Detection of Respiratory Rate of Dairy Cows Based on Infrared Thermography and Deep Learning

Abstract

The respiratory status of dairy cows can reflect their heat stress and health conditions. It is widely used in the precision farming of dairy cows. To realize intelligent monitoring of cow respiratory status, a system based on infrared thermography was constructed. First, the YOLO v8 model was used to detect and track the nose of cows in thermal images. Three instance segmentation models, Mask2Former, Mask R-CNN and SOLOv2, were used to segment the nostrils from the nose area. Second, the hash algorithm was used to extract the temperature of each pixel in the nostril area of a cow to obtain the temperature change curve. Finally, the sliding window approach was used to detect the peaks of the filtered temperature curve to obtain the respiratory rate of cows. Totally 81 infrared thermography videos were used to test the system, and the results showed that the AP₅₀ of nose detection reached 98.6%, and the AP₅₀ of nostril segmentation reached 75.71%. The accuracy of the respiratory rate was 94.58%, and the correlation coefficient R was 0.95. Combining infrared thermography technology with deep learning models can improve the accuracy and usability of the respiratory monitoring system for dairy cows.

1. Introduction

The respiratory behaviour of dairy cows is closely related to their health status, and the respiratory rate (RR) of healthy dairy cows is approximately 12 to 28 times per minute [1]. When the RR fluctuates abnormally, it may be closely related to cow disease, shed comfort, environmental temperature, etc. [2,3,4]. Automatic monitoring of the breathing behavior of cows not only helps breeders grasp the health status of cows, but also ensures the production performance of cows and farm benefits [5]. The detection of the respiratory behavior of cows is mainly completed by manual counting, which has high labor and time costs and low accuracy [6]. Therefore, it is important to realize the intelligent and automatic monitoring of dairy cow respiration to maintain the production performance of dairy cows and promote the development of precision breeding [7]. Today’s respiratory status monitoring methods are mainly divided into contact and noncontact methods [8,9]. For contact detection of breathing behaviour, Eigenberg et al. (2000) fixed a film pressure sensor on the abdomen of dairy cows with a belt, which recorded the regular fluctuations of the abdomen. In a direct sunlight environment, when the temperature changed by 1 °C, the RR change in dairy cows was 6.6 times/min [10]. Milan et al. (2016) fixed the temperature sensor near the nostril of the cow with a halter to monitor the temperature change in the air near the nostril of the cow and calculated the breathing rate according to the number of oscillations of the temperature signal [6]. Strutzke et al. (2019) designed a cow contact breathing monitoring device based on a differential pressure sensor. This method was highly correlated with manual counting results in three different states of cows: sleeping (correlation coefficient, R = 0.92), lying down (R = 0.98), and standing (R = 0.99) [11]. Although the technology of respiratory status monitoring methods based on contact sensors is mature, it is difficult to apply in practice because of the poor compliance of dairy cows, the complex environment, and many interference factors. In addition, the contact sensor easily causes the stress response of cows, affecting animal welfare. Noncontact respiratory status monitoring does not easily cause a stress response in dairy cows because of its fast response and flexible installation; this method is favoured by an increasing number of researchers and is an important development direction of intelligent respiratory status monitoring of dairy cows [12].

For noncontact detection of breathing behaviour, Tang Liang et al. (2015) used machine vision technology to detect the breathing rate of pigs by assessing the changes in their abdominal area while breathing. The results from 10 test videos with low interference showed that the RR detection accuracy of this method could reach 92.0% [13]. Benetazzo et al. (2014) collected and calculated the distance between the chest and the camera based on the Kinect depth camera to detect breathing behaviour. Five sets of test results in different environments show that this method is a feasible alternative for breath detection [14]. Zhao Kaixuan et al. (2014) calculated the relative movement speed of each pixel point of a video frame image by using a Horn–Schunck optical flow method, screened out breathing motion points in the pixel points by using cyclic Otsu algorithm, dynamically calculated the period of a speed direction curve, calculated the RR of the cow and further judged whether the cow breathing was abnormal. The accuracy rate of RR detection was 95.68% and the success rate of abnormal detection was 89.06% [15]. Zeng et al. (2020) proposed a breathing detection system based on WIFI, which uses multiple antennas on the router to receive the object reflection signal and extract and perceive the object breathing signal through blind source separation and independent component analysis [16].

