How computer vision can help create a more animal-centred poultry sector
4 March 2025
Jan Erik Doornweerd, a PhD candidate at Wageningen University & Research, investigated how computer vision can help in assessing movement among broiler chickens. This research could lead to future improvements in the health, welfare and performance of broilers.
Having graduated as an animal scientist from Wageningen University & Research (Breeding and Genetics) in 2019, Jan Erik Doornweerd began his PhD research in 2020 on ‘Digital phenotyping of individual broilers in group-housing: detection, tracking and identification’. His PhD research was commissioned by Wageningen Livestock Research (WR)/Breeding and Genetics using data from Cobb (the global broiler breeding organisation). “As far as I’m concerned this was the perfect fit for me, because it brings together my interests in both poultry and computer vision!”
From subjective to objective scoring
Doornweerd explains how phenotyping – the measurement and analysis of plant and animal traits – works with broilers. “The phenotyping of movement and gait behaviour among broilers for breeding purposes is something that has always been carried out by trained experts. It involves getting the animal to walk through a laterally demarcated walkway of a predetermined length. This gait analysis is performed on a substrate identical to the substrate in the poultry shed and the expert looks at the animal’s movement from the rear. The things they look at include the broiler’s ankles and knees, how high the animal lifts its legs (step height) and the curvature of the toes. These observations are used to produce what’s known as a from 0 to 5 for the broiler’s walking ability. 0 is the highest score that can be achieved, and 5 is the lowest. A score ≤2 is considered optimal and >2 suboptimal. Besides the gait score, the expert also provides a score related to other foot conditions that might affect the animal. This score is a measure of any welfare problems. As you can imagine, this phenotyping is very intensive work and that makes it difficult for someone to perform it objectively and uniformly over a long period of time. Unconsciously, you can’t help relating the scoring of a particular broiler to the chicks you’ve already seen and that affects the gait scores. Moreover, it’s a snapshot and it doesn’t reveal the underlying cause of a certain assessment. And yet these gait and welfare scores are the measures used for the selection and genetic improvement of a population!”
The aim of Doornweerd’s PhD research is therefore to objectify and better quantify movement assessments. “Automating these assessments not only makes them more objective, but also means you can score more animals in less time, which is crucial for breeding management and the genetic development of broilers. This isn’t just a positive outcome for breeders – animals benefit just as much through improvements to their health and welfare, thereby improving the dignity of animal life in the poultry sector.”
RFID has pros and cons
Jan Erik Doornweerd’s PhD research builds on research by his colleague Malou van der Sluis. “Malou van der Sluis investigated the extent to which passive radio frequency identification (RFID) can detect and record the movement of broilers. In that approach, each chick has a unique RFID tag that identifies the animal within a 1.80-metre by 2.61-metre testing environment divided into compartments measuring 36 cm by 42/45 cm. The bottom of each compartment used for the tests has a centrally located antenna to pick up the signal from the RFID tags. Doornweerd found that the movements were accurately recorded in 95.3 per cent of cases. In the remaining instances, the RFID tag skipped one or more compartments because the antennas didn’t register the tag. This could be because an animal crossed the boundary between two compartments or walked through a portion of two compartments. So in 4.7 per cent of cases, the actual movement and the distance actually covered deviated from the recorded values. In most cases, this means an underestimate of the distance travelled. A positive aspect of this method of measurement is that you can identify individual animals, which is necessary in terms of breeding, but at the same time it is impractical to fully equip a poultry shed with antennas.”
Security cameras offer potential
In his PhD research, Doornweerd looked at whether and how a camera with computer vision might be able to detect the movements of broilers, and whether combining the camera with algorithms would enable it to identify individual animals. A literature review revealed that RGB security cameras have been used successfully to detect motion in other animal species. “Partly for this reason, we positioned a standard RGB security camera above the aforementioned testing environment. We use the well-known YOLO algorithm to detect the animals (and the other objects). YOLO stands for ‘You Only Look Once’ and it’s something that a few researchers came up with in 2015. To track animal movements, we use the Simple Online and Realtime Tracking (SORT) algorithm.”
“My PhD research shows that you can record animal movements very effectively with a camera and, in most cases, more accurately than with RFID tags. It’s particularly useful for movements around corners and for walking pace. But because cameras film from above, they’re less successful at tracking individual animals. As soon as one or more broilers lie or stand too close together, you can get ‘identity switches’. The algorithm is then no longer able to distinguish individual animals. This way of working also doesn’t allow us to identify an animal – we can only distinguish animals from each other and follow them within the space.” Follow-up research by Doornweerd has shown that it is also possible to track individual animals in group housing by combining RFID and computer vision.
Using existing data
As well as recording animal movements horizontally across the testing environment, Doornweerd also investigated what role computer vision could play in supporting experts in the gait scoring mentioned above. “For that, the algorithm needs to be able to distinguish the animal’s head, neck, hips, knees and ankles. We trained the algorithm for this by manually annotating images of moving broilers ourselves, and we used existing footage of moving turkeys to train the algorithm. It turns out that the latter is perfectly possible. Initially, computer vision can support the expert in doing gait scoring but I wouldn’t rule out further or full automation eventually.”
Follow-up steps
“For any follow-up steps, we first need to scale up the research, both in terms of technology and in terms of the design and scope. So far, we’ve looked into whether you can use computer vision to record walking movements. That does work, but we cannot yet draw any conclusions about the quality of the measurement. Identity determination is necessary if we want to be able to estimate genetic parameters for the purpose of genetic improvement in animal health and animal welfare, and also because our research focuses on breeding. When you’re looking at farm management in the poultry sector, this may not be the case. In that setting, the data is relevant and interesting as an early indication of problems with feed, water, disease or smothering. And also as a way of analysing the behaviour of broiler chickens and their use of space with regards to barn design to make their housing more animal-oriented. What type of housing is most comfortable for a broiler? In addition, we could also use the same vision technology and algorithms to analyse movement and gait behaviour in other types of poultry such as ducks, turkeys and even other species.”
In terms of (complementary) technology, Doornweerd and his fellow researchers are particularly interested in alternatives to RFID (including non-invasive options), low-cost vision technology and connectivity solutions for the large-scale use of camera technology on poultry farms. “This will allow us to scale up the application of the technology and further improve behavioural research. Personally, I would really like to analyse social interaction between animals. Because knowing what the animal itself wants is a way for us to improve the dignity of animal life in the poultry sector.”
