Spatio-Temporal-Based Identification of Aggressive Behavior in Group Sheep.

Area of Science:

• Agricultural Technology
• Animal Behavior
• Computer Vision

Background:

• Manual observation of group sheep aggression is inefficient and subjective.
• Automated detection systems are needed to improve livestock management and welfare.
• Existing image-based models struggle with occlusion, leading to false detections.

Purpose of the Study:

• To develop and evaluate a video streaming-based model for automated detection of aggressive behavior in group sheep.
• To enhance the efficiency and objectivity of sheep aggression monitoring.
• To address limitations of image-based detection models in handling feature occlusion.

Main Methods:

• Collected and labeled video data of sheep, including daily routines and aggressive incidents.
• Utilized YOLOv5 for sheep detection and coordinate extraction in video frames.
• Implemented a novel sheep tracking heuristic to sort coordinate data.
• Employed a Long Short-Term Memory (LSTM) framework for aggression prediction.
• Optimized model parameters including confidence, batch size, and frame skipping.

Main Results:

• The proposed video streaming model achieved 93.38% precision and 91.86% recall.
• The model effectively handles feature occlusion, a common issue in image-based detection.
• Demonstrated superior performance compared to traditional image-based detection methods for sheep aggression.

Conclusions:

• The video streaming-based model offers a more efficient and objective solution for detecting group sheep aggression.
• This approach overcomes the false detection problem caused by occlusion in image-based systems.
• The developed model has significant potential for improving sheep welfare and farm management through automated monitoring.