Key behavior recognition with tracking for brooding layer breeder chick based on RTM-BHJ

Monitoring the behavior of layer breeder chicks during the brooding period is critical for evaluating their welfare and health status. However, high chick density, frequent occlusions, and rapid fine-grained movements make automated monitoring challenging. To address these issues, this study introduces RTM-BHJ, a real-time framework for multi-chick monitoring based on RTMDet and SlowFast. The framework integrates a Body–Head Joint (BHJ) detector, the NorFast tracker, which combines NorFair-v2 and FastReID, and a SlowFast-based Tracking-Motivated Behavior Recognition (SF-TMBR) model. The BHJ detector incorporates a novel detection head optimized by DSC, DA-SPP-Lite, and Flash Layer to improve detection accuracy while maintaining computational efficiency. For tracking, FastReID extracts discriminative appearance features from a pre-trained model. In addition, the temporal aggregation of detected head features along tracked trajectories enhances feature representation for recognizing key behaviors with rapid motion. Experiments across four chick age groups showed that the proposed framework achieved strong detection and tracking performance, with an mMR −2 of 20.8%, a HOTA score of 66.8%, and 272 identity switches. For behavior recognition, the SF-TMBR model achieved AP scores of 78.3% for preening and 72.1% for pecking, and improved Frame-mAP and Video-mAP by 3.9% and 3.8%, respectively, compared with the original SlowFast network. These results demonstrate that RTM-BHJ provides an effective and practical solution for intelligent monitoring of layer breeder chicks during the brooding period.