Sow posture detection for determining piglet crushing through a camera system

Overview

The contents are those of the author(s) and do not necessarily represent the official views of, nor an endorsement, by CDC/HHS, or the U.S. Government. To create a warning system based on sow postural changes, a posture detection model was trained using the YOLOv8 architecture, selected for its optimal balance of accuracy and size. The model was trained and validated on a sow posture dataset collected from a commercial pig farm. Then, the postural changes in sows were analyzed, involved in crushing events compared to those in non-crushing situations to identify the unique characteristics of each group. Based on this analysis, we established warning criteria that were subsequently validated and tested using video footage from the farm, as illustrated in Fig. 1.

Sow posture detection model

Sow posture dataset preparation

Video data was collected from a small-scale commercial pig farm in Thailand. Two surveillance cameras were positioned to monitor four sow farrowing pens, as depicted in Fig. 2. Footage was recorded by these cameras at a resolution of 1,920 × 1,080 pixels, with a frame rate of 15.15 frames per second (fps) and a total bit-rate of 2,126 kbps. For this study, footage was selected exclusively from the hours of 8:00 a.m. to 4:00 p.m. This period was chosen to ensure consistent natural lighting. It also corresponds to the sows’ most active hours, when more frequent postural changes are typically exhibited.

To create the posture dataset, an annotation pool was first established by sampling images from the video footage at 10-s intervals. Since the data originated from a commercial production farm, the dataset includes images from multiple sows, reflecting standard farm conditions. The DarkLabel annotation tool (darkpgmr, 2021) was used to label each sow’s posture into one of five distinct classes: standing (ST), sitting (SI), ventral recumbency (VR), right lateral recumbency (RR), and left lateral recumbency (LR). Detailed definitions for each class are provided in Table 1, with visual examples shown in Fig. 3. For each annotated instance, a bounding box was generated in the format required by YOLOv8, specifying the class label and the normalized coordinates of the box’s center, width, and height. Following the annotation process, all images were resized to 640 × 640 pixels using bilinear interpolation. Finally, the completed dataset was randomly partitioned into training and validation sets using an 80:20 ratio.

Table 1:

This table describes the definitions of sow postures used for annotation.

Posture	Definition
Standing	Sow stands on four feet
Sitting	Sow uses forelimb or hindlimb to support,
	while another lay down
Ventral recumbency	Sow lays down on the floor using
	chest
Right lateral recumbency	Sow uses right side of body lay down
	on the floor
Left lateral recumbency	Sow uses left side of body lay down
	on the floor

DOI: 10.7717/peerj-cs.3400/table-1

Crushing data

To establish ground truth, daily monitoring of all farrowing pens was conducted by the farm owner via camera footage. A standardized logging form was used to systematically record the pen number, any observed abnormalities, and the confirmation of a crushing event. Although these logs were submitted weekly, a specific data collection protocol was triggered by the confirmation of a crushing event. Following such a confirmation, the corresponding video footage from 8:00 a.m. to 4:00 p.m. was retrieved for a three-day period, covering the day the event occurred and the two preceding days.

Warning solutions

To develop the crushing warning system, three distinct classification approaches were evaluated. The first was a Decision Tree, a supervised learning model that creates a flowchart-like structure. In this model, data is sequentially split based on the most informative features to form a path of decision branches terminating in a final classification. The second approach was a support vector machine (SVM), an algorithm in which an optimal hyperplane is identified to separate data points into different classes by maximizing the margin between them. The third was a custom threshold-based voting method, which is detailed in the following sections.

Postural change dataset

To create the dataset for the warning system, a postural change dataset was derived by applying the trained posture detection model to the collected video footage. This footage included video from confirmed crushing days, the two days prior to a crushing event, and non-crushing days. A postural change event was defined as a difference in the sow’s classified posture between two consecutive 10-s intervals (Fig. 4). The total number of these events was then calculated and aggregated over four time windows including 5, 10, 30, and 60 min.

To identify an effective window for issuing warnings, a descriptive analysis was performed on this aggregated data. The data were structured for two separate comparisons. In the first analysis, data from crushing days and pre-crushing days were combined to form a “high-risk” group and compared against non-crushing days. In the second analysis, a more direct comparison was made between only crushing days and non-crushing days.

Threshold and voting

This method is a custom algorithm in which a statistical threshold is combined with a voting system to issue warnings. In this approach, the median of the postural change (PC) frequency is used as the key threshold. Sow activity is analyzed in 10-min intervals, and if the PC frequency exceeds the pre-determined median, the interval is flagged as a potential high-risk period.

To prevent false alarms from isolated spikes in activity, a final warning is only issued on an hourly basis through a voting process. Over a 60-min period, flags from the six 10-min intervals are collected. A final crushing warning is produced only if a majority of these intervals are flagged as high-risk. Through this voting mechanism, it is ensured that a warning is based on a sustained period of unusual activity rather than a single, brief event (as illustrated in Fig. 5).

