How Lameness Detection Cameras Score Your Cows

Most herds get mobility scored four times a year. A RoMS-registered scorer stands at the parlour exit, watches each cow walk past, marks her 0, 1, 2 or 3 on the AHDB scale and sends you a spreadsheet a few days later. A lameness detection camera over the dairy exit does the same job at every milking. That means two or three scores a day for every cow, and they reach your phone before breakfast.

This post covers how the camera turns a walking cow into a number, how far that number can be trusted against a trained human, and what changes in the yard once you have one. For the wider set of sensors (collars, boluses, rumination tags) see our guide to AI livestock health monitoring. This piece sticks to feet.

What the camera actually sees

The hardware is plain. Most systems use a single camera, often an ordinary 2D security camera, mounted roughly 3 to 4 metres above the exit race and looking straight down or slightly angled. CattleEye, the Belfast-built system, works with standard cameras of this sort. HerdVision, developed by Agsenze with Kingshay, uses a 3D depth camera and also reports body condition from the same pass.

The software does the clever part, in four steps:

1. Finding the cow. Each video frame is checked for animals, and every cow gets a box drawn round her and is followed as she walks through.
2. Placing landmarks. The model marks points on her body: head, withers, the line of the spine, hooks and pins. This is the same “pose estimation” technology that lets a phone app track a person’s skeleton.
3. Measuring the walk. Over the frames where she is fully visible, the software measures how flat or arched her back is, how much her head bobs, how regular her strides are, and her speed compared with the cows around her.
4. Turning measurements into a score. The measurements go into a model trained on thousands of clips that human scorers have already labelled. Out comes a score on the familiar 0–3 scale, or a 0–100 index that maps onto it.

Identity comes either from the RFID reader at the exit or from the parlour sort order. CattleEye can also recognise individual Holsteins from their black-and-white back markings. That doesn’t work on a herd of solid-coloured Jerseys, so check how your system plans to tell cows apart before you sign anything.

The second step is where most of the work goes on. Arched back and nodding head are the cues a human scorer uses, and they are what the camera is measuring too. It simply measures them the same way on every pass and never gets tired at cow number 240.

How accurate is it compared with a person?

People tend to assume mobility scoring by a human is the gold standard. It isn’t really. When two trained scorers watch the same cows, they give the exact same 0–3 score somewhere around 70–80% of the time in most studies. Agreement on the simpler question “is this cow lame (2 or 3) or not?” is better, typically in the high 80s. Almost all the disagreement falls between neighbouring scores, 1 against 2 especially.

In published comparisons, camera systems land in the same band. A cow that a trained scorer calls 2 gets a 2 from the camera about as often as a second scorer would give her one. The usual statistic in these papers is the kappa value. Camera-to-human figures come out similar to human-to-human ones, moderate to good, and nowhere near perfect.

What sets the camera apart is how often it scores and how it treats the trend:

One misjudged score from a human sits on your records for three months. A single odd camera reading, say because the cow got shoved by the one behind, gets swamped by the next twenty. The alert you act on is the rolling trend, not a single walk.

It goes wrong in predictable places. Cows bunching at the exit hide each other. A cow that trots through will look sounder than she is. Wet, slippery concrete makes the whole herd walk carefully, which can push scores up across the board, and dim winter lighting or a low sun across the race at evening milking can both upset 2D systems. When a supplier claims a high accuracy figure, ask what proportion of passes are actually scored. If a third of passes get thrown out as “not visible”, a 90% accuracy figure on the rest tells you much less.

A worked example: 220 cows, Somerset

Take a 220-cow Holstein herd milking twice a day through a 20/40 herringbone. The last quarterly RoMS score found 66 cows at score 2 or 3. That’s 30%, which is close to the UK average from surveys.

AHDB puts the average cost of a lameness case at roughly £330, counting lost yield, fertility and extra culling. That’s £21,780 a year carried by those 66 cows, before counting the ones that went lame and recovered between visits without anyone noticing.

The farm puts in a camera at the exit. What the herdsman sees at 5.30am looks something like this (layout adapted from a typical daily alert list):

MOBILITY ALERTS – 14 Oct, AM milking
Passes scored: 208 / 219 (95%)

NEW ALERTS (moved to score 2, 3-day trend)
Cow  1184   Score 2   Trend 1 → 2 over 4 days   Last trim: 112 days ago
Cow  0937   Score 2   Trend 1 → 2 over 3 days   Calved: 21 days ago
Cow  1402   Score 2   Trend 0 → 2 over 2 days   Check: sudden onset

PERSISTENT (score 2+ for 14+ days)
Cow  0861   Score 3   Treated 03 Oct, no improvement
Cow  1290   Score 2   Treated 06 Oct, improving (2.6 → 2.2)

NOT SCORED THIS MILKING: 11 cows (bunching x7, ID missing x4)

Of those, 1402 is the one to deal with today. A cow that goes from sound to impaired in two days usually has a stone, a foreign body or early digital dermatitis, and those are quick, cheap fixes if caught now. Cows 1184 and 0937 belong on Thursday’s trimming list. Then 0861 needs a vet or the foot trimmer to look again properly, because whatever was done on the 3rd hasn’t worked.

Suppose earlier treatment brings the herd from 30% down to 20% prevalence within a year. That’s 22 fewer lame cows, which at £330 each saves around £7,260. If the subscription for 220 cows runs at £1,000–£1,500 a month (ask for a firm quote, because pricing varies by supplier and herd size), that saving alone doesn’t fully pay for it. What swings the sums is the labour saved on handling, catching severe cases before they become culls, and the use some milk buyers will accept the camera records for. Do the sums with your own figures, not the brochure’s.

What changes in the daily routine

The obvious change is that lame cows get picked up sooner. The less obvious one is that someone has to own the list. A camera that produces an alert every day that nobody reads is just an expensive CCTV system. On farms where it works, one person checks the list after morning milking and drafts cows for the crush that same day or within the week. Every treatment then goes in as a note so the camera’s trend can show whether it worked.

Footbathing and trimming can now be planned around actual numbers. If the herd’s average score creeps up a week after you switch footbath solution, or after cows go into a cubicle shed with new mats, you’ll see it as a line on a chart. Before, you’d have found out at the next quarterly visit.

Milk buyer requirements are worth raising with your field officer before you buy. Red Tractor and most milk contracts still expect mobility scoring by a trained person, and not every buyer accepts camera data instead. Some farms keep the RoMS scorer coming in and treat the camera as the day-to-day early warning in between visits. That’s a sensible way to begin. After two or three visits you can lay your scorer’s sheet next to the camera’s figures for the same week and see how closely they match on your own cows.

Fitting is usually done in a day: a camera on a beam over the exit, a small PC or network box in the dairy office, and a broadband or 4G link. Before the engineer leaves, look at the race. A single-file exit with a straight run of at least five or six cow lengths after the last gate gives the software clean footage. If cows come out of a rapid-exit parlour in a crowd and turn sharply straight away, expect a poor scoring rate unless you add a gate or a short race.

The cows walk out the same way they always have, and nothing about milking changes for them. The difference is that the cow at the back of the queue, quietly shifting her weight off a sore hind foot, now turns up on Tuesday’s list rather than in a scorer’s report in January.