AI Livestock Health Monitoring

Most UK dairy herds still find lame cows the same way they did twenty years ago. Someone notices a cow lagging on the way to the parlour, or the foot trimmer’s list fills up when they come in, or a trained mobility scorer walks the herd once a month and hands over a sheet. Studies of UK herds have repeatedly put lameness prevalence at around 30% of cows at any given time. The usual reason is not that farmers don’t care. It’s that a score 2 cow (impaired mobility, still walking freely enough to go unnoticed) looks fine to a busy person bringing in 250 cows at 5am.

A lameness detection camera on a dairy is meant to close that gap. It watches every cow on every exit from the parlour, scores how she walks, and flags the ones getting worse before they reach score 3. This page covers how those systems work, which ones UK farms are using, what they cost against what lameness costs, and where AI health monitoring for beef and sheep currently stands. The aim is to help you decide whether one is worth installing on your unit, and how to get value from it once it’s in.

What a lameness detection camera actually measures

There are two main designs, and the difference matters when you’re comparing quotes.

2D video systems use an ordinary IP camera mounted above or beside the exit race. Software tracks points on the cow’s body (head, withers, spine, hooks, pins) across a few seconds of walking and works out the features a human scorer looks at: head bob, how arched the back is, stride length, whether the hind feet track up into the front prints, and how even the steps are. CattleEye is the best-known UK example of this approach. It was developed in Belfast and is built to work from a single overhead camera.

3D depth-camera systems use a time-of-flight or structured-light camera that measures the shape of the cow’s back in three dimensions. HerdVision works this way and reports mobility and body condition score from the same unit. DeLaval’s Body Condition Scoring camera is also 3D, though it’s built for BCS rather than mobility.

Either way, the system has to know which cow it is looking at. Some read the cow’s EID tag through a race-side reader or take the ID from the parlour’s own records in exit order. Others identify cows from their coat markings. For black-and-white Holsteins, marking-based ID can be very accurate. On a herd of solid-black or solid-red cows it can struggle, so ask about this if your herd isn’t predominantly patched.

The output is normally a score per cow per milking, rolled up into a daily or weekly trend. Here’s what a useful alert list looks like once it has been reduced to the cows that need action:

Herd: 248 in milk   Week 41   Mobility alerts (rolling 7-day score)

Cow    Lact  DIM   Score(7d)  Score(prev 7d)  Trend   Action
1142   3     87    2.4        1.6             ▲▲      Lift foot this week
0981   2     201   2.1        2.0             ▬       On trimmer list already
1310   1     34    1.9        1.1             ▲▲      Check: fresh heifer, early warning
0756   5     260   2.8        2.7             ▬       Chronic: discuss cull/dry-off
1207   2     143   1.8        1.2             ▲       Watch

Herd mean score: 1.12 (last week 1.15)
Cows at score 2+: 19 (7.7%)   Cows at score 3: 3 (1.2%)

The scoring scale is the AHDB Dairy Mobility Score, 0 to 3. Most systems report a decimal average because they’re combining dozens of walks. Score movement matters more than the absolute number. A cow going from 1.1 to 1.9 in a fortnight is the one to catch. A chronic cow sitting at 2.7 is a management decision, not an alert.

Camera, collar, leg or bolus: matching the sensor to the problem

Cameras aren’t the only route to finding lame cows earlier, and many herds already have a collar or tag system that partly covers it. Before spending money, check what the sensors you already own can tell you.

If you run SenseHub or a similar collar for heats, you’ll sometimes see a lame cow show up as an activity drop. That happens late. By the time eating and rumination fall, she’s usually score 3. Leg sensors do better because lying time rises measurably in lame cows, and research from IceRobotics’ Scottish work tied lying-bout changes to lameness. Cameras are the only option that looks at the gait itself, so they’re the only ones that reliably separate score 1 from score 2.

The UK systems worth shortlisting

If you’re seriously considering a camera, our detailed comparison at /automated-mobility-scoring-dairy-cows/ goes through how the camera systems compare on accuracy, installation and subscription terms. In brief:

CattleEye. A single overhead 2D camera on the exit race, with processing done in the cloud. It scores each cow every time she passes and sends alerts to a phone or web dashboard. It suits rotary and herringbone parlours where cows leave in single file. Because the hardware is a standard camera, the installation work is mostly about the mount, power and a reliable network connection.

HerdVision. A 3D camera unit that reports mobility and body condition together. If you’re already paying someone to body condition score around drying off and early lactation, getting both from one unit changes the sums. It needs a clean, single-file walk under the camera, and lighting and gate position affect it more than some 2D setups.

Parlour-integrated options. DeLaval, Lely and GEA all sell health and condition tools that plug into their own herd management software. They’re worth asking about if you’re already in one of those ecosystems, but check whether the product actually scores mobility or only BCS and activity. Sales literature often blurs the two.

