Characterising responses in group-housed pigs to Salmonella typhimurium infection through integrated computer vision–based behavioural monitoring and statistical analyses
Eddiemar B. Lagua, Hong-Seok Mun, Keiven Mark B. Ampode, Md Sharifuzzaman, Md Kamrul Hasan, Young-Hwa Kim, Chul-Ju Yang
TL;DR
This study uses computer vision to monitor pig behavior and detect Salmonella infection, showing it can track health changes over time but may miss early signs.
Contribution
Integrates computer vision and statistical analysis to detect Salmonella-induced behavioral changes in group-housed pigs.
Findings
TRT pigs showed significant health deterioration 4 days post-inoculation, confirmed by growth decline and elevated health indicators.
Automated behavior monitoring revealed subtle behavioral anomalies as early as 0 days post-inoculation using time-specific Z-score analysis.
Compensatory growth was observed in TRT pigs during the recovery period after antibiotic treatment.
Abstract
Health monitoring is crucial for early disease detection and prompt intervention to mitigate the disease. Computer vision is one of the novel methods for disease detection, but a significant gap remains in its application for detecting behavioural deviations associated with disease. This study employed YOLOv8s-based behavioural monitoring combined with statistical analysis to evaluate disease detection efficacy in group-housed pigs. Two groups of pigs (Control [CON] and Treatment [TRT]), 9–10 weeks old of a (Large White × Landrace) × Duroc cross, were raised for 21 days. The growing period was divided into three periods (adaptation, challenge, and recovery) and evaluated based on growth performance, health indicators (ear base temperature and faecal score), and behaviour (postures, feeding, and drinking). The TRT group was challenged with Salmonella typhimurium during the challenge…
Genes, proteins, chemicals, diseases, species, mutations and cell lines named across the full text — each resolved to its canonical identifier and authoritative record.
Click any figure to enlarge with its caption.
Figure 10
Figure 11
Figure 12
Figure 13
Figure 1
Figure 2
Figure 3
Figure 4
Figure 5
Figure 6
Figure 7
Figure 8
Figure 9Peer Reviews
No public reviews on file for this paper yet. If you reviewed it on a platform where reviews are public (OpenReview, ICLR, NeurIPS, ICML), you can paste yours below so the community can read it here.
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsAnimal Behavior and Welfare Studies · Animal Nutrition and Physiology · Salmonella and Campylobacter epidemiology
Introduction
Health monitoring is one of the critical components of livestock management for the early detection of discomfort and disease. In commercial livestock production, health status is commonly assessed through a combination of clinical signs and behavioural observations, which provide essential information for disease control and management decisions [1–3]. Behavioural indicators are particularly valuable because subtle deviations may come before clinical symptoms become evident [4–7]. Nutritive activities and postural behaviors are particularly informative, as they are closely associated with an animal’s physiological state and its response to environmental and health challenges. For instance, decreased intake of feed and water is frequently observed during infectious diseases [8, 9] and heat stress [10–12], reflecting altered metabolic demands and discomfort. Similarly, alterations in postural behavior, such as increased lying time [13], increased sternal lying [14], or decreased standing and exploratory activity [4, 15], have been linked to infections, thermal stress, pain, and fatigue. These behavioral responses thus serve as sensitive, early indicators of compromised health and welfare in pigs.
However, conventional behavioural monitoring relies heavily on human observation, which is skill-dependent, labour-intensive, non-continuous, and limited in scope, often resulting in insufficient observation time per animal [16–18]. As a consequence, early symptoms are frequently overlooked, and behavioural anomalies are typically detected only when the disease has progressed significantly, sometimes not until the animal is near death or already deceased [19].
To address these limitations, automated and real-time health monitoring technologies have increasingly been integrated into livestock production systems [20–22]. Among these technologies, wearable sensors are commonly used, particularly in cattle farming, to track individual animals’ behavioural and physiological parameters [23–25]. In pig production, wearable sensors are typically ear tags equipped with radio-frequency identification (RFID) [12, 26, 27], accelerometers [28], and, in research settings, sensors for body temperature, sound, and piezoelectric activity [29, 30]. In commercial applications, RFID systems are most commonly implemented when integrated into automatic feeding systems, enabling tracking of individual feed intake and feeding behavior [31, 32]. When integrated with machine learning approaches, such systems have the potential to detect deviations in feeding patterns associated with disease or stress [33]. However, their application is limited by high installation and maintenance costs, and the RFID-integrated feeding system is limited to feeding behavior. Moreover, wearable sensors are not widely adopted in pig farming because pigs tend to reject foreign objects on their bodies and may chew or damage them, especially in group housing conditions.
Computer vision offers a promising alternative, as it enables the remote evaluation of animal conditions and allows for the simultaneous monitoring of multiple animals in the same environment using a single device [34–38]. Furthermore, there is increasing interest in its applications in pig farming, particularly for the automatic detection and monitoring of animal behaviour using object detection techniques. Computer vision–based systems can capture a wide range of behaviours, including nutritive activities [39–42] and postural patterns [36, 43, 44], without requiring animal-mounted devices. Drinking and feeding can be accurately identified using various object detection models, including convolutional neural networks (CNNs) [41, 42, 45], and advanced approaches can further distinguish between nutritive visits (entering the feeding area with feed consumption) and non-nutritive visits (entering the feeding area without consuming feed) with up to 99.4% accuracy [39]. Postural behaviours can similarly be detected using binary (e.g., lying and not lying) [37] or multi-class classification models (standing, sitting, sternal lying, and lateral lying) [46–48]. Recent work has achieved posture classification accuracies exceeding 99% using improved YOLOv5 architectures, even in pigs housed in group settings [44]. Furthermore, several studies have successfully integrated posture, nutritive activities, and other behaviours into a unified model, demonstrating the feasibility of comprehensive behaviour monitoring using computer vision [49–51].
Despite these technological advancements, most existing studies focus primarily on improving behaviour-detection accuracy. There remains a clear knowledge gap in evaluating whether computer vision–derived behavioural data can reliably capture behavioural deviation responses associated with infectious disease in pigs. Therefore, this study aimed to evaluate the effectiveness of computer vision combined with statistical analyses in detecting disease-related behavioural anomalies in group-housed pigs infected with Salmonella typhimurium, based on their postures, drinking, and feeding patterns.
