Artificial vision for real-time identification of goat behavior

THE NEED FOR CONTINUOUS MONITORING OF BEHAVIOR IN GOATS

Behavior is key to assessing the welfare and health of goats, as detecting early changes in their activity patterns allows for anticipating potential health issues.

In this way, it is possible to improve management, efficiency, and profitability of the farm (Matthews et al., 2016). However, manual observation of the herd is a costly, subjective process and not well-suited for large intensive farms.

In recent years, Precision Livestock Farming (PLF) has revolutionized the sector with tools to continuously, automatically, and in real-time monitor and control the herd.

Among PLF technologies, artificial vision represents a promising solution that offers an automatic, economical, and non-invasive method to monitor herd behavior.

Computer vision, a branch of artificial intelligence, seeks to mimic biological vision through the use of algorithms applied to images. This approach has been previously adopted in pig farming and cattle (Nasirahmadi et al., 2017), but its application remains very limited in goat farming, despite its potential benefits.

The experimental activity was carried out at the small ruminants farm of the Universitat Politècnica de València, whose schematic plan is shown in Figure 1.

Pen 1 was selected for video collection, conducted continuously for three weeks, from July 8 to 28, 2024.

During this period, eight Murciano-Granadina goats were selected and housed in this pen.

For video recording, a commercial RGB camera (2.8 mm lens and CMOS sensor) with QHD resolution (2,560 x 1,440 pixels), 20 Hz frame rate, and infrared LED for recording in the absence of visible light was used.

The camera was installed in a zenithal position, centered over the pen, at a height of 5 meters.

In Image 1, an example of a frame recorded by the camera is presented.

From the 500 hours of recordings, a set of 1,247 significant frames was extracted, which were classified into subsets for training (70%), validation (20%), and testing (10%) of the model.

All frames were manually labeled, identifying each goat with a bounding box and classifying its behavior according to the following categorical classes (Figure 2):

Eating: action of inserting the head into the feeders.

Drinking: action of inserting the head into the water dispensers.

Standing: action of standing on four legs, also includes walking and urinating.

Lying down: action of being on the ground in any position.

The training subset was used to develop a computer vision model based on YOLOvX (Ge et al., 2021), an algorithm consisting of a neural network that finds different applications in computer vision, such as autonomous guidance (Zhang et al., 2025) or wildlife detection (Delplanque et al., 2023).

In Image 2 an example of a frame analyzed by the model is shown, which identifies each goat through a red box and indicates its behavior and the reliability level of the detection.

THE MODEL AND THE EVALUATION OF ITS ROBUSTNESS

After the training phase, the model was evaluated through the comparison of its predictions with the labeled data, which represent the ground truth.

For a prediction to be considered correct, two conditions must be met simultaneously:

1 The goat indicating box, defined by the model, must overlap at least 50% with the ground truth box. This measure, known as intersection over union (IoU), is illustrated in Figure 3.

2 The model must identify the same goat behavior as the labeling, which represents the ground truth.

As shown by the metrics in the Table 1, the model is fast and accurate, since it processes more than 50 frames per second (allowing real-time application) with an average AP of 96%.

The model achieves AP equal to or greater than 95% in all behaviors, with the only exception being drinking (AP = 94%).

The lower accuracy of the model in detecting goats drinking is due to several reasons:

• There is a data imbalance, with fewer frames of goats drinking compared to other classes.
• The fisheye distortion significantly affects the corners of the frames, where one of the drinkers is located.
• There is a great visual similarity between goats standing and drinking.

As shown in Image 3, where the areas of the frame that the model focuses on are highlighted, it relies on contextual elements to identify behaviors, such as the fence to identify goats eating or the legs to detect goats lying down.

The only element that allows distinguishing between standing and drinking goats is the relative position between the head and the waterer.

NEXT DEVELOPMENTS

The developed model allows reliable, continuous, and real-time monitoring of the group behavior of goats, facilitating the acquisition of different indicators, such as behavioral patterns, heat maps of pen usage, and herd activity history.

Currently, work is underway to achieve more detailed monitoring through the integration of tracking algorithms to follow each goat and obtain detailed information at the individual level.

Acknowledgments

The research team would like to thank technicians José Vicente Martí Vicent and José Luis Palomares Carrasco for their contribution to the farm activities. This work was funded by the First Research Projects Grant (PAID-06-23), Vice-Rectorate for Research of the Universitat Politècnica de València (UPV) and part of the research team benefited from the TED2021-130759B-C31 project grant, funded by MCIN/ AEI/10.13039/501100011033/ and by the “European Union NextGenerationEU/PRTR”.