Robust Tracking and Quantification of Open-Beak Behavior for Heat Stress Assessment in Multi-Tier Caged Rearing Broiler Breeders via Mobile Inspection

Panting is one of the earliest and most prominent behavioral responses of chickens under heat stress. More importantly, panting is inherently a temporally continuous state rather than an instantaneous event. However, existing vision-based methods usually determine panting from beak opening in individual frames, which contradicts the nature of panting and therefore cannot provide reliable assessment. To address this limitation, this study proposes a Cumulative Panting Ratio (CPR), which quantifies sustained open-beak behavior associated with panting by tracking and accumulating instantaneous open-beak events within a temporal sliding window. To support consistent estimation of sustained open-beak behavior and CPR, a cascaded Detection–Tracking–Classification pipeline is constructed to explicitly decouple body-level tracking from fine-grained beak behavior inference, thereby reducing the conflict between global localization and local feature extraction. In addition, a Near-Online Trajectory Stitcher is proposed to reconnect fragmented tracklets using unified temporal, spatial, scale, and directional constraints induced by platform movement. Experiments on practical inspection datasets demonstrated robust tracking continuity and stable quantification performance. Compared with state-of-the-art frameworks, the proposed method reduced identity switches from 10 to 4 and improved IDF1 by 2.43%, while minimizing counting errors across diverse detector architectures. With TensorRT acceleration and multi-stream parallel processing on the Jetson Orin NX, system throughput increased from 7.89 FPS to 51.43 FPS without compromising accuracy. Overall, this work provides a practical solution for quantitative heat stress assessment and supports the deployment of physiological state monitoring systems for precision livestock farming.