In the above methods, the costs of the depth camera and the requirements for the compliance of a detection object and the environment are relatively high. The Horn–Schunck optical flow method is computationally expensive, and its detection speed and real-time performance are not good. In dairy farms, the environment is complex, and the noise is serious, so it is not suitable to use WIFI for breath detection. Infrared thermography (IRT) [17] is a noncontact method that has been validated for the measurement of RR in adult dairy cattle [18]. In this study, to achieve accurate, rapid, and automatic monitoring of the respiratory status of dairy cows, a system based on IRT was proposed. First, a YOLO v8 object detection model was used to train a detector of a cow’s nose in IRT images and three kinds of instance segmentation models were used to segment the nostrils from the nose area. Second, the hash algorithm was used to extract the temperature of each pixel in the nostril area of the cow to obtain the temperature change curve. Finally, the temperature change curve was filtered, the wave crest was detected, and the RR was obtained. This method can provide a reference for the accurate and automatic detection of RR and realize automatic monitoring.

2. Materials and Methods

2.1. Test Data

2.1.1. Data Acquisition

The experimental video was collected in Nanyang Sansege Dairy Farm, China, on a sunny day in July 2021. The experimental objects were Holstein cows with an average age of 26 months, that were in lactation and had no history of respiratory disease. The cows did not drink water for half an hour before image acquisition. The temperature of the barn was around 32 °C and the humidity was 65%. When a cow ruminates, it will stretch its head out of the railing, and the individual activity is relatively stable. The FLIR A615 infrared thermal camera was fixed at a distance of 1.9 m from the cow object and 0.75 m from the ground. During the cow ruminating process, the camera recorded the videos of five cows: the collection time of each cow was approximately 3 min, the total time was approximately 15 min, the frame rate of the video was 25 frames per second (FPS), and the total frame was 22,485. The video resolution is 640 pixels (horizontal) × 480 pixels (vertical), and the shooting took place under natural lighting conditions.

2.1.2. Dataset Production

The preparation process of the test dataset is as follows:

• Data preprocessing: approximately 15 min of video is cut into 10 s of video data, totalling 92 segments;
• Picture acquisition: each frame picture is extracted from the recorded IRT video file, and the picture is saved in a JPG format;
• Dividing the dataset: to better train the parameters and predict the performance of the model, the dataset is divided into a training set and a test set at a ratio of 7:3 by random sampling. The training set is used as a training object to extract the neural network model. When the amount of data in the training set is sufficient, it is helpful to enhance the recognition ability of the model application. The test set is used to verify the accuracy of the model.
• Data labelling: the nose, left nostril, and right nostril of the cow in each frame image of the training set are manually labelled. After the picture is annotated, a corresponding file is generated, which records the annotated area, object category, and other information;
• Dataset format: the file is converted into a standard format dataset for model training and testing.

2.2. Nostril Detection Based on Cascaded Deep Learning

2.2.1. Nose Detection Based on YOLO V8

Intelligent video surveillance technology has been widely used in intelligent breeding, and it has become a research hotspot in the field of agricultural engineering to improve the level of dairy cow behavior monitoring by using computer vision technology [19]. To realize the automatic monitoring of a cow’s respiratory status in video images, accurate nose positioning is essential.

Nose detection refers to locating the position of a cow’s nose from an IRT image. To facilitate subsequent algorithm development and RR detection, the first part is to detect the object in the IRT images and extract the cow’s nose as the basis for further analysis.

YOLO (you only look once) [20] is a general object detection model based on a convolutional neural network. It is characterized by using regression ideas to predict the rectangular frame of the detected object during detection, so it can greatly improve the computational efficiency compared with other detection methods that need a selective search to extract a large number of candidate areas. YOLO v8 is the latest version of the YOLO series models, providing a new SOTA model, replacing the C3 structure of YOLO v5 with a C2f structure with richer gradient flow, which greatly improves model performance. The model can complete the tasks of object detection, instance segmentation, and image classification, and the loss function of the model adopts the positive sample allocation strategy. The model is faster and more accurate. At the same time, the data enhancement part of the training introduces the operation of turning off Mosaic enhancement in the last 10 epochs of YOLO X, which can effectively improve accuracy.

Therefore, YOLO v8 (you only look once v8) [21], an object detection model based on a deep learning neural network, is selected in this paper because of its detection accuracy and efficiency. As a detector of a cow’s nose, the detection result is represented by a rectangular box in the image. In addition, even if the nose can be accurately detected, how to obtain the breath temperature flow is still a significant problem. Therefore, it is necessary to further analyse the results of nose object detection.