Robotic herds. On a Lely Astronaut or DeLaval VMS unit, you already get box-level data: weight distribution across the four legs on some models, kick-offs, milking frequency and refusals. A cow who visits less often and spreads her weight unevenly is a lameness suspect. Before adding a camera, pull a report of cows with milking frequency down more than 0.5 visits a day over a week and compare it with your trimmer’s findings.

Does it score as well as a trained person?

This is the right question, but the comparison needs some care.

Two trained human scorers using the AHDB scale don’t agree perfectly either. Published studies put inter-observer agreement on a four-point scale anywhere from moderate to good. Kappa values are often quoted between about 0.4 and 0.7, and most disagreement is between scores 1 and 2, which is exactly where early detection happens. So when a supplier says their system “agrees with an expert X% of the time”, ask these four things:

1. Which statistic? Raw percentage agreement looks high because most cows are score 0 or 1, and anyone who called everything score 1 would agree a lot of the time. Weighted kappa, or sensitivity and specificity for detecting score 2+, are more honest.
2. Against whom? One RoMS-registered scorer, or a consensus of several?
3. On which herds? Trials on one well-lit, all-Holstein research unit tell you less than results from twenty commercial farms.
4. Single walk or averaged? A camera averaging 14 walks a week should beat a single snapshot score, because a cow who stepped on a stone on Tuesday doesn’t swing the result.

Where cameras clearly win is frequency. A monthly human score gives you 12 data points per cow per year. A twice-a-day camera gives you more than 600. Even if each camera score is a little noisier than an expert’s, the trend is far more reliable, and the trend is what tells you to lift a foot.

Costing it: a worked example on a 250-cow herd

Prices change and suppliers quote differently (some charge per cow per month, others an install fee plus a flat subscription), so the figures below are assumptions for you to replace with your own quote. Use the structure, not the numbers.

The herd: 250 cows, year-round calving, herringbone parlour, Red Tractor assured, monthly mobility scoring by a contracted scorer.

Cost of lameness now. The figure most often used in the UK is roughly £330 per case, from AHDB-funded work. It covers lost milk, treatment, extra culling and fertility losses. Newer estimates vary, and some go considerably higher once fertility is costed properly, but £330 is a cautious working number. Assume an incidence of 50 new cases per 100 cows per year, which is not unusual:

Annual cases       = 250 cows × 0.50          = 125 cases
Annual cost        = 125 × £330               = £41,250

What earlier detection is worth. UK trial work (including studies by Leach and colleagues, and by Thomas and colleagues at Nottingham) found that cows treated promptly at score 2 recover better than cows found at score 3, and that early trimming plus an anti-inflammatory improved cure rates. Assume the camera cuts the cost of the average case by 20% through earlier treatment, and avoids 10% of cases outright because repeat offenders are caught before they relapse:

Cases avoided      = 125 × 0.10               = 12.5 → 12 cases
Saving (avoided)   = 12 × £330                = £3,960
Remaining cases    = 113
Saving (earlier)   = 113 × £330 × 0.20        = £7,458
Total benefit      =                            £11,418 / year

Labour you might drop. If the camera replaces some or all of your monthly scorer visits, count that too. Many milk buyers and Red Tractor still expect a trained person’s score on record, though, so don’t assume it all goes. Suppose you cut from monthly to quarterly visits at £0.80 per cow per visit:

Before: 12 visits × 250 × £0.80               = £2,400
After:   4 visits × 250 × £0.80               = £800
Saving                                         = £1,600

System cost (assumed). £3,500 installation including mount, cabling and network, plus a £2.50 per cow per month subscription:

Year 1: £3,500 + (250 × £2.50 × 12)           = £11,000
Year 2 onward:   250 × £2.50 × 12             = £7,500

The result. Year 1 benefit is £13,018 against £11,000 of cost. From year 2 it’s £13,018 against £7,500. That works on paper. But look at what’s carrying it: the £7,458 from earlier treatment only appears if someone actually lifts the flagged feet within a few days. If alerts sit unread for three weeks, the benefit collapses to the avoided-case line and the system loses money. Run the sums with your own incidence (your vet or trimmer records will give you this), and halve the benefit figure to see whether the case still holds when things go less well.

Getting the installation right

Most disappointing camera installs fail on the race, not the software. Before the engineer arrives:

• Single file, steady walk. The camera needs 4 to 6 metres of straight race where cows walk at their own pace. A sharp turn straight out of the parlour, a scraper running across the walkway or a person standing at the exit gate will all distort gait.
• Surface. Scores are only comparable over a consistent surface. A race that’s grooved concrete in one section and rubber matting in another will shift readings when cows step onto the matting.
• Lighting. 2D systems need even light. Low winter sun across the race at afternoon milking causes trouble, and a cheap LED floodlight on a timer often fixes it.
• Connectivity. Cloud-processed systems need a stable connection from the parlour. If your Wi-Fi drops when the vacuum pump starts, run a cable or fit a point-to-point link before installation day rather than after.
• ID matching. If the system takes cow order from the parlour software, check that sort gates, cows returning through the parlour, and the occasional cow who backs out don’t scramble the sequence. Get the supplier to show you their mismatch rate after the first fortnight.

