Camera-Based Sow (Sus scrofa domesticus Erxleben) Posture Analysis and Prediction of Artificial Insemination Timing

Abstract

Determining sow (Sus scrofa domesticus Erxleben) estrus status requires considerable labor investment, and continuous real-time monitoring is impractical. Workers typically identify estrus at scheduled intervals and determine artificial insemination timing based on experience. However, experience-based methods are subjective, vary with operator expertise, and impede standardized management in large-scale farms. This study employs cameras and deep learning to detect sows and analyze postural changes, enabling estrus detection and optimal insemination timing prediction. Experimental results indicate that the proposed method achieved an accuracy of 70% (42/60), where the recommended insemination timing differed by less than 24 h from human decisions. This approach facilitates data-driven estrus detection and insemination scheduling, potentially reducing labor intensity and improving reproductive outcomes, particularly beneficial for labor-intensive and large-scale swine production systems.

1. Introduction

Artificial insemination in sow (Sus scrofa domesticus Erxleben) reproduction systems has the potential to enhance productivity, fertility, and labor efficiency [1]. Successful artificial insemination must be performed within 24 h of ovulation. Ovulation occurs after approximately two-thirds of the estrus period and lasts for an average of 40–60 h [2]. Estimating the optimal ovulation and insemination timing requires first determining whether the sow is in estrus by conducting a backpressure test (BPT) or observing changes in the vulva. Sows in estrus exhibit a fixed posture during the BPT, with their hind legs and back muscles becoming rigid [3] and their vulva becoming swollen and reddened [4].

Determining the estrus status of sows is laborious. As continuous real-time monitoring is impossible, workers can identify estrus at specific times during their working hours. For sows identified as being in estrus, workers perform artificial insemination after a certain period based on their personal experience. However, this approach is subjective and varies according to the worker’s expertise, and makes the uniform management of large-scale farms challenging.

Therefore, precision livestock farming has attracted increasing attention. Automatic monitoring of sow activity can support data-driven decision-making, improving production efficiency [5]. Sows exhibit distinct behavioral characteristics, such as increased activity, reduced feed intake, erected ears, and frequent visits to boars, during estrus [6,7]. Unlike previous sensor-based approaches that require physical attachment to animals, our vision-based system provides non-invasive, continuous monitoring capability suitable for large-scale commercial operations. To observe these behavioral traits, sensors can be attached to the sow’s body to effectively measure activity levels. However, body-attached sensors require a sufficiently long battery life and must be periodically replaced [2,8].

In our previous study [9], sows were detected, and their postures were estimated using camera footage and an object detection model. The detected sow posture changes were accumulated as time series, and a recurrent neural network (RNN) was used to classify the estrus status. The ground truth was defined as the time window of 24 h preceding the worker’s estrus assessment. Despite using only postural changes, the model achieved an accuracy of 86% in estrus classification. These results are consistent with those of a previous study [10], which reported that sows tend to stand or sit for longer durations during estrus.

However, the timing of a worker’s estrus assessment is subjective and varies depending on individual experience and skill level. Moreover, estrus detection was only conducted at specific times (08:00 and 16:00) during working hours. Consequently, estrus prediction accuracy in the previous study was evaluated using a dataset labeled according to worker assessments, introducing uncertainty regarding whether the detected state accurately reflected the sow’s actual estrus status. Furthermore, the previous model treated estrus detection as a binary classification task, thus failing to account for the diverse behavioral trajectories exhibited during estrus. This oversimplification limits the model’s ability to determine the optimal timing for artificial insemination, which requires recognizing subtle temporal patterns and individual variability in estrus behavior.

Similarly to the previous study [9], we observe changes in sow posture using cameras and an object detection model. After constructing a time series of sow postures, the estrus status is determined, and the optimal timing for artificial insemination is predicted. However, an RNN is not used in the study. The proposed model is compared with the actual insemination timing selected by the workers, and the number of piglets born is analyzed to evaluate its effectiveness.

2. Related Work

2.1. Sow Behavior and Estrus

During estrus, estrogen levels in sows increase rapidly, leading to heightened activity and reduced rest time. The rise in estrogen increases cortisol and physiological changes, such as elevated heart rate, respiratory rate, and blood pressure, which peaks just before estrus [2,11,12]. The estrus period in sows lasts for an average of 52.6 h, with ovulation occurring approximately 44 h after the onset of estrus [13]. Sows inseminated 0–24 h before ovulation exhibit high conception rates [14].

