Automated mating monitoring and occlusion-aware identity attribution in breeding chickens using multi-stage deep learning

Description

Monitoring mating behavior is essential for evaluating reproductive health and optimizing fertilization rates in poultry breeding. However, conventional manual observation is labor-intensive, prone to fatigue, and insufficient for accurately identifying individuals within dense flocks. To address these limitations, we introduce a dataset and corresponding framework for individual-level mating event attribution in poultry systems. Unlike prior work that focuses on flock-level or event-level detection, our dataset enables persistent identity tracking across mating events, supporting per-individual and longitudinal analysis of reproductive behavior. The dataset is collected under challenging real-world conditions, including a uniform single-breed flock, top-down surveillance views, and frequent severe occlusions caused by rooster mounting. Under these conditions, appearance-based re-identification methods are unreliable. To handle this, we provide annotations supporting occlusion-aware identity inference, where the identity of occluded hens is resolved through interaction-aware temporal and spatial reasoning. Each subject is equipped with a lightweight wearable vest-tag system designed with a robust color–symbol encoding scheme to provide deterministic identity signals. This design ensures identity consistency even under partial or full occlusion. The dataset is accompanied by a processing pipeline that integrates YOLO11 for rooster detection and tracking, and UniFormer for spatiotemporal mating behavior recognition. We further provide an occlusion-aware spatiotemporal matching module that combines temporal gap detection, last-observed-position estimation, and spatial validation within dynamically defined interaction zones to infer participant identities during mating events. Experimental evaluations on real-world farm surveillance videos demonstrate the effectiveness of the proposed dataset and framework. The proposed method achieves a mean F1-score of 91.10% across three independent test days under natural class imbalance. The identity matching module correctly resolves both participating individuals in 97.6% of evaluable mating events. Overall, this dataset provides structured, identity-aware mating event annotations and enables automated generation of longitudinal mating logs, facilitating quantitative analysis of reproductive behavior and supporting data-driven poultry breeding management.

Institutions

• Nanjing Agricultural University

  Jiangsu, Nanjing

Categories

Licence