Materials and methods
Experimental animals and facilities
The methodology of this study was reviewed and approved by the Institutional Animal Care and Use Committee (IACUC) of Sunchon National University (SCNU IACUC-2024–23). A total of 64 apparently healthy growing pigs (32 gilts and 32 barrows) of a (Large White × Landrace) × Duroc cross, with body weights ranging from 25.56 to 34.70 kg and aged 9 to 10 weeks, were used. This study was conducted at the Sunchon National University pig experimental farm during the summer months (July and August 2024) in South Korea.
Two identical climate-controlled houses were utilised, each with one room, equipped with air-conditioning units, as described in the previous studies [15, 52]. An electrical problem with the air-conditioning unit occurred in one of the rooms at the beginning of the study, leading to the decision not to use the air-conditioning units in either room during the entire growing period. Nevertheless, the room temperature was set at 24 °C, but the actual conditions exceeded the set temperature. The environmental conditions were regularly monitored and recorded using an environmental monitoring system (Farm Note, NareTrends Inc., South Korea), and the ventilation system was adjusted to ensure uniformity across all the rooms. Furthermore, each room had four fully plastic slatted pens measuring 2.35 × 2.9 m. Each pen was equipped with an automatic wet-dry feeding system (LFS-120, IONTECH Co., Ltd., South Korea), a water trough, and an automatic weighing scale (not used for body weight measurement) (Figs. 1a and 1b). The feeding system was programmed to dispense feed and water when the feeder trough is empty, and the pig touches the sensor. The pigs were already accustomed to the feeding system, as they had been kept in a room with a similar design to the experimental rooms. Additionally, each room was equipped with an RGB camera (PNO-A6081R, Hanwha Vision Co., Ltd., South Korea) with 2560 × 1920 resolution, installed 2 m from the floor between pens 1 and 2, and angled to capture the animals, feeding system, and water trough in pen 2 (Fig. 1a). The video data were stored and accessed via cloud storage.Fig. 1. Equipment in the experimental rooms and sample images. (a) Sample image of pen number 2 displaying different equipment. (b) Sample image taken by the camera (PNO-A6081R, Hanwha vision Co., Ltd., South Korea). (c) Sample prediction image using posture detection Model (PDM). (d) Sample prediction image using feeding and drinking detection Model (FDDM). 1 = camera; 2 = water trough; 3 = feeding system; 4 = weighing scale; 5 = environmental sensors
Bacterial culture conditions
Salmonella typhimurium strain KCTC 2515 was procured from the Korean Collection for Type Cultures (KCTC) and was used to challenge the growing pigs. The bacterium was initially cultured in nutrient broth (Becton, Dickinson and Company, France) containing 3.0 g/L beef extract and 5.0 g/L peptone for 24 h aerobically at 37 °C in a shaking incubator set at 100 rpm. Then, the bacterium was routinely cultured overnight with a 1:100 dilution in fresh nutrient broth for six consecutive times.
Experimental design and management
The two rooms were designated as experimental groups, a control group (CON), and a treatment group (TRT). 16 gilts and 16 barrows (32 pigs) were housed in each group, with 8 pigs comprising 4 gilts and 4 barrows in each pen, replicated four times. The pigs were weighed and grouped according to their body weight, and they were housed in the experimental rooms one week prior to the commencement of the study. This procedure facilitated acclimatization with their penmates and established social hierarchy. The pigs were raised for three weeks, with an adaptation period in the first week, a challenge period in the second week, and a recovery period in the third week (Fig. 2). The CON group pigs were raised under standard farm practices. During the adaptation period, factors were similar across all groups. However, during the challenge period, all pigs in the TRT group were orally inoculated with 10 mL overnight S. typhimurium KCTC 2515 culture for 6 consecutive days to induce infection. The inoculum was administered by oral drenching using disposable needleless syringes, with pigs gently restrained only when necessary to ensure accurate delivery while minimizing stress. Then, an antibiotic (containing Enrofloxacin 50 mg/mL and Butyl Alcohol 30 mg/mL) was administered against S. typhimurium through submuscular injection in the neck at a dose of 1 mL per 10 kg of body weight on the last day of the challenge period (noon time). During the entire growing period, disinfectant foot baths and alcohol sprays were provided in each room to prevent cross-contamination. All routine farm activities and data collection were carried out from the CON room to the TRT room as additional preventive measures. The pigs were fed a corn-soybean-based commercial diet containing at least 3350 kcal digestible energy and 0.90% digestible lysine. The diet also contained animal fat, molasses, mineral sources (salt, limestone, calcium phosphate, magnesium oxide, and mineral premix), pure amino acids (DL-Methionine, Lysine HCl, L-Tryptophan, and L-Threonine), choline chloride, vitamin premix, phytase, and direct-fed microbial complex (strains not specified). The bins of the automatic feeding system were refilled each afternoon, allowing the pigs ad libitum access to diet and water. Water usage per group, referred to here as water intake, was automatically recorded using the same system for environmental monitoring. The recording was in 5-minute intervals. The data on growth performance parameters were measured and calculated on a weekly basis [52]. Artificial lights were turned on 24 h throughout the growing period, and no environmental enrichment material was provided.Fig. 2. The experimental design was divided into three periods: adaptation, challenge, and recovery, each lasting one week. The treatment group was orally inoculated with 10 mL of an overnight culture of Salmonella typhimurium KCTC 2515 for 6 consecutive days, then treated with an antibiotic containing Enrofloxacin 50 mg/mL and Butyl alcohol 30 mg/mL at a dose of 1 mL per 10 kg of body weight at the end of the challenge period
Faecal sample collection and bacterial analysis
Fresh faecal samples, not touched on the floor, were collected from 3 random pigs per pen at the start of the study and on 6 DPI during the weighing of the pigs. The samples were stored in an ice chest and immediately delivered to the laboratory to determine the presence of Salmonella. The faecal samples per pen were thoroughly mixed, and 1 g of faeces was mixed with 9 mL PBS, vortexed, and 20 µL was collected and plated on BBLTM Salmonella Shigella Agar (Becton, Dickinson and Company, France). The plates were then incubated aerobically for 24 h at 37 °C. Plates positive for Salmonella would contain colonies with black centres. All samples collected at the first start of the growing period were negative for Salmonella.
