Direct video-based spatiotemporal deep learning for cattle lameness detection.

Area of Science:

• Veterinary Medicine
• Artificial Intelligence
• Animal Science

Background:

• Cattle lameness significantly impacts animal welfare and farm productivity.
• Early and accurate detection of lameness is crucial for timely intervention and economic loss mitigation.

Purpose of the Study:

• To develop and evaluate a spatiotemporal deep learning framework for automated cattle lameness detection using video data.
• To compare the performance of 3D Convolutional Neural Networks (3D CNN) and Convolutional Long-Short-Term Memory (ConvLSTM2D) for this task.

Main Methods:

• A dataset of 50 video clips of cattle was curated and labeled as lame or non-lame.
• Data augmentation techniques were applied to improve model generalization.
• Two deep learning architectures, 3D CNN and ConvLSTM2D, were trained and evaluated for video classification.

Main Results:

• The 3D CNN model achieved 90% accuracy, with 92% precision, 90% recall, and 90% F1 score.
• The 3D CNN outperformed the ConvLSTM2D model (85% accuracy).
• The end-to-end approach demonstrated comparable accuracy to existing methods with a simpler, single-stage design.

Conclusions:

• Spatiotemporal deep learning, particularly using 3D CNNs, is effective for automated cattle lameness detection.
• The proposed framework offers a computationally efficient and simpler alternative to multi-stage methods.
• This approach is suitable for real-time deployment in livestock farming environments.