To prevent sows from fighting as they approach the estrus cycle and to facilitate estrus detection, they are confined to individual stalls and exposed to boars. Sows housed in individual stalls exhibit an increased duration of active postures during estrus, spending less time lying down or resting [15].

Several sensor-based approaches have been explored for sow estrus detection. For example, study [16] attached a digital accelerometer (STMicroelectronics-LIS3L02DS (STMicroelectronics, Geneva, Switzerland)) with a battery to the sow’s neck to classify activity types such as feeding, walking, and resting with high accuracy, showing potential for early detection of estrus signs or health problems. However, these devices are prone to damage due to exploratory behaviors and become increasingly burdensome to maintain in large herds because of sensor management and battery replacement. In addition, study [17] employed a LiDAR-based 3D camera system positioned 0.7–1.0 m behind the sow to capture point-cloud data and extract morphological features of the vulva, including length, width, and volume. The estimated volume showed a strong correlation with ground-truth measurements ($R^2=0.92$), enabling precise monitoring of estrus-related changes. Nevertheless, LiDAR systems are expensive and difficult to install and maintain, which limits their widespread use in commercial farms.

2.2. Object Detection Model

Object detection involves detecting instances of visual objects of a certain class (such as humans, animals, or cars) in digital images [18]. With the rapid advancements in deep learning neural networks and graphics processing units (GPUs), object detection systems have exhibited significantly improved performance [19].

Deep learning-based object detection models can be broadly categorized into two approaches: two-stage and single-stage. Faster R-CNN, which first extracts regions of interest (RoIs) and then classifies each RoI, is an example of a two-stage approach. In contrast, the single-stage approach detects objects in a single pass through a neural network, with You Only Look Once (YOLO) and single-shot multibox detector being common examples [20].

YOLOv5 is a relatively recent version of the YOLO series, offering a lightweight architecture, improved training speed, and a more flexible deployment environment than previous models [21]. YOLOv5 incorporates scalable model sizes based on a cross-stage partial structure and an efficient data-augmentation technique called mosaic augmentation, ensuring both high accuracy and resource efficiency [22]. Additionally, YOLOv5 performs well even in environments with limited hardware resources and is effectively utilized in research and applications where real-time processing is not strictly required.

Most deep learning-based object detection models can be used for sow detection and posture estimation. We selected YOLOv5 because of speed, accuracy, and model size.

3. Proposed Method

The study in this manuscript only included non-invasive observation on animals through ceiling-mounted cameras. No physical contact, handling, or intervention was involved. According to Article 3 of the Korean Laboratory Animal Act, ethical approval from an IACUC is not legally required for this study. In addition, continuous video monitoring is considered minimally intrusive, as ceiling-mounted cameras do not interfere with the animals’ daily activities or cause additional stress.

Figure 1 shows the proposed method for predicting the optimal timing of artificial insemination. First, the system receives video input, detects the sows, and estimates their postures. Using the bounding box (Bbox) coordinates of the detected sows, each sow is assigned an identification (ID) based on its location. When a sow’s posture changes, both the posture and the duration for which it is maintained are stored as time series data ($TSD$). These data are sampled at fixed intervals (t) to construct a regularly spaced time series dataset (TSD).

A double moving average is then applied to the generated $TSD$ to produce the ${TSD}_{DMA}$. If the most recent values in the generated $TSD$ (${TSD}_{DMA}$) show a continuously increasing trend, the system determines that the sow is in estrus.

3.1. Data Collection and Preprocessing

We detected sows in video footage recorded by a ceiling-mounted rear-angle camera, following the approach used by [9].

Sow postures were categorized as either not-standing or standing (Figure 2). Objects were distinguished based on the center coordinates of the detected bounding box, and whenever a sow’s posture changed, both the posture and the duration for which it was maintained were recorded.

This event-based sequence of posture changes constituted the time series data ($TSD$), which was inherently unevenly spaced because posture transitions occurred at irregular intervals (Figure 3). To facilitate quantitative analysis, the $TSD$ was resampled at fixed 1 min intervals, thereby generating a regularly spaced time series (${TSD}_r$).