Clinical signs determination
The four corners of each pen were examined for the presence of new faeces and scored for consistency (0 for normal faeces, 1 for soft faeces, 2 for mild diarrhoea, and 3 for severe diarrhoea) [53] and averaged per pen. The scoring was performed once daily at around 10:00 AM. A study showed that the maximum temperature of the ear base, measured using infrared thermography, was highly correlated with rectal temperature and serves as a reliable indicator of body temperature in pigs [54]. In the present study, a handheld thermal camera (FLIR E76, FLIR Systems Inc., Boston, MA, USA) was used to determine the body temperature of pigs non-invasively, minimising human-caused behavioural responses. The infrared thermographs of the ear base of the pigs were taken from approximately 1 m, and the maximum values were recorded. 3 pigs per pen (12 per group) were randomly selected, marked, and monitored for the ear base temperature. During the adaptation and recovery periods, ear base temperature was recorded daily alongside faecal scoring. During the challenge period, measurements of ear base temperature were taken at 6-hour intervals from the time of inoculation.
Overview of the computer vision models
The images were extracted from video recordings of pig batches housed in the same experimental facilities. A total of 5714 unique images were obtained (Table 1). Of these, 3684 were used to construct training and validation datasets. The remaining 2030 images, collected on different dates and distinct from the first set, were used for testing. These images were saved in separate folders.Table 1. Number of images in the datasetImagesTrainingValidationTestTotalRaw2,5831,1012,0305,714Augmented^1^5,1665,166Total7,74910,880Horizontal flip, rotation between −15° and +15°, grayscale 25%, and noise up to 1.65% of pixels
The collected image sets were uploaded and annotated within a single project in Roboflow. Initially, pigs were categorised into 10 distinct classes based on posture and activity: Lateral Lying, Sternal Lying NFD, Sternal Lying F, Sternal Lying D, Standing NFD, Standing F, Standing D, Sitting NFD, Sitting F, and Sitting D. Here, NFD (not feeding/drinking) denotes pigs that are neither feeding (F) nor drinking (D). Image labelling was conducted by the first five trained authors. All annotators followed a standardized protocol based on the ethogram described in Table 2 and visualized in Fig. 3. Although inter-observer reliability metrics and external validation were not conducted, all labels were reviewed and, where necessary, adjusted by the first author to ensure internal consistency.Table 2. Descriptions of pig postures and nutritive activitiesModels (Classes)DescriptionPosture Detection Model Lateral LyingThe pig is lying on its side with legs extended. Sternal LyingThe pig is lying on its sternum or belly with at least two legs folded under the body. StandingThe body is elevated and supported by four legs. SittingThe posterior part of the body rests on the floor, while the front legs remain upright and extended, elevating the anterior part.Feeding and Drinking Detection Model FeedingThe pig is in various postures, except for ‘Lateral Lying,’ with its head entering the feeder and reaching toward the bottom. DrinkingThe pig is in various postures, except for ‘Lateral Lying,’ with its head entering the water trough and reaching toward the bottom. Not Feeding/DrinkingThe pig is neither feeding nor drinking.Fig. 3. Different classes in the posture detection Model (PDM) and the feeding and drinking detection Model (FDDM). NFD = not feeding/Drinking; F = Feeding; D = drinking
After annotation, the training and validation datasets were split into a training set (70%) and a validation set (30%). To facilitate model development, the annotated dataset folder was duplicated to create two subsets: one for the Posture Detection Model (PDM) and another for the Feeding and Drinking Detection Model (FDDM). For the PDM, the classes were modified and simplified into four categories: Lateral Lying, Sternal Lying, Standing, and Sitting. In contrast, for the FDDM, all classes of various postures except the ‘Lateral Lying’ indicating ‘F’ were merged into a single ‘Feeding’ class, those indicating ‘D’ into ‘Drinking’, and neither of these two was classified as ‘Not Feeding/Drinking’.
Before downloading the datasets from Roboflow, pre-processing and augmentation procedures were conducted. Auto-Orient, static crop (0–100% Horizontal Region, 0–97% Vertical Region), and resizing to 800 × 600 were included in the pre-processing. Furthermore, to diversify the datasets and improve robustness across both experimental rooms and varying video quality, augmentation methods such as horizontal flip, rotation by ±15°, 25% grayscale reduction, and noise up to 1.65% of pixels were applied to the training dataset. This resulted in an additional 5166 images, increasing the dataset to 10,880 images (Table 1).
Both models were trained using a pretrained YOLOv8s implemented in the Ultralytics framework and trained in Google Colaboratory using a Tesla T4 GPU (15 GB VRAM), Python 3.10.12, and PyTorch 2.2.1. Training was conducted for 50 epochs with a batch size of 16, and input images were resized to 640 × 640 pixels. The optimiser used was AdamW, which was automatically selected by YOLOv8’s default configuration, with a learning rate of 0.00125 and a momentum of 0.9. All other hyperparameters were set to their default values. The best models were used. Performance metrics for both models were summarised in Table 3. Figure 1c and d, and Supplementary Videos 1 and 2, provide sample predictions using the models.Table 3. Prediction performance of the models used for behavioural monitoringModels (Classes)MetricsPrecisionRecallF1 ScoremAP50Posture Detection Model Lateral Lying92.492.092.296.3 Sternal Lying91.290.090.694.8 Sitting92.090.491.294.9 Standing97.998.298.099.2 Average93.492.693.096.3Drinking and Feeding Detection Model Feeding96.093.894.996.9 Drinking89.490.690.095.3 Not Feeding/Drinking96.896.796.798.5 Average94.193.793.996.9
Computer environment and behaviour analysis
Video recordings from pen 2 in each experimental room were processed to extract behavioural data using the pretrained YOLOv8s models. The inference was conducted offline due to restricted access to the live CCTV feed managed by a third-party provider. The data extraction was conducted on a desktop computer running Windows 10 Education 22H2, equipped with an Intel(R) Core(TM) i5-9400F CPU at 2.90 GHz, 8 GB of RAM, and an NVIDIA GeForce GTX 1050 Ti GPU with 4 GB of VRAM. The trained models were implemented in Python 3.11.6 and executed in Visual Studio Code (VS Code) version 1.100.2 using PyTorch (v2.1.0) [55] and OpenCV (v4.8.0) [56]. To fasten the inference, video frames with a resolution of 2560 × 1920 were resized to 800 × 800, sampled at 1 frame per second (fps), and with a detection confidence threshold of 0.45. The output data were stored in CSV format, containing group names, dates, timestamps, and per-class counts. Using custom Python scripts, multiple CSV files were consolidated into a single dataset, and behaviour frequencies were aggregated on an hourly basis. The percentage of each behaviour class per hour was then calculated using the pivot table function in Microsoft Excel, applying the following equation:
\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${P_c}\left( \% \right) = \left( {{{{N_c}} \over {{N_{total}}}}} \right) \times 100$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${P_c}\left( \% \right)$$\end{document} is the percentage of the behaviour class in a particular hour, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N_c}$$\end{document} represents the total number of observations for that class within that hour, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N_{total}}$$\end{document} is the total number of observations across all classes in the same hour. Total lying was added by adding lateral lying and sternal lying.