The model training parameters are set as follows: (1) The batch size is 8 with 10,000 iterations, the initial learning rate is 0.01, with 3000 warm-up iterations, optimizer is SGD, and the learning rate schedule is linear; (2) the input image width and height of the network are 640 pixels × 480 pixels; and (3) the hardware platform configuration for model training involves one NVIDIA GTX 2080 Ti graphics card (Nvidia Corporation, Santa Clara, CA, USA) with 16 GB memory and one Intel ^® Core ™ i5-10400F processor (Intel Corporation, Santa Clara, CA, USA).

2.2.2. Nostril Segmentation Based on Deep Learning

With the deepening of research, only obtaining the nose position is far from meeting the needs of precise breeding. Therefore, it is important to accurately obtain the nostril information of dairy cows to improve the level of respiratory status monitoring.

In this paper, the instance segmentation algorithm is introduced to obtain the nostril area. In recent years, with the rapid development of deep learning technology, E2EC, Mask R-CNN, YOLO v1-v8 and other semantic segmentation, instance segmentation or object detection networks have been proposed. Among them, the object detection speed is relatively fast, but the object and part of the background will be selected at the same time, so additional processing is needed [22]. Semantic segmentation and instance segmentation can directly obtain more accurate contour information. Instance segmentation is an advanced task that combines object detection and semantic segmentation. It not only needs to distinguish different instances but also needs to mark them with masks. This can further distinguish different individuals of the same category, but it is also more time consuming. The development of instance segmentation started with Mask RCNN [23].

Mask2Former is a new architecture capable of handling any image segmentation task (panoramic, instance, or semantic). Its key components include hidden attention, which extracts local features by constraining cross-attention within the predicted mask area. In addition to reducing the research effort by at least a factor of three, it significantly outperforms the best-dedicated architectures on four popular datasets. Most notably, Mask2Former has proposed new and more advanced techniques for panoramic segmentation (57.8 PQ on COCO), instance segmentation (50.1 AP on COCO), and semantic segmentation (57.7 mIOU on ADE20K) [24].

To better obtain the temperature change information in the nostril area of dairy cows, this study uses the method of extracting the object from each frame image and calculating the temperature of the nostril area of each frame image to complete respiratory status monitoring.

2.3. Temperature Extraction of the Nostril Area Based on the Hash Algorithm

When a cow breathes, the temperature of the nostril area will change regularly with the change in respiratory status. During inhalation, cool air is drawn in from the environment, resulting in a cooling of the nostrils and a subsequent darker appearance of the nostrils in infrared recordings. In contrast, during exhalation, warm air is expelled into the environment, resulting in a warming of the nostrils and a subsequent warmer reading with a brighter appearance of the nostrils in infrared recordings. According to the obtained information of the nostril area of the cow, the temperature of each pixel in the nostril area of each frame image when the cow breathes is obtained by using the hash algorithm. The specific steps are shown in Figure 1.

RGB is a common colour space, and its colour mode includes three colour channels: red (R), green (G), and blue (B). The “colour space” for any colour can be defined as a fixed number or variable. Among them, the methods of coding a colour are collectively referred to as “colour space”. In this experiment, the RGB value of each pixel has its corresponding colour coding.

According to the colour–temperature scale on the right side of the sampled video, the temperature value is determined by obtaining the mapping relationship between the RGB value of each pixel in the scale and the temperature and applying a hash algorithm. A hash map can map one value to another through a certain functional relationship. The hash map has the advantages of high efficiency, fast searching, insert and delete operations, and relatively easy programming. The disadvantage is that it is based on arrays, which are difficult to extend after they are created.

A portion of the pixels in the nostril area in each frame image can obtain the corresponding temperature value by traversing the colour–temperature scale and using the hash table, but the RGB value in the colour–temperature scale cannot completely match the RGB value of each pixel. To calculate RR through the temperature change in the nostril area, this paper adopts the method of calculating the minimum distance between the RGB value that is not hit and the known RGB value in the colour–temperature scale to obtain the temperature corresponding to the RGB value that is not hit.

The common distance measures are the Euclidean distance, Hamming distance, Manhattan distance, Chebyshev distance, Jaccard distance, etc. In this study, the classical Euclidean distance is used to deal with the pixels that are not hit, and the Euclidean distance has the advantages of simplicity and rapidity.