Decision tree

The dataset used to train the decision tree models was derived from the 10-min postural change (PC) frequency data. This dataset was composed of 51 records from crushing days, 61 from pre-crushing days, and 132 from normal (non-crushing) days. The data were partitioned into training and validation sets using a 70:30 split. Then, two separate classification models were trained and evaluated.

Model 1: Crushing vs. normal. This model was trained to distinguish between crushing days and normal days. In the resulting tree, shown in Fig. 8, all extracted features were utilized except for the mean PC frequency. The optimal hyperparameters, selected via grid search, were a criterion of entropy or log_loss, a max_depth of 11, and a min_samples_split of 0.05. On the training data, a sensitivity of 43% and a specificity of 71% were achieved.

Model 2: High-risk vs. normal. The second model was trained to distinguish high-risk days (combining crushing and pre-crushing data) from normal days. In this tree, depicted in Fig. 9, all seven of the descriptive statistical features were incorporated. The best-performing hyperparameters for this model were a criterion of entropy or log_loss, no maximum depth (max_depth=None), and a min_samples_split of 0.2. With this configuration, a sensitivity of 63% and a specificity of 41% were achieved on the training set.

Support vector machine

The SVM models were trained and validated using the same dataset and 70:30 split as the Decision Trees. During the grid search process, it was noted that some hyperparameter combinations resulted in models that were unable to identify any true positives, leading to a sensitivity of zero. The optimal results for the two classification scenarios are detailed below.

Model 1: Crushing vs. normal. For the task of distinguishing crushing days from normal days, a polynomial kernel with a regularization parameter of C = 1.0 was used by the best-performing SVM model. On the training data, a sensitivity of 24% and a specificity of 96% were achieved.

Model 2: High-risk vs. normal. For the classification of high-risk days against normal days, a sigmoid kernel with C = 1.0 was utilized by the optimal model. With this configuration, a sensitivity of 51% and a specificity of 55% were achieved on the training set.

Threshold and voting

The threshold-based voting method was evaluated using the median of the 10-min postural change (PC) frequency as the activation threshold. As with the other models, its performance was assessed on the 28-day test set using two different classification scenarios.

Model 1: Crushing vs. normal. When tasked with distinguishing crushing days from normal days, the system’s performance was evaluated on both an hourly and daily basis.

• Hourly performance: The sensitivity was 62.50% and the specificity was 60.25%.

• Daily performance: The performance improved with daily aggregation, yielding a sensitivity of 71.42% and a specificity of 84.61%.

Model 2: High-risk vs. normal. In the second scenario, pre-crushing days were included in the positive class (high-risk). This change resulted in a general decrease in performance.

• Hourly performance: The sensitivity and specificity fell to 53.38% and 59.74%, respectively.

• Daily performance: On a daily basis, the specificity remained consistent while the sensitivity dropped to 66.67%.

Unlike the decision tree and SVM models, which were evaluated on a test set (30% of the data), the threshold and voting technique was evaluated using the entire dataset. A comprehensive comparison of the performance of all three methods (Decision Tree, SVM, and Threshold-Voting) is provided in Table 5.

Posture detection model performance and limitations

In this study, a high-performing sow posture detection model was successfully developed, and an overall mAP@50 of 0.976 was achieved with the computationally efficient YOLOv8n. This result is highly promising, demonstrating the viability of using deep learning for posture analysis even with a modestly sized dataset collected under the constraints of a commercial small-scale farm. The success was likely attributable to the use of a pre-trained YOLOv8 model with transfer learning.

Despite the high overall accuracy, several limitations were identified. Slightly lower performance was obtained for the sitting posture, which was attributed primarily to class imbalance in the training data, as there were fewer sitting instances compared to other postures. This issue was composed by annotation challenges that occur when the sow’s body occasionally obscured its legs, particularly from certain camera angles. As a result, the sitting posture was most frequently misclassified as ventral recumbency, as these two postures can appear visually similar depending on the viewing angle (Fig. 10). These challenges, including class imbalance and the impact of camera perspective, are consistent with findings from other studies (Chen, Lu & Wang, 2023). Furthermore, the model’s generalization is currently limited to the environment of a single farm. Therefore, future work should focus on expanding the dataset with images from diverse farm settings (e.g., both open and closed farrowing house systems) to enhance robustness and precision.

Comparison with alternative methods

When compared to other technological approaches for monitoring sow behavior, the image-based posture detection method presented in this study is competitive. For instance, its performance is comparable to systems that use vibration energy to classify postures, particularly for detecting standing (Dong et al., 2023). An alternative approach involves the use of audio data to detect piglet distress sounds, such as screaming, as an indicator of a crushing event. For example, an AI-based IoT system was developed by Chen, Lin & Chen (2021) in which a convolutional neural network (CNN) was used to detect piglet screams and trigger an actuator to encourage sow movement. In that study, it was found that using min-max scaling for audio pre-processing yielded a higher and more stable classification accuracy (99.4%) than the more common Mel Frequency Cepstral Coefficient (MFCC) method.

