This paper proposes a spatiotemporal deep learning framework for automatic detection of lameness in cattle. Using 50 publicly available video clips (42 cattle), we train and evaluate two deep learning architectures: a 3D CNN and a ConvLSTM2D. Data augmentation techniques are applied to improve generalization performance, and the 3D CNN achieves 90% accuracy (90.9% precision, recall, and F1 score, respectively), outperforming ConvLSTM2D (85% accuracy). Unlike existing multi-stage pipelined approaches that incorporate object detection and pose estimation, this study demonstrates the effectiveness of a direct end-to-end video classification approach. By successfully extracting and learning spatiotemporal features from diverse video sources, we enable scalable and efficient cattle lameness detection in real-world farm environments.
