Puzzle: taking livestock tracking to the next level

Area of Science:

• Animal behavior research
• Computer vision applications in animal science
• Artificial intelligence for livestock management

Background:

• Understanding animal behavior is crucial for health and welfare management.
• AI-powered camera systems offer non-invasive monitoring of traits like activity and space utilization.
• A key challenge is individualizing animal data, known as the multiple object tracking problem.

Purpose of the Study:

• To introduce an novel AI-based solution, "Puzzle," for the multiple object tracking problem in animal behavior analysis.
• To train a Convolutional Neural Network (CNN) for deriving animal appearance clues and associating detections with individual IDs.
• To assess the efficacy of the "Puzzle" method in outdoor goat tracking and its impact on behavioral studies.

Main Methods:

• The "Puzzle" method begins with straightforward video sequences to train a CNN on animal appearance.
• The trained CNN is applied to the entire video, combined with distance metrics, for robust animal identification.
• Evaluation involved outdoor goat tracking, analyzing ID association criteria (location, appearance, or both), and impact on space utilization and activity estimations.

Main Results:

• The "Puzzle" method achieved over 90% successful tracking in outdoor goat scenarios.
• Relying solely on appearance-based tracking, after tailoring the CNN, yielded satisfactory results.
• Behavioral estimations (space utilization, activity) showed a low error rate, below 10%.

Conclusions:

• The "Puzzle" paradigm offers an effective solution for the multiple object tracking challenge in animal behavior studies.
• Tailoring appearance-based CNNs to specific video data is a viable strategy for accurate animal individualization.
• The open-source nature of the method and linked data-paper promote accessibility and advancement in deep learning for livestock research.