To obtain the distance between the colour pixel to be acquired and the known colour pixel, the Euclidean distance ${dist}_{a_k}$ between the colour vector (R, G, B) of pixel $M_a$ that is not hit and the colour vector (R, G, B) of pixel $N_k$ that is known in the colour–temperature scale is calculated. Selecting a minimum value for the distance, the RGB value of the known pixel corresponding to the minimum value is the RGB value of the colour pixel to be obtained. $M_a$ is the a-th not hit pixel, $N_k$ is the k-th known pixel, $M_a$ and $N_k$ are three-channel RGB pixel points, and the calculation formula for ${dist}_{a_k}$ is as follows:

$${dist}_{a_k}=\sqrt{{\left(M_a^r-N_k^r\right)}^2+{\left(M_a^g-N_k^g\right)}^2+{\left(M_a^b-N_k^b\right)}^2}$$

(1)

In Formula (1), $M_a^r$, $M_a^g$ and $M_a^b$ are the R, G, and B components of the a-th not hit pixel, respectively, and $N_k^r$, $N_k^g$ and $N_k^b$ are the R, G, and B components of the k-th known pixel, respectively. ${dist}_{a_1}$, ${dist}_{a_2}$, ${dist}_{a_3}$……${dist}_{a_k}$ from $M_a$ to $N_1$, $N_2$, $N_3$……$N_k$ can be obtained through Formula (1). The smallest of them ${\left({dist}_{a_k}\right)}_{min}$ is selected, corresponding to the known pixel $N_k$. The temperature value corresponding to $M_a$ is obtained by using the hash algorithm.

By using the above method, the RGB value of each pixel in the nostril areas of each frame of IRT pictures can be obtained, and then the corresponding temperature value can be obtained.

2.4. Detection of RR

This module processes the obtained temperature data and detects the RR of dairy cows according to temperature changes.

The temperature extracted by the above method is a discrete signal consisting of the average value of all temperatures in each frame. First, the signal is subjected to spectral analysis by using a Fourier transform. Second, the data need to be filtered. Finally, the RR was measured by the temperature curve. The specific steps are shown in Figure 2.

There are mainly four kinds of filters according to the characteristics and different applications of the filter: the Butterworth filter, Chebyshev I filter, Chebyshev II filter, and elliptic filter [25]. Among them, the Butterworth filter is a more commonly used signal processing filter, and its role is to make the frequency response of the signal as flat as possible in the passband, so it is also called the maximum flat amplitude filter. This filter is used in this test. A typical Butterworth filter is a low-pass filter.

In this test, peaks are marked according to their prominence. The prominence of a peak measures how much the peak protrudes from the baseline around the signal and is defined as the vertical distance between the peak and its lowest contour, called the prominence of the peak (P), which is a measure of the independence of the peak. It can also be understood as the minimum height at which the peak needs to fall to reach a higher point. In this study, it was defined that a wave transformation with a small protrusion in the temperature change curve of dairy cows during respiration did not belong to a respiration process. In addition, the purpose of not marking a small wave peak could be achieved by setting the P value; that is, when the respiration temperature difference was less than the P value, it was not defined as respiration. In this paper, the time of one breath is defined as the time between two adjacent peaks. Using the sliding window method to traverse the signal and mark each peak (extreme point), the number of peaks is the number of breaths.

2.5. Evaluation Index of the Model

To verify the accuracy of the respiratory status detection algorithm studied in this paper, two evaluation indices, the instance segmentation evaluation parameter average precision (AP) and accuracy ω, were used to evaluate the model performance. Note that for the nose, if the parallel cross-ratio (intersection over union, IOU) between the detection rectangular frame and the precalibrated real rectangular frame is greater than 0.5, then the detection is defined as correct.

The mean value of precision AP is used to measure the localization accuracy of nose detection and nostril instance segmentation in this study. The instance segmentation evaluation parameters AP and AP₅₀ in the model are selected. AP₅₀ refers to the measured value of AP when the IOU threshold is greater than 0.5.

The accuracy of RR detections ω is used to verify the accuracy of the algorithm. In each test video, the actual number of breaths of the cow obtained through manual counting is n₁, and the number of breaths detected by the algorithm in this study is n₂. The detection accuracy ω of this research algorithm is defined, as shown in Formula (2):

$$\mathsf{\omega }=(1-\frac{\left|n_1-n_2\right|}{n_1})\times 100\%$$

(2)

3. Results and Discussion

3.1. Analysis of Test Results

3.1.1. Nose Object Detection Result

Figure 3 is an example of a cow’s nose detection result in a test set. In the figure, the results of a cow’s nose detection is marked by blue rectangular boxes. Due to the high detection accuracy of the YOLO v8 model, the cow nose in most of the images can be accurately detected. However, in some special cases, such as fast motion leading to blurring and other factors, there will still be some cases of missed detection. Alternatively, there is a nose false alarm detection; that is, there is no nose in the picture but the nose is detected.

