Video Analytics for Animal Behaviour and Welfare: Areas of Interest

Overview

This page outlines our areas of interest in applying computer vision to the monitoring of animal behaviour, welfare and condition. The common premise is that many early indicators of stress, discomfort or disease are behavioural: changes in activity, posture, feeding, spatial distribution or gait often precede clinical signs.

Continuous visual observation by staff is costly and limited in duration. Camera-based monitoring, by contrast, can in principle run around the clock and over the whole production or life cycle. It also produces quantitative time series that can be compared with an individual or group baseline.

The directions described below cover poultry, horses, zoo animals and pigs. They differ in maturity, from conceptual and exploratory work to field-tested prototypes; the descriptions focus on the problems and methods rather than on specific implementations.

The main research questions across these directions are:

• which behavioural patterns carry diagnostic or welfare-relevant information for a given species and housing system;
• how reliably those patterns can be detected from video under real farm or enclosure conditions;
• how detections can be aggregated into indicators that are interpretable for farm managers, veterinarians, animal keepers and researchers.

Poultry

In broiler production, mortality from microclimate failures, equipment faults and handling errors can reach several percent of the flock, according to industry estimates. Flock behaviour is a sensitive and early indicator of such problems. Two application contexts are of interest: commercial poultry houses (floor and cage housing) and experimental vivaria, where the effects of feeds, additives or vaccines are studied.

Detection of individual broiler chicks in an experimental pen (overhead wide-angle camera)

Behavioural and welfare indicators under study:

• Stress responses. Sudden collective changes in behaviour in response to noise, staff entry or other disturbances. These can be accumulated into a stress history over the growing cycle and, in principle, correlated with human-presence detection.
• Activity level. Tracking positions and trajectories to estimate mean flock activity and to identify the most restless individuals.
• Spatial distribution. Uniformity of the birds' distribution across the house. Huddling or uneven distribution may indicate a drop in temperature or other microclimate issues; both short-term deviations and long-term trends are relevant.
• Feeder and drinker occupancy. Number of birds around feeders over time, used to identify periods of highest and lowest feeding activity.
• Immobile and dead birds. Detection of birds that remain still for an extended time (for example, more than three hours), with their approximate location in the house.
• Aggressive interactions within the flock.
• Body weight. Contactless estimation of live weight from video, avoiding manual weighing and the associated stress.

Commercial poultry house: feeders (blue), immobile birds (red) and interacting birds (green)

Wide view of a commercial poultry house: feeder lines (blue circles), individual birds (green), immobile birds (red and orange markers)

Dense-flock detection in a commercial poultry house

In a research vivarium, the same indicators may support the assessment of new feeds, vitamin supplements and vaccines. Continuous recording over the whole life cycle can capture behavioural changes that fall outside the scope of periodic human observation.

Contactless broiler weight monitoring: Poultry Guard (working prototype)

Of the indicators listed above, body weight estimation has reached the stage of a working prototype. Poultry Guard estimates the average weight of a broiler flock continuously from routine video, without catching or weighing the birds.

• Video sampling. Short video samples are recorded manually or on a schedule (time window and number of samples).
• Estimation. Birds are detected in each sample, and a statistical average flock weight is estimated.
• Growth dynamics. Estimates are tracked by day of the cycle against a target growth curve.
• Calibration. The camera pose is calibrated geometrically. Periodic reference weighing on physical scales provides farm-specific calibration.
• Deployment. The system can run locally or remotely; edge operation on a Raspberry Pi 5 single-board computer is supported.

Prototype interface: detected birds, top-view positions, growth curve against the target, measurement table and camera controls

The prototype's accuracy under real operating conditions is the next question to be answered. The intended validation design covers one poultry house or experimental unit over a growth cycle. It includes camera set-up and geometric calibration, automatic video measurements, periodic reference weighing, and a final comparison of video-based estimates with the reference data.

Horses

In equine management, two conditions are of particular concern: colic and lameness. Both are treated more successfully when recognised early, yet their initial signs are subtle and may go unnoticed between routine checks. Three related monitoring concepts are of interest.