Statistical analyses
Growth performance, behavioural variables, ear base temperatures, and faecal scores were analysed using linear mixed-effects models (LMMs) implemented in SPSS (v20). For growth performance variables, group was included as a fixed effect, with pen specified as a random effect to account for the experimental unit. For behavioural data, ear base temperatures, and faecal scores, which were collected repeatedly over time, group, time-related factors (day and collection points), and their interactions were included as fixed effects, and an autoregressive covariance structure of order 1 [AR(1)] was applied to account for temporal autocorrelation. Model assumptions were assessed using residual diagnostics, including normal Q–Q plots indicating approximate normality and a good fit. Behavioural data did not fully meet normality assumptions and were therefore transformed using a Box–Cox power transformation (λ = 0.1) to improve residual normality. Following the transformation, residuals were approximately normally distributed. All models were fitted using restricted maximum likelihood (REML), and statistical significance was declared at p < 0.05. For behavioural and clinical sign data, estimated marginal means were obtained, and pairwise comparisons were performed using a Bonferroni adjustment to assess group differences at each time point (day) when significant group-by-time interactions were detected.
Pearson’s correlation analyses were conducted to evaluate the relationships between daily faecal scores, ear base temperatures, and pig behaviours within each group. In addition, correlation analyses were performed for each group concerning hourly and daily room temperature, water intake, and pig behaviours. Correlation results were interpreted based on both statistical significance (p < 0.05) and the magnitude of the correlation coefficient. Correlation strength was classified by the absolute value of the coefficient (|r|), with 0.20–0.40 considered weak, 0.40–0.60 moderate, 0.60–0.80 strong, and >0.80 very strong [57]. The sign of the coefficient indicated the direction of the relationship (positive or negative). Independent two-sample tests, linear mixed-effects models, and correlation analyses were performed using SPSS (version 20).
Principal component analysis (PCA) was used as an exploratory multivariate method to visualise overall behavioural patterns and changes over time across experimental periods. The correlation analysis confirmed sufficient inter-relationships among behavioural variables and the suitability of the dataset for PCA. Only the averaged daily behaviour data within each group were analyzed. All behavioural variables were standardised (z-scores) before PCA, and the analysis was performed using the correlation matrix. Two principal components were retained for visualisation based on the proportion of variance explained. PCA scores were plotted with observations coloured by treatment group and distinguished by phase-specific markers, enabling visual assessment of multivariate behavioural trajectories over time. Loading vectors were overlaid to indicate the relative contribution and direction of individual behaviors to each principal component.
Additionally, a multiclass XGBoost classifier integrated with SHAP (Shapley Additive exPlanations) values was employed to identify the key behaviours that distinguish different periods of the TRT group. SHAP and feature importance score plots were created and exported. These analyses were performed in Python using Visual Studio Code, with the packages pandas (v2.3.1) [58], Numpy (v2.2.6) [59], scikit-learn (v1.7.1) [60], matplotlib (v3.10.5) [61], seaborn (v0.13.2) [62], XGboost (v3.0.3) [63], and SHAP (v0.48.0) [64].
The temporal behaviours of each group were analysed for anomalies per group using Z-score analysis, with the adaptation period serving as the baseline. Two approaches to Z-score calculation were implemented for comparison: First, daily Z-score analysis (DZA) was computed across the full 24-hour period to allow a more holistic view of daily behavioural deviation. Second, time-segmented Z-score analysis (TSZA) was conducted by dividing each day into two time windows: 00:00–11:00 and 12:00–23:00. This segmentation accounted for diurnal behavioural variation, as pig activity levels often differ between morning and afternoon hours. This method aimed to assess the potential of time-segmented Z-scores in enabling earlier or clearer detection of anomalies compared to daily-level calculations. The Z-score was computed using the following equation:
\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Z - score = {{{P_{d,c,t}} - {\mu _{adapt,c,t}}} \over {{\sigma _{adapt,c,t}}}}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${P_{d,c,t}} $$\end{document} is the percentage of behaviour class c on day d during time window t (or entire day for DZA), and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mu _{adapt,c,t}}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\sigma _{adapt,c}} $$\end{document} are the mean and standard deviation of behaviour class c during the adaptation period for the same time window t (or entire day for DZA), respectively. A 95% confidence interval (CI) threshold of ±1.96 was used to determine behavioural anomalies.
