CageDAR-DeepSORT: An improved real-time multi-object tracking algorithm for caged chickens based on DeepSORT

• Proposed CageDAR-DeepSORT for caged chicken tracking • YOLOv11-AD: high precision, 2.09M params, 63.3 FPS • DARNet achieves 95.7% mAP and 80.1% Rank-1 in ReID • optimization track matching strategy for chicken environment • Reduced ID switches to 2, IDF1 score of 81.69% The egg-laying chicken farming industry is a crucial pillar of the global agricultural economy and protein supply; however, the large-scale, high-density cage-rearing model poses significant challenges for individualized precision management. Achieving stable and continuous identity tracking for each individual chicken is a key prerequisite for analyzing individual behavior, providing early disease warnings, implementing precision feeding, and assessing welfare. This is of great practical significance for improving breeding efficiency and safeguarding food safety. However, due to the highly similar appearances of chickens and severe occlusion in the cage environment, traditional tracking methods suffer from the problem of frequent identity (ID) switching. To address this, this study proposes an improved real-time multi-object tracking model, CageDAR-DeepSORT. This model synergistically optimizes three aspects: detection, re-identification (ReID), and data association. First, in the detection phase, the YOLOv11 detector is rendered lightweight by introducing the ADown module, significantly increasing detection speed while maintaining high precision. Second, targeting the discriminative features of chicken heads, a Dual-Aligned Re-Identification Network (DARNet) is designed to enhance the ability to distinguish similar individuals. Finally, in the tracking association phase, a strategy combining an improved association metric (CIoU) with trajectory rank matching is proposed, effectively mitigating ID confusion caused by brief occlusions. Experimental results indicate that the lightweight detector achieves a real-time detection speed of 63.3 FPS while reducing the parameter count to 2.09M. DARNet achieves a mean Average Precision (mAP) of 95.7% on the chicken re-identification dataset. The integrated tracking system demonstrates outstanding performance on key metrics: the number of ID Switches (IDSw) is reduced to 2, and the IDF1 score reaches 81.69%, significantly outperforming various baselines and state-of-the-art algorithms. This study provides a highly efficient and reliable solution to the challenge of stable individual chicken tracking in cage-reared environments, laying a solid technical foundation for subsequent precision breeding management and intelligent analysis.