Multi-camera detection and tracking for individual broiler monitoring

Abstract

Welfare concerns in poultry farming have driven the need for advanced monitoring solutions to study broiler activity and health. However, existing research predominantly relies on single-camera setups, which are prone to occlusions from equipment such as feeders and lighting, limiting their effectiveness. To address this, we propose a multi-camera setup that enables comprehensive broiler localization and tracking from a top-down view of the pen. To support this approach, we introduce MVBroTrack,1 an open-source dataset containing real-world data with annotations for various subtasks critical to broiler studies. We demonstrate robust performance of our multi-view detection pipeline throughout the six-week broiler lifespan despite significant changes in visual appearance. Additionally, we present a novel unsupervised tracking method that surpasses the traditional tracking by detection paradigm, improving the IDF1 score by 3% and increasing the proportion of mostly tracked broilers by 5%. By reducing the need for manual observation, our multi-camera pipeline facilitates exhaustive studies of broiler behavior and welfare, paving the way for significant advancements in poultry research and farming practices.