Automated 24-h cattle feeding behaviour analysis using Raspberry Pi, infrared imaging, and deep learning-based head position and ID classification

Monitoring cattle feeding behaviour is essential for assessing animal wellbeing. While recent vision-based systems advance behavioural analysis, many cannot link behaviour to feed intake, operate day and night, or detect interactions like muzzle-to-feed contact. They also rely on small, unrepresentative datasets and complex architectures that hinder collaboration. This study presents a low-cost, scalable framework using infrared imaging and motion-triggered video capture to identify cattle feeding behaviours and presence at the feed bunk. Using a Raspberry Pi with an infrared camera and open-source motion detection software, 503 video recordings were collected across varied light conditions. From these, two supervised models were trained: a convolutional neural network (CNN) for head-position classification, including muzzle-to-feed contact (test accuracy: 94.86%, loss = 0.17) and another for individual identification. Trained on 10,103 frames, the animal identification model distinguished feeder occupancy and specific individuals with high test-set accuracy (98.52%, loss = 0.05), dropping to 89.5% when evaluated on an external dataset of 25,453 manually annotated frames. This discrepancy highlights challenges in generalising to real-world conditions. Together, the models link feeding events to individual animals and feed bin weights to estimate intake. Future work will refine ingestion behaviour detection, while modularity allows potential deployment on edge or server-based farm management platforms. Findings show simple systems enable robust behavioural monitoring and interdisciplinary collaboration by lowering technical barriers for animal scientists. The framework supports scalable, data-driven livestock management and integration with precision farming systems.