Results
Growth performance
All growth performance parameters did not show significant differences between groups during the adaptation period (Fig. 4). However, during the challenge period, pigs in the TRT group exhibited a significant and drastic reduction in growth, with a difference of −90.51% (3.16 vs. 0.30, p = 0.001) in body weight gain compared to the CON group. Although the feed intake was 1.07 kg/head, the FCR was very high (25.00) because 43.75% (14/32) (data not shown) of the pigs in the TRT group experienced weight loss. Additionally, although the water intake in both groups was almost similar, the water:feed ratio was higher by 43.57% (6.61 vs. 9.49) compared to the CON group. This indicated that the inoculation of S. typhimurium in pigs was successful despite the diet containing a complex of probiotics. Moreover, compensatory growth was observed in the TRT group during the recovery period, exhibiting a significant improvement in body weight gain (4.10 vs. 5.38 kg, p = 0.006), average daily gain (0.58 vs. 0.78 kg/day, p = 0.006), and FCR (2.90 vs. 2.22, p = 0.037). Additionally, the feed intake (1.66 vs. 1.71 kg, p = 0.337) was higher but not statistically significant. Overall, for the 3-week growing period, the TRT group had significantly lower final body weight (42.61 vs. 40.54 kg, p = 0.004), body weight gain (12.91 vs. 10.91 kg, p = 0.026), ADG (0.61 vs. 0.52, p = 0.026), feed intake (1.59 vs. 1.38 kg, p = 0.012), and higher FCR (2.59 vs. 2.73, p = 0.025).Fig. 4. Comparison of weekly growth performance: (a) body weight, (b) body weight gain, (c) average daily gain, (d) feed intake, (e) feed conversion ratio, (f) water intake, and (g) water:feed ratio. ^1^Collected by group, no statistical test applied. Means with different superscripts were significantly different (p < 0.05)
Health conditions
Generally, the ear base temperatures measured were lower than the rectal temperature of the healthy pigs, according to the initial assessments and the study of Xie et al. [54]. Additionally, the thermal readings were influenced by the ambient temperature, resulting in daily fluctuations. Nevertheless, room temperatures remained similar in both groups. Significant differences in ear base temperature were observed between the groups. The TRT group exhibited significantly higher ear base temperatures 1 DPI, with differences noted at 12 and 18 HPI (Fig. 5a). There were no differences afterwards, but significant differences re-emerged from 3 DPI (60 and 78 HPI). Highly significant differences (p < 0.01) were noted from 4 to 6 DPI. No significant differences were observed at 7 DPI and throughout the recovery period. Furthermore, scouring was first observed at 1 DPI, coinciding with the increase in ear base temperature, resulting in a marked increase in the faecal score from nearly 0 to 1 in the TRT group (Fig. 5b). It continued to rise to 2.69 until 8 DPI, then suddenly dropped to 0.94 at 9 DPI and gradually normalised in the succeeding days. The faecal bacterial test also confirmed that faecal samples from the TRT group, collected at 6 DPI, were 100% positive for Salmonella; however, no further tests were conducted to confirm the strain in the faecal samples. Nevertheless, no samples from the CON group were positive, indicating the effectiveness of preventive measures.Fig. 5. Comparison of clinical signs. (a) Daily and 6-hour intervals ear base temperature measured using FLIR E76 (FLIR systems Inc., Boston, MA, USA). (b) Daily faecal score (0 = normal faeces, 1 = soft faeces, 2 = mild diarrhoea, and 3 = severe diarrhoea). DPI = days post-inoculation; HPI = hours post-inoculation. Asterisks indicate statistically significant group differences at specific days for behaviours with a significant group × day interaction (*p < 0.05; **p < 0.01)
Pigs’ behavioural response
Effects of room temperature
The diurnal pattern of behaviour in pigs of both groups was not clearly defined until −4 DPI, displaying multiple peaks in lying, drinking, and feeding frequencies as shown in Figs. 6 and 7. Room temperature was not the primary factor investigated in this study, but it influenced the behaviour of the pigs. Starting at −3 DPI, the room temperature exceeded 30 °C from 11:00 to 19:00, and overall room temperatures went above the set temperature (24 °C), with a high total lying observed during this period. Additionally, high total lying was also observed approaching midnight and early morning, when room temperatures dropped. Room temperature and total lying time were weakly and non-significantly associated (Fig. 8). However, these periods can be distinguished by the lying postures and drinking frequencies, where sternal lying and drinking were high during periods of high room temperatures, and high lateral lying and low drinking were observed during the coldest periods. The differences in postures may conflict with existing studies; however, pigs mostly slept during the coldest period (but remained above the set room temperature), and pigs drank from the trough while lying in a sternal position during high temperatures. The correlation analysis between hourly room temperature and pig behaviours in CON (Fig. 8a) showed that sternal lying and drinking were moderately correlated (R = 0.429), and drinking was strongly correlated (R = 0.626) with water intake. However, the degree of correlation between these variables varied on a daily basis. Furthermore, the degree of association between room temperature and daily behavioral patterns differed between groups.Fig. 6. Time course of defined behaviors and room temperature in the Control group. A high-resolution version of this figure is provided as supplementary Figure 1. DPI = days post-inoculationFig. 7Time course of defined behaviors and room temperature in the treatment group. The red arrow is highlighting the days when feeding and drinking gradually decreased and then increased again. A high-resolution version of this figure is provided as supplementary Figure 2. DPI = days post-inoculationFig. 8Pearson’s correlation analyses between pigs’ behaviours and house temperature on daily and hourly bases: (a) Control group and (b) treatment group. Correlation coefficient (|r|) with 0.20–0.40 considered weak, 0.40–0.60 moderate, 0.60–0.80 strong, and >0.80 very strong. Asterisks (*) indicate significant correlation values. * = P < 0.05; ** = P < 0.01
Group comparison
Figure 9 illustrates the daily behavioural dynamics of pigs in the CON and TRT groups, while Table 3 summarises the overall effects derived from linear mixed-effects models. During the adaptation period (−7 to − 1 DPI), no clear behavioural differences were observed between groups. Following inoculation, transient day-specific behavioural changes were evident. Drinking behaviour differed between groups at 0 DPI, with higher values in CON. A pronounced behavioural disruption occurred at 4 DPI in the TRT group, characterised by increased total lying and sternal lying, reductions in standing, sitting, feeding, and drinking, and increased NFD. These changes correspond to significant day effects.Fig. 9. Effects of Salmonella typhimurium inoculation on daily behavioural patterns of pigs: (a) lateral lying, (b) sternal lying, (c) total lying, (d) standing, (e) sitting, (f) feeding, (g) drinking, and (h) not feeding/drinking. DPI = days post-inoculation. Asterisks indicate statistically significant group differences at specific days for behaviours with a significant group × day interaction (p < 0.05; p < 0.01). Daggers (†) denote exploratory day-specific group differences (p < 0.05) for behaviours without a significant group × day interaction and should be interpreted cautiously Table 4. Average behaviours (%) of control and treatment group pigs for a 21-day growing periodBehavioursControlTreatmentP-value*^1^GroupDayGroup × DayTotal Lying84.6884.290.823<0.0010.003Lateral Lying60.3456.660.0470.3100.965Sternal Lying24.3427.630.0030.3840.805Standing11.4111.940.4920.0080.333Sitting3.913.780.431<0.0010.148Feeding5.074.710.699<0.0010.061Drinking4.214.640.155<0.001<0.001Not Feeding/Drinking90.7290.650.6720.0030.161P-value from transformed data using a Box–Cox power transformation (λ=0.1)