While such audio-based systems are highly accurate at detecting an ongoing event, the posture-based approach presented here is aimed at identifying high-risk behaviors before a crushing event occurs, offering a preventative rather than a reactive solution. This preventative focus makes the method particularly well-suited for providing actionable pre-warnings on small-scale farms. The high accuracy of image-based models for this task has also been demonstrated in other studies (Alameer, Kyriazakis & Bacardit, 2020; Riekert et al., 2020; Zheng et al., 2018, 2020).

Interpretation of postural change analysis

The frequency of postural changes was confirmed to be a key indicator of crushing risk, as a higher frequency of these changes was observed on crushing and pre-crushing days compared to non-crushing days. This finding is in alignment with the broader literature, in which sow restlessness and movement patterns have been identified as critical risk factors. For instance, it has been noted in previous studies that while non-crushing sows may exhibit more standing and walking, more high-risk behaviors like rolling tend to be displayed by crushing-prone sows. Similarly, the transition from standing to lying is a well-documented critical moment for crushing events, particularly in primiparous sows (Valros et al., 2003; Wischner et al., 2009; Illmann et al., 2013; Manteuffel et al., 2017).

It is important to note that this study was conducted in a system with farrowing crates, which restrict sow movement to reduce crushing risk but also limit natural behaviors. The distinct postural patterns that were detected, even within this constrained environment, underscore the robustness of this metric as a warning indicator.

Justification of the 10-min warning interval

For the development of the warning system, a 10-min interval was selected for compiling postural changes. This choice was based on a balance between responsiveness and reliability. As shown in the descriptive analysis (Table 4), the median postural change frequency in this interval during high-risk periods was approximately double that of non-crushing periods. A clear separation and a sufficient margin were thereby provided for the threshold to be effective while generating minor detection errors, an advantage not as stated in the 5-min interval.

Furthermore, the 10-min interval allows six data points to be included in the hourly voting mechanism. A more robust basis for a warning decision is thus provided compared to the two and one data points available from the 30- and 60-min intervals, respectively. This reduces the likelihood of false alarms being triggered by short, isolated bursts of activity.

Evaluation of warning system performance

In this study, three different models for the piglet crushing warning system were evaluated, with performance being assessed on both an hourly and daily basis to accommodate different user needs. A notable discrepancy between the models’ performance on the training and testing sets was observed. While promising performance was shown by the decision tree and SVM models on the training data, their performance dropped significantly on the unseen testing data for hourly predictions. In contrast, their daily-based performance was considerably higher (Table 5).

This was due to the limited number of crushing events in the dataset. With few positive instances, complex models like decision trees and SVMs are prone to overfitting the training data, leading to poor generalization on an instance-by-instance (hourly) basis. Through daily basis, which considers the entire data, these instance-level prediction errors are smoothed out, resulting in higher performance. In contrast, more balanced and robust performance on the testing set for both hourly and daily warnings was demonstrated by the simpler threshold-and-voting method. The simpler method worked better because it used a basic statistical rule instead of trying to find complex patterns in the small dataset. Also, looking at the data visually (Fig. 11) showed that the different groups were mixed together, which made it hard for complex models like SVMs to draw a clear line to separate them.

Practical implications and justification of the chosen method

Given these results, the threshold-and-voting method is identified as the most practical and reliable solution for the target application, with its primary advantage being straightforward interpretability. For smallholder farmers, who may check on animals infrequently, a reliable warning is critical, as a system with a high rate of false positives or low sensitivity would quickly be ignored.

The goal of this research was to develop a proactive tool for farmers with limited labor and experience. Through an automated warning system based on postural change frequency, awareness can be effectively raised, and interventions can be prompted, such as adjusting the crate, increasing sow comfort, or fostering piglets to another sow. Although the hourly sensitivity is moderate, a valuable layer of monitoring, which would otherwise be absent, is provided.

Future work and system enhancement

While postural change frequency is a promising indicator, the warning system could be significantly enhanced by incorporating additional data streams, through which both sensitivity and specificity could be improved. For example, an audio-based piglet scream detection system could be integrated to act as a secondary confirmation of a crushing event. High accuracy for vocalization-based detection, with specificity up to 95%, has been shown in previous research (Manteuffel et al., 2017), and a high probability of success has been demonstrated by IoT-based rescue systems triggered by screams (Chen, Lin & Chen, 2021).

Furthermore, the model’s precision could be improved by including environmental and health-related features. Factors such as temperature and humidity are known to influence sow behavior. For instance, a lower frequency of postural changes is often exhibited by heat-stressed sows (Liu et al., 2021). By accounting for such external variables, as well as the sow’s health status and parity, a more dynamic and adaptive warning threshold could be developed, moving beyond the fixed threshold used in this study (Yi et al., 2019; Kaiser et al., 2022). Through the integration of these relevant features, it is believed that a more precise and timely crushing warning system can be developed.