CSP-YOWO-TrajNet: A spatio-temporal detection method for laying hen heat stress behavior

Area of Science:

• Animal Science
• Computer Vision
• Artificial Intelligence

Background:

• Laying hen welfare and production efficiency are significantly impacted by heat stress.
• Monitoring heat stress-induced behavioral changes in battery-cage systems is challenging due to occlusion, coarse labels, and limited edge device resources.

Purpose of the Study:

• To develop a real-time detection system for heat stress behavioral changes in laying hens.
• To address challenges in battery-cage systems, including bird occlusion, coarse behavior labels, and edge device limitations.

Main Methods:

• Proposed a three-dimensional spatiotemporal action detector, CSP-YOWO-TrajNet.
• Integrated a 2-D CNN (CSPDarknet53-SPA) for spatial feature extraction and a 3-D CNN (3-D ResNeXt-50) for temporal modeling.
• Incorporated a TrajNet module for trajectory prediction, enhancing perception and tracking.

Main Results:

• Achieved 94.1% precision and 96.1% mAP@50 on a dataset of 281 hen-behavior clips.
• Outperformed the YOWO baseline by increasing precision by 3.0% and mAP@50 by 3.6%.
• Reduced model size from 134.0 MB to 78.8 MB, with tracking F1 score at 95.4% and accuracy at 89.2%.

Conclusions:

• CSP-YOWO-TrajNet enables real-time recognition of heat-stress behaviors in laying hens.
• The detector is suitable for deployment on edge devices, offering a practical solution for welfare monitoring.
• The proposed method effectively overcomes challenges in battery-cage systems, improving detection accuracy and efficiency.