Total lying did not differ overall between groups (p = 0.823) but varied significantly across days and showed a significant group × day interaction (p = 0.003), indicating distinct temporal responses between CON and TRT. In contrast, lying postures exhibited consistent group-level differences, with no significant interactions. Sternal lying was higher in the TRT group than in CON (27.63 vs. 24.34%, p = 0.003), whereas lateral lying was lower in TRT (56.66 vs. 60.34%, p = 0.047). Correlation analysis further supported the behavioural relevance of lying posture changes. In the TRT group, sternal lying showed a moderate positive correlation with faecal score (R = 0.529), whereas lateral lying exhibited a weak and non-significant negative association (R = −0.103) (Fig. 10b). Although Fig. 9a and b show day-specific fluctuations in these postures, particularly during the challenge period, these should be interpreted as descriptive patterns rather than confirmatory effects, given the absence of significant group × day interactions.Fig. 10. Pearson’s correlation analysis between pigs’ behaviours, faecal score, and ear base temperature of Control (a) and treatment (b). Asterisks (*) indicate significant correlation values. Correlation coefficient (|r|) with 0.20–0.40 considered weak, 0.40–0.60 moderate, 0.60–0.80 strong, and >0.80 very strong. * = P < 0.05; ** = P < 0.01
Furthermore, standing, sitting, and feeding behaviours were not affected by group (p > 0.05) but exhibited significant day effects (p ≤ 0.008). Drinking behaviour displayed a significant day effect and a strong group × day interaction (p < 0.001), despite no overall group difference, which is reflected in Fig. 9gby marked deviations during the challenge and recovery phases, including higher drinking activity in TRT at 11–12 DPI. NFD similarly did not differ between groups but varied significantly across days (p = 0.003), peaking during the acute phase of infection.
The PCA plot (Fig. 11) highlights the behavioural variation between groups across the different periods. The TRT group displayed greater variability, as shown by the wider dispersion of its data points, suggesting unstable behavior patterns. Overlaps between adaptation, challenge, and recovery periods indicate transitional responses likely associated with infection stages. An XGBoost classifier was applied to extract feature importance scores and SHAP values, allowing the identification of key behaviours distinguishing the different periods in the TRT group (Fig. 12). The results indicated that high levels of sternal lying and lateral lying were the main discriminating behaviours during the challenge period (Figs. 12b and e), supporting the correlation analysis results. In contrast, sitting emerged as the most influential behavior in both the adaptation (Fig. 12a and d) and recovery periods (Figs. 12c and f), with elevated sitting particularly associated with the recovery period.Fig. 11. Principal component analysis (PCA) based on daily average behaviors. Cyan markers indicate Control, and pink markers show treatment. Sphere markers represent the adaptation period, square markers represent the challenge period, and diamond markers represent the recovery period. Only the pigs in treatment were challenged with Salmonella typhimurium and treated with antibioticsFig. 12SHAP values (a-c) and feature importance scores (d-f) across growing periods of the treatment group. A and d represent the adaptation period; b and e represent the challenge period; c and f represent the recovery period. In the SHAP plot, the colour gradient (red = high feature values, blue = low feature values) shows the magnitude of features (behaviours); positive values indicate that a feature drives the prediction toward the target class (period), while negative values push it away
Detection of behavioural anomalies
The CON group exhibited more stable behaviour patterns than the TRT group in both DZA and TSZA, as shown in Fig. 13. Diurnal behavioural anomalies were detected in TSZA but were missed using DZA, such as feeding at 2 DPI during the 0:00–11:00 period in the TRT group. Additionally, a significant increase in lateral lying was detected from 12:00–23:00 at 9 DPI in the CON group, but it was only detected at 11 DPI using DZA. Consequently, the subsequent result comparisons utilised the TSZA findings.Fig. 13. Behavioural anomalies detection using Z-score analysis based on the mean and standard deviation during the adaptation period (−7 to − 1 DPI). (a) Control group daily Z-score analysis. (b) Treatment group daily Z-score analysis. (c) Control group time-segmented Z-score analysis. (d) Treatment group time-segmented Z-score analysis. DPI = days post-inoculation; CI = confidence interval. Z-scores exceeding the threshold at a 95% confidence interval (±1.96) were considered significant or anomalous, indicated with asterisk marks (*)
Table 5 summarises the counts of behavioral anomalies detected. In the CON group, 55 anomalies were identified using TSZA, with 20 during the challenged period and 35 during the recovery period. Among the behaviours, sitting and drinking during 0:00–11:00 were significantly increased, which was observed starting in the late adaptation period (Fig. 13c). Furthermore, lateral lying in the 12:00–23:00 period was significantly increased for 5 days during the recovery period, while sternal lying decreased significantly for 7 days. On the other hand, the behavior patterns in the TRT group were significantly disturbed during the challenge and recovery periods, with 43 and 68 (total 111) anomalies detected during those periods. These disruptions reflect acute behavioural instability in response to the infection and gradually stabilised toward the end of the recovery period. Major changes in behaviour occurred at 4 DPI, as also observed in the group comparison (Fig. 9). However, the earliest significant change in behaviour post-inoculation was noted on 0 DPI during the 12:00–23:00 period, marked by an increase in sitting posture. This was followed by a significant increase in drinking during the 0:00–11:00 period and lateral lying during 12:00–23:00, with a concurrent significant decrease in sternal lying at 2 DPI, which continued until 3 DPI. Total lying and standing increased and decreased, respectively, in the 12:00–23:00 period on 3 DPI. Feeding behaviour started to decline at 2 DPI and decreased significantly at 4 to 5 DPI in both time windows, but continued until 7 DPI in 12:00–23:00. Sternal lying started to increase at 4 DPI, peaking at 8 DPI, and decreased significantly from 11 DPI. Lateral lying and total lying both