Stall monitoring and daily routine

A camera installed in the stall can provide a structured record of the horse's day, presented to the owner or stable manager through a user application:

• Feeding: start time, duration and completeness of each meal.
• Body position: posture and total time spent lying down.
• Human presence: time and nature of staff visits to the stall.
• Emergency situations: abrupt behavioural changes such as fright or aggression, with notification.

Early signs of colic

The veterinary literature describes a set of behavioural signs associated with the onset of colic. A system that first learns the individual horse's normal behaviour could flag deviations such as:

• disrupted feeding pattern;
• frequent stretching and looking back at the flank;
• frequent neck stretching and head shaking;
• excessive or unusually prolonged lying;
• kicking at the abdomen with the hind limbs, pawing;
• repeatedly lowering the muzzle into the water trough;
• aimless movement around the stall.

This concept can share hardware with stall monitoring.

Posture and action recognition

Stall monitoring and colic detection both rely on a common layer: classification of the horse's postures and actions from video. Two levels are distinguished:

• Basic postures: lying on the back, lying on the side, sternal recumbency, standing.
• Complex actions: eating or drinking, pawing the ground, stretching the legs backwards, drawing the legs towards the abdomen, looking back.

Several of the complex actions coincide with the behavioural signs of colic listed above, so a general action classifier can serve as the basis for condition-specific detectors.

Gait analysis for lameness

A high-resolution camera records the horse in motion, for example during dressage work. Skeletal keypoints are then estimated and tracked to derive biomechanical parameters of each stride:

• duration of the stance phase;
• duration of the swing phase;
• trajectory of each limb's keypoints.

Asymmetries in these parameters between limbs are potential indicators of lameness.

Zoo animals: large raptors

In zoological collections, video analytics can support keepers' visual supervision of animals, especially outside working hours. One direction of interest concerns open-air enclosures of large raptors, such as Old World vultures and sea eagles. Here the aim is to detect emergencies and early signs of deteriorating condition.

Situations considered for detection:

• Entanglement in the enclosure mesh. A 3D model of the enclosure is reconstructed, birds are detected, and their distance to the mesh is estimated to identify a bird hanging on the netting.
• Fights. Identification of the aggressor and of behavioural patterns characteristic of a fight.
• Foreign objects. Detection of atypical objects in the enclosure, with suppression of repeated false-positive detections.
• Atypical lying postures. Lying on the back, lying with the head on the substrate, or prolonged lying, combined with a timer that measures time spent in that position.

Detection of an aggressive interaction between two birds (orange frame) with pose lines

Bird detection and pose estimation in an outdoor enclosure under winter conditions

Spring conditions: birds detected in rock-wall niches, on perches and on the ground, with pose lines on a walking bird

Summer conditions: detection of birds on perches and ledges

Small-scale detection of birds on a rock wall; the orange frame marks a bird classified as lying

Pigs

In pig production, the tasks of interest are mainly zootechnical:

• automated animal counting;
• breed identification;
• contactless body weight estimation;
• as a further step, detection of behavioural patterns.

Open challenges and limitations

Several methodological issues are common to these directions and remain open questions:

• Domain shift. Models trained under controlled conditions (an experimental vivarium) lose accuracy in commercial houses, where lighting, density and bird size differ. Retraining on field data is therefore required.
• Coverage versus resolution. A single wide-angle camera can cover an industrial poultry house, but the image may lack the detail that individual-level analytics need. Multi-camera set-ups add system complexity.
• Dense groups and growing animals. Detection that is reliable for small chicks becomes harder as birds grow and cluster closely together.
• Scarcity of rare-event data. Video examples of early-stage conditions, such as the onset of colic, are difficult to obtain. Validation requires long observation periods and systematic feedback from owners and veterinarians.
• Precision requirements for biomechanics. Gait analysis depends on high-quality video and very accurate keypoint localisation. The camera must also be placed appropriately for each recording session.
• Interpretability. Raw detections need to be translated into indicators and alerts that practitioners can act on without excessive false alarms.