A double moving average was applied to the ${TSD}_r$, as shown in Equation (1), to generate a new $TSD$, ${TSD}_{DMA}$. ${\beta }_1$ and ${\beta }_2$ represent the window sizes for the moving averages, set to 1440 and 720, respectively. These values correspond to 24 h and 12 h windows, chosen to smooth daily cyclical noise and capture the underlying multi-hour activity trends effectively. $P$ denotes a single value in the $TSD$. By applying a double moving average, the noise was reduced, and the overall trend of the data was extracted.

$${TSD}_{DMA}={\displaystyle \frac{1}{{\beta }_2}}{\sum }_{j=0}^{{\beta }_2-1}\left({\displaystyle \frac{1}{{\beta }_1}}{\sum }_{i=0}^{{\beta }_1-1}{P}_{t-j-i}\right)$$

(1)

3.2. Estrus Detection and Optimal Timing Prediction for Artificial Insemination

When the length of the $TSD$ with the applied double moving average (${TSD}_{DMA}$) exceeds a certain threshold ($\gamma$), the average value of the most recent ${\displaystyle \frac{\gamma }{2}}$ period is compared with the average value of the preceding ${\displaystyle \frac{\gamma }{2}}$ period. If the recent average is higher, a binary array is generated with a value of 1; otherwise, it is set to zero. Thus, a value of 1 in the binary array indicates that activity has increased in the most recent ${\displaystyle \frac{\gamma }{2}}$ period compared to the preceding ${\displaystyle \frac{\gamma }{2}}$ period. If the activity continues to increase, the binary array takes the form [1, 1, 1, 1, ..., 1].

If the number of consecutive 1 s in the binary array, starting from the most recent time point, exceeds ${\theta }_1$, the sow is classified as being in estrus. For example, if the binary array is [0, 0, 1, 1, 1], the number of consecutive 1 s at the most recent time step is 3. After determining estrus, the optimal timing for artificial insemination is predicted as ${\theta }_2$ hours after the estrus detection point. The values of ${\theta }_1$ and ${\theta }_2$ are determined based on the hypothesis discussed in Section 3.3.

3.3. Problem Definition and Hypothesis

This study aims to identify sow estrus and predict the optimal insemination timing based on posture-derived activity trends

As shown in Figure 4, sows tend to seek boars more frequently during estrus [23]. The estrus period ($E$) lasts 40–60 h [2].

$$40h\le E\le 60 h$$

(2)

Ovulation ($O$) occurs after two-thirds of the estrus period has passed [2].

$$O=\lambda +\left({\displaystyle \frac{2\times (\lambda +E)}{3}}\right)$$

(3)

The optimal timing for artificial insemination ($I$) is within 24 h of ovulation ($O$).

$$\mathit{max}\left(O-24 h, \lambda \right)<I<O$$

(4)

Workers determine sow estrus status using the BPT and predict $I$ based on the onset of estrus (first detected time, $\lambda$). To predict $I$, we estimate the onset of estrus ($\widehat{\lambda }$) by observing changes in sow posture and use it to determine $I$.

To estimate $\widehat{\lambda }$ and predict $I$, the following hypotheses are established, and the values of ${\theta }_1$ and ${\theta }_2$ are set accordingly: 1. Sows sleep for approximately 8 h. 2. There is a warm-up period of approximately 24 h before $\lambda$, during which activity levels increase.

Based on this hypothesis, the activity levels were expected to increase for approximately 16.7 h, excluding sleep time. Therefore, ${\theta }_1$ is set to 16.7 h. If activity levels continuously increase for 16.7 h, the sow is classified as being in estrus ($\widehat{\lambda }$). The optimal timing for artificial insemination ($I$) is then predicted 20 h after $\widehat{\lambda }$, excluding the 8 h sleep period (${\theta }_2=8+20h$).

If $E$ lasts 60 h, $I$ falls within the range of $\lambda +16.7 h$ to $\lambda +40 h$. If $E$ lasts 40 h, $I$ falls within the range of $\lambda +2.7 h$ to $\lambda +26.7 h$. Setting $I$ at $\lambda +20h$ satisfies Equations (2)–(4), as well as Hypotheses 1 and 2. Thus, ${\theta }_1$ is set to 1000, and ${\theta }_2$ is set to 28 h. Since our data is sampled at 1 min intervals, this corresponds to 960 data points. We set ${\theta }_1$ to 1000, slightly longer than 960, to conservatively ensure a sustained increase in activity is detected. Therefore, if activity levels continuously increase for approximately 16.7 h (1000 min), the sow is classified as being in estrus ($\widehat{\lambda }$).

