Real-time detection of dead broilers in high-density poultry houses using high-resolution video and semantic segmentation

This study proposes a dead broiler detection method that combines lightweight semantic segmentation with temporal modeling. In high-density poultry house scenarios, the proposed approach achieves an overall detection accuracy of 88.64 % on a mixed test set containing both dead-broiler and non-dead-broiler samples. Unlike previous studies that rely on single-broiler state analysis, the proposed method is designed for practical commercial farming environments and is capable of reliably localizing 0-4 dead broilers among approximately 1,200 broilers within the field of view of a single camera. The proposed broiler segmentation model maintains a favorable balance between segmentation accuracy and computational efficiency, achieving real-time inference at 34.12 frames per second on 1600 × 2880 resolution images. Experimental results demonstrate that the proposed method enables robust and reliable localization of dead broilers in real-world poultry farming environments, highlighting its practical applicability and providing a feasible foundation for future deployment under higher-resolution wide-angle camera systems.