peaked at 4 DPI and gradually reduced toward the end of the growing period, but remained at significantly higher levels compared to the adaptation period. Lastly, pigs began to recover at 6 DPI, exhibiting a gradual increase in nutritive-related behaviours, particularly in the morning. A significant increase in feeding was observed within 24 hr (0:00–11:00 at 7 DPI) after antibiotic administration and peaked at 9 DPI.Table 5. Counts of significant behavioural deviations based on Z-score analysesGrowing PeriodTime WindowChangeControlTreatmentLSTStSiDFNLSTStSiDFNDaily Z-score AnalysisAdaptation (−7 to −1)+1–1Challenge (0 to 6)+3121222–12222Recovery (7 to 14)+271261–112Time-Segmented Z-score AnalysisAdaptation (−7 to −1)0:00–11:00+11111–111112:00–23:00+11–111Challenge (0 to 6)0:00–11:00+76212131–1132112112:00–23:00+116412–4243Recovery (7 to 14)0:00–11:00+18837262–112541712:00–23:00+5761–72764L = Lateral Lying; S = Sternal Lying; T = Total Lying; St = Standing; Si = Sitting; D = Drinking; F = Feeding; N = Not Feeding/Drinking. A positive (+) change shows significant behavioral deviations above the positive threshold at 95% confidence (1.96), while a negative (–) change indicates deviations exceeding the negative threshold
Discussions
Two models based on the YOLOv8s architecture were trained and applied to quantify behaviour in experimentally infected group-housed pigs with S. typhimurium, focusing on postures, drinking, and feeding patterns. Although the raw images were relatively small and the dataset was imbalanced, the models achieved excellent performance, with mAP50 scores reaching 96.30% for the PDM and 96.90% for the FDDM. Notably, the highest accuracy was observed for classes Standing (99.2%) and Not Feeding/Drinking (98.5%). These values exceed the accuracy reported in similar studies [45, 65], demonstrating the reliability of the models in quantifying behaviour. This improvement in performance could be due to data augmentation and the nature of the data collected from identical rooms. While not applied in real-time, the output enabled precise, high-throughput behavioural measurements that were subsequently analysed through statistical methods. Although future architectural refinements could push performance even higher [39, 44], the current models provided a robust foundation for detecting significant behavioural deviations and anomalies between infected and uninfected pig groups.
Pigs orally infected with S. typhimurium exhibited classical clinical signs such as lethargy, inappetence, and decreased drinking during the challenge period, leading to reduced growth performance, findings consistent with those of Ahmed et al. [4]. These observations confirm that inoculation successfully induced a sickness response in the pigs, while behaviour monitoring captured measurable behavioural deviations during this period. However, unlike the earlier study that relied on manual and not continuous observations (4 hr/day, weekly averages), the AI-based models facilitated automated and continuous behaviour monitoring, enabling a detailed temporal analysis of pig responses to infection, therapeutic interventions, and environmental fluctuations (Figs. 6 and 7). Additionally, the combined statistical analyses enabled the identification of days showing significant differences in behaviours, particularly at 4 DPI. This capability demonstrates a strong potential for real-time application in disease and welfare surveillance and early response systems.
In terms of specific behaviours, lying posture was particularly revealing. While the total lying was similar between groups, sternal lying was significantly higher in the infected group and increased during the challenge period. Correlation analysis further revealed a strong association between sternal lying and faecal score, suggesting that sternal lying is an important indicator of enteric infection, as confirmed by the SHAP analysis. Previous studies have also linked high sternal lying to inflammatory or immune challenges (e.g., LPS, Poly I:C, PRRSV) [66, 67], and to thermoneutral [15] and cold-stress conditions [58], compared with heat-stress conditions. However, lying posture alone may not provide reliable predictions, as increased sternal lying was also observed during periods of high room temperature in the present study, contradicting the findings of Mun et al. [15]. This discrepancy likely reflects behavioural adaptation, as pigs often adopt a sternal posture while drinking during heat exposure. Consequently, incorporating nutritive activities alongside postural behaviours may improve discrimination between physiological stressors. Notably, the relationship between room temperature and daily behaviours differed between groups, potentially due to infection-related effects. These distinct behavioural responses could be leveraged to differentiate infectious challenges from heat-stress environments. Such conditions may be automatically distinguished using machine learning approaches that exploit differences in temporal behavioural patterns or activity, as demonstrated by Yin et al. [28]. Such advancement could improve precision in health monitoring.
In field conditions, there is no CON group as a basis for comparison to detect behavioural deviations. Therefore, Z-score analysis was employed, a commonly used method for detecting anomalies in time-series data [31, 68, 69], and using the adaptation period of each group as a baseline. This method is computationally simple yet effective, as demonstrated in the present findings. Furthermore, the results showed that behavioural anomaly detection was more sensitive when data were analysed in finer time windows, such as 12-hour intervals, compared to single daily summaries. This approach enhanced the detection of subtle and transient behaviour deviations that may occur before clear clinical signs emerge. Sensitivity could be further improved by dividing the day into shorter intervals, such as 6-hour windows. Targeting specific periods of known activity, such as feeding times, may also improve anomaly detection accuracy, particularly in systems with controlled, scheduled feeding. A drop in activity during these periods could serve as an early indicator of illness or environmental stress [70].
Although the primary objective was to detect behavioural anomalies associated with infection, house temperature also appeared to influence pig behaviour, particularly their diurnal activity patterns. During the adaptation period, behavioural patterns differed from those in the other periods. This was supported by Z-score analysis in the control group (Figs. 13a and 13c), which revealed increased drinking, sitting, and lateral lying behaviours corresponding with rising temperatures toward the end of the growing period. While Z-score analysis proved effective for detecting anomalies, it does not account for underlying causal factors, limiting its interpretability. Consequently, it was difficult to determine whether the significant behavioural deviations observed between 0 and 3 DPI in the treatment group were attributable to infection or environmental influences. Machine learning techniques were initially explored to detect behavioural anomalies based on clinical signs. However, due to the short duration of the experiment, limited dataset size, and the likely influence of multiple interacting factors, the models’ performance was poor.