While these biological principles provide a theoretical range for insemination timing, we adopted a data-driven approach to determine the final, precise values for our model’s key parameters, ${\theta }_1$ and ${\theta }_2$. We analyzed the activity patterns from a preliminary dataset of successfully inseminated sows (those categorized in the high-similarity groups $L_1$ and $L_2$, in our results). This empirical analysis revealed that a sustained increase in activity, lasting approximately 16–17 h, consistently preceded the optimal insemination window determined by experienced farm workers.

Based on this data-driven evidence, we set ${\theta }_1$ to 1000 min (~16.7 h) to conservatively capture this sustained activity. Furthermore, the analysis showed that the average time from the start of this detected activity increase to the actual insemination event was 28 h, leading us to set ${\theta }_2$ to 28 h. This dual approach ensures our model is not only biologically plausible but also empirically validated against the specific behavioral patterns observed on the farm, strengthening the foundation of our prediction method.

Workers determine sow estrus status using the BPT and predict insemination timing based on the first detected estrus onset (λ). In this study, λ was defined according to the farm’s routine workflow, where sows were exposed to a boar twice daily (08:00 and 16:00) for a limited duration. Our objective is to demonstrate that automated posture-based monitoring could provide insemination timing results comparable to those determined by skilled workers in real farm settings. Therefore, we assessed our model against the insemination timing determined by skilled workers using established farm practices, focusing on sows that eventually farrowed.

4. Experimental Results

This section compares the recommended optimal timing for artificial insemination using the proposed method with the actual insemination timing selected by the workers.

4.1. Experimental Environment Setup

Our experiments were conducted using cameras with an ingress protection rating of 66 or higher. Each camera was positioned considering the field of view, installation height, and sow-detection accuracy, and the setup was designed to monitor up to four sows simultaneously. The farm where the experiment was conducted had a concrete floor equipped with ventilation fans for air circulation. The sows were housed individually in stalls, and the observed sows were F1 crossbreeds of Landrace and Yorkshire breeds. The sows were fed a diet of corn, soybean meal, and wheat, with additional ingredients, including vitamins, amino acids, sweeteners, flavor enhancers, and dried distillers grains. Water was available via automatic drinkers. The sows were introduced to the sow barn on Thursdays during lunchtime and transferred to the farrowing barn on weekends during lunchtime.

To check for estrus, workers exposed the sows to a boar twice daily at 08:00 and 16:00 for no more than 15 min, from Sunday morning to Wednesday noon. During exposure, workers observed the vulva and conducted a BPT by applying pressure on the sow’s back. Estrus sows typically showed muscle rigidity, with stiffened legs and back when mounted, and mucus discharge around the vulva was also assessed. Among the sows expected to be in estrus, gilts (sows that have never given birth) underwent immediate artificial insemination. In contrast, parous sows (sows with previous birthing experience) underwent insemination after a certain period, as determined by the worker’s empirical judgment.

4.2. Sow Detection

In the experiment, the YOLOv5s model was trained using approximately 20,000 datasets on a GeForce RTX 2080 Ti GPU. The training and test data were collected in the same environment with the camera fixed, achieving a detection accuracy of 98.6% based on a mean average precision of 0.5. In the graphs (Figure 5), the x-axis represents time in 1 min intervals, and the green line, aligned with the left y-axis, indicates estrus when it reaches 1.0. The optimal timing for artificial insemination is recommended 28 h (${\theta }_2$) after the detected estrus point. Additionally, the blue line, aligned with the right y-axis, represents the posture-based activity of the sow, calculated from posture transitions and smoothed using a double moving average. The blue curve therefore reflects the overall activity level of the sow, whereas the green curve denotes whether the activity level is increasing compared with the previous state. A rise in the green curve indicates increased activity relative to earlier periods, and when it reaches 1.0, it means that this increase has been sustained for approximately 16 h. At this point, estrus is considered to have occurred (i.e., the sow is predicted to be in estrus).

4.3. Optimal Timing Prediction for Artificial Insemination

To evaluate the proposed method for predicting the optimal artificial insemination timing, the difference between the actual insemination timing selected by farm workers (${Pred}_0$) and the insemination timing predicted using the proposed method (${Pred}_1$) was calculated, as shown in Equation (5). At the farm where the experiment was conducted, estrus detection and artificial insemination were performed twice daily at 08:00 and 16:00. Given that the interval between estrus detection and insemination can vary up to 16.7 h, a similarity range was defined as expressed in Equation (6).

