Cow key point detection in indoor housing conditions with a deep learning model

Area of Science:

• Animal Science
• Agricultural Engineering
• Computer Vision

Background:

• Lameness in dairy cattle is a significant economic and animal welfare issue.
• Current lameness detection methods are labor-intensive and infrequent.
• Automated gait assessment offers potential for frequent, low-labor locomotion monitoring.

Purpose of the Study:

• To automate the detection of key anatomical points on dairy cows in an indoor farm setting.
• To develop postprocessing steps for key points to enable gait feature calculation.
• To lay the groundwork for camera-based automated locomotion monitoring in real farm environments.

Main Methods:

• Utilized side-view video footage of 34 Holstein-Friesian dairy cows.
• Trained a deep learning model (T-LEAP) to detect 17 key points on cows.
• Assessed model performance using Percentage of Correct Key points (PCKh0.2), achieving 89% overall on the test set.

Main Results:

• The T-LEAP model achieved 89% PCKh0.2 for detecting 17 key points on dairy cows.
• Key points on the cow's back showed the poorest detection performance (59% PCKh0.2), especially in the horizontal direction.
• Identified necessary postprocessing steps including camera distortion correction and erroneous key point rectification.

Conclusions:

• Automated key point detection is feasible in practical indoor farm settings, even with occlusions.
• Postprocessing techniques are crucial for translating detected key points into reliable gait features.
• This study provides essential prerequisites for developing automated locomotion monitoring systems to improve dairy cattle welfare.