Computer vision for animal welfare on farms: what it can (and can’t) do
4 June 2025
Animal scientist and big data research associate Claudia Kamphuis graduated from Wageningen University in 2004, going on to complete her PhD on her research into the use of sensor data and machine learning for the automatic detection of mastitis (udder infection) in dairy cows. After 20 years, how does she view the use of computer vision, big data and AI in farm animal husbandry and breeding? Has any progress been made? What can vision do for animal welfare and what can’t it do?
3 Key applications of vision technology in modern animal husbandry
I feel that the use and potential of vision technologies in keeping and particularly managing farm animals can be divided into three main themes. Firstly, you can use vision – cameras and sensors – to detect, identify and track animals. Either individual animals or groups of animals, indoors and outdoors, on land and in water, and take management actions accordingly. For example, if you detect poultry huddling together, that means the animals are cold. I think vision has a lot of potential to track animals individually through time and space. However, the holy grail for vision is the identification (not detection) of animals. Once we can do that, vision will really take off.
The second aspect is determining behaviours. On the one hand, behaviours you can determine better with vision than with existing techniques, such as walking behaviour and social interaction. Take a dairy cow spending a lot of (too much) time at the feeder or lying in a stall for too long, for example. Vision adds a component that is not possible with radio frequency identification (RFID) and pedometers. An additional advantage of vision is that the technique is non-invasive and independent of battery life. Another example is demonstrated by my colleague Jan Erik Doornweerd, who uses vision to assess the movement of broilers.
For me, the third theme is using vision to make processes more efficient or to automate them. For example, estimating the weight of fish with vision is now faster, more accurate and more animal-friendly than the traditional weighing of (individual) fish. Vision can also help determine whether animals have one chromosome too few or too many. Determining that is crucial for breeding purposes because animals with chromosomal abnormalities may be infertile or diseased. You don’t want those in your breeding population.
24/7 Animal welfare: the power of always-on vision technology
The main advantage of vision technology is that, in most cases, it can continuously detect and measure and thus contribute to the continuous monitoring of behaviours considered important for animal-friendly husbandry. Human observations often provide snapshots and other sensors deliver derived information which is not always an adequate basis from which livestock farmers or researchers can draw a correct conclusion.
The main difference is the fact that I think, expect, that we will be able to use image processing to measure much more accurately, get much closer than with current sensors. We can do things better. Not recognising heat status by increased activity, but by seeing which cow is jumping and which cow stays standing. Not recognising lameness in a cow by reduced activity, but because we can see that the back is curving, that the head is bobbing, that her pattern of getting up is different. Not measuring heat stress in poultry by a single THI value, but also because we can see chickens lying stretched out. That, I think, is where the added value of vision lies. I feel that detection with vision technology is a much more reliable option than with other sensors, but it still needs to be proven.
Challenges of vision systems in livestock farming: what’s (still) holding us back?
Looking back over the past 20 years, I must honestly say that I am also slightly critical of the advance of vision, sensor data and machine learning. I took part in my first conference on precision livestock farming in 2007 and some aspects haven’t changed or improved significantly since then. For example, sensors for detecting mastitis are still not 100 percent reliable and there are still too many false-positive results, where there is ultimately nothing wrong with the animal. And the economic aspect is often still ignored. Expensive technology is being developed but what does it actually deliver? Those same challenges exist for computer vision where we have an even bigger challenge to ensure that image processing works in the challenging practical conditions. And having to deal with naturally curious animals – especially pigs – who destroy sensors.
However, I do believe and I’m convinced that vision is going to help people and society get a better and more objective image of a good life for farm animals and manage them accordingly. Particularly thanks to the continuous and non-invasive nature of vision technologies. Wouldn’t it be great if, as a (dairy) farmer, you could be alerted as soon as your grazing animals are suffering heat stress? Or when a drone alerts you as a sheep farmer when a wolf or other predator is threatening the flock? Perhaps that drone could then simultaneously chase that wolf away with a sound or light signal.