$$\Delta d=\left|{Pred}_0-{Pred}_1\right|$$

(5)

$$L_k\left(\begin{array}{c}{L}_{1 }, \Delta d<16.7 \mathrm{h}\\ \begin{array}{c}{L}_2, 16.7 \mathrm{h}\le \Delta d<24 \mathrm{h}\\ L_3, 24 \mathrm{h}\le \Delta d<32 \mathrm{h}\\ L_4, 32 \mathrm{h}\le \Delta d<40 \mathrm{h}\end{array}\\ L_5, 40 \mathrm{h}\le \Delta d\end{array}\right.$$

(6)

Sixty sows were tracked from 15 February 2024, to 5 June 2024. Figure 6 shows the similarity range calculated using Equation (6). Among the predicted values, 70% (42/60 sows) had a difference in less than 24 h compared to the actual insemination timing selected by farm workers. Given that farm workers perform estrus detection at 8 or 16 h intervals (08:00 and 16:00), we defined the highest similarity level ($L_1$) as a prediction difference of less than 16 h, aligning our evaluation metric with the operational realities of the farm.

5. Discussion

Table 1. Sample dataset showing sow (Sus scrofa domesticus Erxleben) parity, actual and predicted insemination timing, similarity range, and litter size.

ID	Parity	Pred_0	Pred_1	DELTA_d	L_k	Piglet_T	Piglet_L
50–60	6	24y02m19d 16 h	24y02m18d 19 h	20 h 28 m	2	16	16
32–33	4	24y02m20d 16 h	24y02m20d 08 h	7 h 50 m	1	18	18
60–72	8	24y02m19d 8 h	24y02m18d 18 h	13 h 50 m	1	19	9
23–82	5	24y02m19d 8 h	24y02m18d 18 h	13 h 21 m	1	16	14
58–87	8	24y02m20d 16 h	24y02m20d 22 h	6 h 10 m	1	17	17
45–74	6	24y02m18d 16 h	24y02m18d 18 h	2 h 45 m	1	16	11

The meaning of each variable in the table is as follows: ID—Unique identification number for each sow; Parity—Number of parities (number of times the sow has given birth); Pred_0—Actual insemination timing performed by farm workers; Pred_1—Recommended insemination timing using the proposed method; DELTA_d—Difference between the actual insemination timing and recommended timing, calculated using Equation (5); L_k—Similarity range calculated using Equation (6); Piglet_T—Total number of piglets born; Piglet_L—Number of live-born piglets.

presents a sample dataset on tracked sows, including parity, insemination timing, and litter size.

Experimental validation demonstrated that the proposed method achieved 70% concordance with farm worker decisions, with predicted insemination timings differing by less than 24 h from actual implementation times. This level of agreement suggests practical viability for field deployment in commercial swine operations. Among the $L_1$ to $L_3$ groups, the total litter size was consistently 14 or more (

Table 2. Average litter size by similarity range.

Similarity Range	Average
	Total Litter Size	Live-Born Litter Size
$L_1$	14.07	13.07
$L_2$	14.78	13.5
$L_3$	14.4	13.2
$L_4$	13.25	12.75
$L_5$	13	13
Fail	15.75	11.75
Total Average	14.25	13.05

). The ‘Fail’ category represents cases where the prediction error exceeded the maximum defined range ($L_5$), indicating a significant deviation from the worker’s decision.

It is important to reconcile the non-significant ANOVA result (p = 0.207) with the significant negative correlation. We intentionally employed two complementary statistical approaches to evaluate the model from two different perspectives: macroscopic utility and microscopic precision. The ANOVA test assesses the model’s utility at a macroscopic level by categorizing prediction errors into broad groups, such as L1 (a window of <16.7 h). The non-significant result from this analysis suggests that, from a practical standpoint, predictions falling within these wide windows yield comparable reproductive outcomes. In contrast, the correlation analysis performs a microscopic assessment, using the precise, continuous prediction error (Δd) in minutes. The significant negative correlation revealed by this analysis (r = −0.23, p < 0.05) demonstrates that even within a ‘successful’ window, smaller errors (i.e., higher precision) are systematically associated with a higher number of live-born piglets.