Ask for a two-to-four-week calibration period in which your usual scorer scores the herd on the same days. Compare the lists. Expect disagreement on individual score 1 and score 2 cows, but the cows ranked worst should largely overlap.

Turning alerts into trimmed feet

A camera produces a list. Whether it does anything useful is down to what happens to that list. Farms that get value from these systems usually follow a routine like this:

Monday morning, 10 minutes. The herd manager opens the dashboard and filters for cows whose score has risen by 0.5 or more over 14 days, plus any new score 3s. That’s usually 4 to 10 cows on a 250-cow herd once the backlog is cleared.

Same day. Those cows are drafted at the next milking using a sort gate, or marked with tail paint. Somebody trained, whether the herdsperson, a staff member on a foot trimming course, or a professional trimmer on a weekly contract, lifts every flagged foot within 48 hours.

Record the lesion. Sole ulcer, white line, digital dermatitis or sole haemorrhage. Put it in the same software the camera feeds, or in your herd software (Uniform-Agri, Herdwatch, NMR Herd Companion). Over a year this tells you whether you have a mostly infectious problem, which points to footbathing and slurry, or a claw horn problem, which points to cubicles, standing time and transition.

Check in 14 days. Treated cows should drift back down. A cow still above 2 two weeks after treatment needs a second look or a vet discussion.

The first month is always heavy because the camera finds the backlog of chronic cows nobody had listed. Plan for an extra trimmer day in month one. Don’t conclude the system is “too sensitive”.

Compliance, milk contracts and who owns the data

Red Tractor dairy standards require routine mobility scoring and an action plan for lame cows. Several milk buyers go further. Arla, Müller and the retailer-aligned contracts for Tesco, Sainsbury’s and M&S have all pushed for more frequent scoring and for herd lameness figures to be reported. Before you assume a camera satisfies your contract, ask your milk buyer’s field staff directly whether automated scores are accepted on their own, accepted alongside a periodic RoMS score, or not accepted. Answers differ, and they have been changing.

On data, read the subscription terms for three things:

1. Export. Can you download per-cow scores as CSV whenever you like? If you change supplier, you want your history.
2. Third-party use. Some agreements let the supplier use anonymised herd data to train models or share it with partners. That’s often fine, but you should know.
3. Termination. What happens to the hardware if you cancel? Some installs leave the camera in place, while subscription-hardware models take it away.

Beef and sheep: where AI health monitoring actually is

For beef and sheep the honest answer is that options are thinner, and mostly built around weight rather than health directly.

Beef. The most practical AI-adjacent tools are weight-tracking platforms. Breedr, a UK app, combines EID reads, weigh-crate data and growth predictions, so a store or finishing animal falling behind its batch’s daily liveweight gain shows up within a couple of weighings. That’s a health signal, even though nothing in it was built to diagnose. Walk-over weighing on water troughs gives you near-daily weights without handling. In housed finishing units, rumen boluses such as smaXtec are sometimes used to catch pneumonia through temperature spikes. Whether that pays depends on cattle value and how bad your BRD history is.

Sheep. Lameness is the big welfare and cost issue in UK flocks, and the Five Point Plan (cull persistent offenders, quarantine, treat promptly, avoid spread, vaccinate) is still the core. Automated lameness detection for sheep is mostly at research stage. Accelerometer ear tags and camera systems on handling races have been trialled at UK universities and agri-tech centres, but nothing yet matches the dairy camera market in commercial maturity. What works now is less glamorous: EID tag readers (Shearwell, Tru-Test, Allflex) linked to flock software, so every footrot treatment is recorded against the ewe and the cull list builds itself. A ewe treated three times in a season is your persistent offender. Software that tells you that, without anyone flicking through a notebook, is worth more than any experimental camera today.

Questions to put to any supplier before you sign

Print these and take them to the stand at the Dairy Show or UK Dairy Day:

• What’s the agreement with trained human scorers, as weighted kappa or as sensitivity and specificity for score 2+, and on how many commercial UK herds?
• How does it identify cows, and what’s the mismatch rate on a herd like mine (breed, colour, parlour type)?
• What does the race need to look like, and will you survey it before quoting?
• Is processing on-farm or in the cloud? What happens to scoring when the internet drops for a day?
• What does my milk buyer accept, and have you got that in writing from them?
• Can I export every score for every cow, at any time, as CSV?
• What’s the minimum contract term, and who owns the camera if I leave?
• Can I speak to two farms who’ve had it running for more than a year, one of them not chosen by you?

That last question tells you more than the brochure will. A farmer two years in can tell you whether the alert list still gets opened every Monday, and that’s the part that pays for the camera.