We observed notable limitations in the system’s sensitivity for detecting early signs of infection, particularly in group-based comparisons and ZDA. Although clinical indicators such as elevated ear base temperature and increased faecal scores were evident at 1 DPI, in agreement with findings in piglets [71, 72], corresponding behavioural deviations were not statistically significant until 4 DPI. This delay may be attributed to the nature of group-based monitoring, which can mask early deviations, especially in larger groups where a single infected pig has a minimal impact on the overall average behaviour. Pig behaviour is influenced by social dynamics and group size [70], and it could be hypothesised that the sensitivity of group-level monitoring decreases as the population size increases. This limitation is especially relevant under field conditions, where pigs are often chronically exposed to low doses of pathogens in a non-uniform manner [73], and pigs respond at different rates depending on their immune status, genetic background, or prior exposures [74–77], further complicating group-level detection. In such scenarios, individual-based monitoring offers a more accurate and sensitive approach for detecting early health deviations and improving precision in health management. This can be accomplished in computer vision by integrating object detection with tracking models, an area of growing interest in precision pig farming. Although current tracking accuracy remains a challenge, especially in large groups, ongoing research is steadily improving performance [78–80]. Nevertheless, group-based monitoring can be a cost-effective strategy for the surveillance of highly contagious and lethal diseases, such as African swine fever, where rapid detection is crucial for making timely decisions for disease containment [19].
In conclusion, this study highlights the benefits of AI-based computer vision models for automated behaviour monitoring in pigs. Continuous monitoring combined with statistical analyses enabled the identification of behavioural deviations related to health and welfare conditions. However, further research is required to fully leverage longitudinal data and to robustly classify behaviours as normal or abnormal across different production stages. This may be achieved by integrating additional sensors, such as automated thermal cameras, for continuous detection of clinical signs, thereby enhancing the identification of infection-related behavioural deviations. While challenges remain, particularly in early detection and group-level sensitivity, this approach offers a promising foundation for advancing precision livestock health monitoring and improving animal welfare through timely disease detection.
Conclusions
This study demonstrated that computer vision–based behavioural monitoring using pretrained YOLOv8 models, when integrated with statistical analyses, can effectively detect disease-related behavioural deviations in group-housed pigs infected with Salmonella Typhimurium. Deviations in postures, feeding, and drinking behaviours were successfully captured and analysed, highlighting the potential of AI-driven, non-invasive monitoring systems for early disease detection and real-time health surveillance in precision livestock farming.
However, several limitations should be addressed in future studies, including:
- Suboptimal camera placement due to limited ceiling height and partial occlusion by housing equipment;
- Behavioural monitoring restricted to a single pen, with limited individual-level clinical assessment;
- Feeding and drinking detection based on posture and head position rather than confirmed intake;
- Potential influence of feeder and drinker design on observed behaviours, limiting generalisation;
- Continuous 24-hour lighting was applied, which may affect behavioural and physiological responses to infection by disrupting the circadian rhythm; and
- Use of a fixed adaptation period as a behavioural baseline, which may not fully account for age-related behavioural deviations.
Despite these limitations, the findings support the applicability of computer vision and behavioural analytics as scalable tools for early disease detection and data-driven decision-making in modern pig production systems, thereby supporting improved animal welfare and health.
Electronic supplementary material
Below is the link to the electronic supplementary material.
Supplementary Material 1
Supplementary Material 2
Supplementary Material 3
Supplementary Material 4
The reference list from the paper itself. Each links out to its DOI / PubMed record.
- 1Putz AM, Harding JCS, Dyck MK, Fortin F, Plastow GS, Dekkers JCM, et al. Novel resilience phenotypes using feed intake data from a natural disease challenge Model in wean-to-finish pigs. Front Genet [Internet]. Frontiers. 2019. 10.3389/fgene.2018.00660. cited 2025 Dec 15];9.10.3389/fgene.2018.00660 PMC 633168930671080 · doi ↗ · pubmed ↗
- 2Bonneau M, Poullet N, Beramice D, Dantec L, Canario L, Gourdine J-L. Behavior comparison during chronic heat stress in Large White and creole pigs using image-analysis. Frontiers in animal science [Internet]. 2021 [cited 2023 Sept 25];2. https://www.frontiersin.org/articles/10.3389/fanim.2021.784376. Accessed 25 Sept 2023.
- 3Lagua EB, Mun H-S, Ampode KMB, Chem V, Kim Y-H, Yang C-J. In: Artificial intelligence for automatic monitoring of respiratory health conditions in smart swine farming. Animals. Vol. 13. Multidisciplinary Digital Publishing Institute; 2023. p. 1860. 10.3390/ani 13111860.10.3390/ani 13111860 PMC 1025186437889795 · doi ↗ · pubmed ↗
- 4Yu Z, Han Y, Cha L, Chen S, Wang Z, Zhang Y. Design of an intelligent wearable device for real-time cattle health monitoring. In: Front robot AI [Internet]. Frontiers; 2024. cited 2025 July 4];11. 10.3389/frobt.2024.1441960.10.3389/frobt.2024.1441960 PMC 1161736639640911 · doi ↗ · pubmed ↗
- 5Gómez Y, Stygar AH, Boumans IJMM, Bokkers EAM, Pedersen LJ, Niemi JK, et al. A Systematic Review on validated precision livestock farming technologies for pig production and its potential to assess animal welfare. In: Front Vet Sci [Internet]. Frontiers; 2021. cited 2025 Dec 16];8. 10.3389/fvets.2021.660565.10.3389/fvets.2021.660565 PMC 816024034055949 · doi ↗ · pubmed ↗
- 6Mora M, Piles M, David I, Rosa GJM. Integrating computer vision algorithms and RFID system for identification and tracking of group-housed animals: an example with pigs. J Anim Sci. 2024;102:skae 174. 10.1093/jas/skae 174.10.1093/jas/skae 174PMC 1124569138908015 · doi ↗ · pubmed ↗
- 7Funk TH, Rohrer GA, Brown-Brandl TM, Keel BN. Online feeding behavior monitoring of individual group-housed grow-finish pigs using a low-frequency RFID electronic feeding system. Transl Anim Sci. 2024;8:txae 051. 10.1093/tas/txae 051.10.1093/tas/txae 051PMC 1114349138827162 · doi ↗ · pubmed ↗
- 8Huang L, Xu L, Wang Y, Peng Y, Zou Z, Huang P. Efficient detection method of pig-posture behavior based on multiple attention mechanism. Comput Intel Neurosci. 2022, 2022;1759542. 10.1155/2022/1759542.10.1155/2022/1759542 PMC 930852235880056 · doi ↗ · pubmed ↗