Figure 7 shows the clustering results of time series activity data from 60 sows. In Figure 7, the x-axis represents the time axis over one week, identical to the x-axis used in Figure 5, with activity values smoothed using a double moving average. The y-axis indicates the activity level of each sow, calculated from posture transitions in the same manner as the blue line in Figure 5. Clusters 1 and 3 exhibited a pattern of sharply increasing activity at an early stage. The proportion of (L₁ + L₂) in Clusters 1 and 3 was relatively low, at 66.66% and 59.25%, respectively. Our clustering analysis of activity patterns reveals distinct estrus phenotypes among sows. Notably, the model demonstrated the high accuracy for sows in Cluster 2, which exhibited a classic rise-and-fall activity pattern consistent with established biological models of estrus (Figure 4). This suggests that our posture-based model is particularly effective for identifying the optimal insemination window in sows with this ‘classic’ estrus behavior. For other patterns, such as the early-peaking activity in Clusters 1 and 3, the model was less precise, indicating that different behavioral phenotypes may require tailored prediction algorithms in the future. Although the clustering analysis is descriptive, it underscores that estrus behaviors vary across sows. This finding highlights the potential of our system not only to predict insemination timing but also to classify sows according to their estrus behavior profiles, thereby providing a foundation for developing phenotype-specific prediction models.

Several limitations warrant consideration in interpreting these findings. Environmental factors including dust, ammonia, and humidity periodically compromised camera visibility, potentially affecting detection accuracy. Furthermore, the relatively modest sample size (n = 60) and uneven distribution across similarity groups ($L_3$~$L_5$: 5, 8 and 1 sow(s), respectively) limit the statistical power for comprehensive subgroup analyses. Furthermore, as our clustering analysis revealed, the model’s prediction accuracy varied across different estrus behavior phenotypes, showing higher precision for sows with ‘classic’ activity patterns.

Furthermore, correlation analysis between the prediction error (Δd) and the number of live-born piglets, focusing on sows with $L_1$ (Δd within 16 h), revealed a negative relationship in Figure 8. This suggests that higher temporal precision in estrus prediction contributes positively to reproductive outcomes, reinforcing the practical value of our approach for field deployment.

It is important to reconcile the non-significant ANOVA result (p = 0.207) with the clear negative correlation found between prediction error and live-born piglets (Figure 8). The ANOVA test categorized prediction errors into broad groups (e.g., $L_1$ spans errors from 0 to 16 h), which may have masked the more subtle, continuous relationship between insemination timing accuracy and productivity. In contrast, the correlation analysis uses the precise, continuous prediction error (Δd), revealing that even within the ‘good’ prediction window (e.g., under 24 h), smaller errors are systematically associated with higher numbers of live-born piglets. This suggests that the model’s strength lies not only in identifying the general 24 h window but more importantly in pinpointing the optimal timing within that window, but more importantly, in pinpointing a more precise, optimal time within that window. Therefore, the granular correlation analysis provides stronger evidence for the practical, reproductive benefits of our system than the binned ANOVA analysis.

Although this study analyzed 60 sows, which may be considered a limited dataset, the results nevertheless suggest the potential of camera-based deep learning for non-invasive estrus detection and insemination timing prediction. The study was also conducted in a single farm under controlled conditions, which may restrict the external validity of the findings. Furthermore, only Landrace × Yorkshire F1 crossbred sows were included, which may limit the applicability of the results to other breeds. Future work should validate these findings with larger-scale and multi-farm datasets, including diverse layouts, lighting, and management practices, and also extend to multiple breeds to improve generalizability.

In addition, practical constraints such as the economic cost of deployment, limited internet connectivity, and the need for constant electricity may hinder adoption in smallholder or resource-poor farms. Behavioral variability related to sow age or body condition (e.g., gilts vs. multiparous sows) could also affect posture-based predictions, and occasional misclassification of postures may be mistaken for estrus. Addressing these issues in future research will be essential for robust and widely applicable implementation.

6. Conclusions

We proposed a method for predicting sow estrus and recommended the optimal timing for artificial insemination using camera footage and deep learning-based object detection technology. Sow posture information and posture duration, detected using camera footage and the YOLOv5 model, were stored as time series data. The stored $TSD$ were analyzed to observe sow activity and predict estrus.

Our experimental validation demonstrated that 70% of predictions (42/60 sows) achieved temporal concordance within 24 h of farm worker decisions, indicating substantial promise for automated estrus detection systems. Furthermore, our analysis suggests that higher temporal precision in estrus prediction contributes positively to reproductive outcomes, reinforcing the practical value of our approach. These findings underscore the potential of computer vision-based behavioral monitoring to enhance reproductive management in intensive swine production systems while reducing labor dependency.

Future work should focus on (1) expanding the dataset to include multiple breeds and seasonal variations, (2) developing adaptive algorithms that can learn individual sow behavior patterns, and (3) integrating multiple behavioral indicators beyond posture analysis for improved accuracy.