United KingdomAgriculture and food production2023 – ongoingHighfields Farm
AI camera mobility scoring for dairy cows
A 1,000-cow dairy farm used an overhead AI camera to catch lame cows earlier, at about £1 per cow a month
Problem
Lameness hurts milk yield and cow welfare, and it is usually spotted by people scoring each cow's walk by eye, which is periodic and subjective. On a farm milking 1,000 cows three times a day, manual scoring cannot follow every animal.
Approach
Highfields Farm in Audlem, Cheshire, mounted a CattleEye camera 4 m above the exit of its 50-point rotary parlour. Software marks key points on each cow as she walks out, gives a gait score from 0 to 100, groups cows into quartiles in a cloud app and links scores to electronic ear tags. Footage is uploaded after milking; set-up cost about £500 and the service about £1 per cow per month.
Results
Takeaway
The farm reported catching lame cows earlier and lameness falling to the low teens in percent. It could not separate the camera's effect from the culling of persistently lame cows.
Categories
- Size
- Small business
- Industry
- Agriculture and food production
- Business function
- Production and quality
- Innovation type
- Process innovation
- AI technology
- Computer vision
- Adoption
- Vendor solution
- Evidence
- Independent study
Sources
Source summaries
The farming weekly visits Highfields Farm, which milks 1,000 cows three times a day, and describes how the CattleEye camera scores each cow's gait after milking. The farm's office manager says lame cows are caught earlier and prevalence has fallen to the low teens, though culling makes a lasting effect hard to confirm. Farmers Weekly reports costs of about £500 for set-up and £1 per cow per month, and cites Liverpool University tests showing an 89% match with independent scorers.
A peer-reviewed study compares the camera system's scores with trained assessors using 6,040 mobility scores from three UK dairy farms. Percentage agreement and Gwet's coefficient stayed above 80% and 0.80, while Cohen's kappa was only fair to moderate. For detecting painful foot lesions the system had a sensitivity of 0.52 and specificity of 0.81, against 0.29 and 0.89 for a trained vet.
Full details are in the original sources.
