Lameness detection in dairy cows from overhead view: high-precision keypoint localization and multi-feature fusion...

Area of Science:

• Animal Science
• Computer Vision
• Agricultural Engineering

Background:

• Overhead view lameness detection in dairy cows is challenging due to subtle back movement features and individual variability.
• Existing methods from an overhead perspective struggle with accuracy and generalization.
• RGB-D data offers potential for improved lameness detection by capturing depth information.

Purpose of the Study:

• To develop and evaluate an overhead-view lameness detection approach for dairy cows using RGB-D data.
• To create a high-precision keypoint detection method for cow's back to analyze posture and motion abnormalities.
• To design and validate novel lameness-related features and employ an unbiased feature selection framework.

Main Methods:

• Developed a keypoint detection network for the cow's back, modeling long-range spatial dependencies.
• Designed six lameness-related features, including four new indices, to capture posture and motion abnormalities.
• Utilized Random Forest Gini importance and permutation importance correction (PIMP) for unbiased feature selection.

Main Results:

• The keypoint detection network achieved 100.00% PCK@0.02 and 95.89% average precision.
• Back curvature, movement asymmetry index, and vertical oscillations of the back and head showed strong discriminative ability.
• The multi-feature fusion lameness detection model reached an overall accuracy of 91.00%.

Conclusions:

• Overhead RGB-D imaging combined with precise keypoint detection and feature fusion is a reliable strategy for dairy cow lameness detection.
• The proposed method significantly improves accuracy and robustness compared to baseline models.
• This approach provides valuable support for health monitoring and intelligent management in modern dairy farming.