metrics. This work provides a scalable, non-intrusive solution for real-time poultry monitoring, with strong potential for applications in behavior analysis, welfare optimization, and automated breeding selection.