To verify the effectiveness of the proposed method, experiments are carried out on the test set. Then, the video is screened according to the nose detection result in each video. If the nose is not detected in the video frame, it is considered that the nose is not in the field of view in the frame. To avoid too large errors in RR detecting, if the total time of not detecting the nose in each video is more than 1 s, then the video is rejected. Finally, 11 videos are eliminated, leaving 81 videos.

Table 1. Detection results of cows’ nose object.

Algorithm	AP50	AP75	AP50–95
YOLO v8	98.6	84.3	81.2

shows the results of cow nose detection using the YOLO v8 model. After analysis, the reasons for the accuracy error are missing detection and false alarm detection, but the impact on accuracy is very small. Combined with Table 1, it can be seen that the YOLO v8 model can obtain ideal results in the detection of a cow’s nose.

3.1.2. Comparison of Instance Segmentation Algorithms

To verify the reliability and stability of the Mask2Former network in nostril object extraction, Mask R-CNN and SOLOv2 [26], which are the mainstream instance segmentation algorithms at present, are selected. It is compared with Mask2Former, which is an instance segmentation algorithm used in this paper. Among them, Mask R-CNN belongs to the two-stage instance segmentation algorithm of “first detection and then segmentation”. In the comparison of the segmentation accuracy of many models, the average accuracy of the two-stage framework is better. This is because two-stage methods such as Mask R-CNN are more suitable for area-based detection; although the model framework is not flexible, the training time is long and it is not suitable for real-time applications [27]. The single-stage SOLOv2 model can perform segmentation and detection in parallel, so it can achieve good performance and reduce processing time, which is usually faster and more real-time than two-stage [28]. However, it is difficult to detect small objects in a single-stage framework, and a two-stage framework is better in this respect.

Based on the nose detection results, three instance segmentation algorithms, Mask2Former, Mask R-CNN, and SOLOv2, are used for training, and the detection results of the obtained model on the test set are shown in

Table 2. Comparison of nostril segmentation evaluation parameters.

Algorithm	AP50
Mask2Former	75.71
Mask R-CNN	67.32
SOLOv2	68.27

. It can be seen that the instance segmentation algorithm Mask2Former used in this paper has obvious advantages over Mask R-CNN and SOLOv2.

Based on the nose detection results, the Mask2Former instance segmentation method is used to extract the cow nostril area from the cow IRT images numbered 2855, and the effect is shown in Figure 4. The nostril segmentation result of the cow in the figure is indicated by a mask. The size of the nostril area is approximately 20 × 16 pixels, and objects with pixel areas less than 32 × 32 pixels are defined as small objects in the COCO dataset. One of the difficulties of small object detection is that the RGB information of the nostrils is too small, so it contains too few discriminative features. Second, the dataset is unbalanced, and there is a serious image-level imbalance in COCO [29]. Third, small objects have different degrees of occlusion, ambiguity, and incompleteness. Therefore, there are different degrees of false detection and missed detection in the segmentation results of the nostril area, and the results are shown in Figure 5.

Table 3. Proportion of false detection and missed detection of nostril.

Cow Number	Total Detection (/Frame)	False Detection (/Frame)	Missed Detection (/Frame)	Proportion
1386	4871	297	3	6.16%
1563	3500	171	12	5.23%
1568	4500	215	17	5.16%
1809	3449	156	10	4.81%
2855	3750	134	6	3.73%

is obtained after the traversal analysis of the results after nostril detection.

The analysis shows that one of the reasons for the error is that the nostril is a small object. Moreover, the method in this paper must be based on the correct detection of the cow nose, and the final segmentation of the nostril area is highly dependent on the accuracy of nose detection. If the missing rate of the nose is high or there are a large number of false alarm objects, then the accuracy of nostril segmentation will be reduced. According to Table 3, the average proportion of false detection and missed detection is only 5.02%, which shows that the algorithm has high accuracy for cow nostril detection.

3.1.3. Results of RR Detecting

Figure 6 depicts the result of visualizing the extracted temperature of the nostril area of each frame after extracting the temperature of cow No. 1386.

In this study, the temperature obtained in the previous step was further analysed according to the method described in Section 2.4, and the results are shown in Figure 7. Figure 7a is a spectrum analysis diagram before filtering; Figure 7b is a spectrum analysis diagram after Butterworth filtering; and Figure 7c is the result of filtering the Fourier transformed data with the Butterworth filter and traversing the signal marker peaks with the sliding window method in this study.

The results of detecting the number of breaths on the test data by using this model are shown in Figure 8. The number of breaths of cows in 81 IRT videos were counted and compared using the model and manually. Through the error analysis of manual counting and model counting, 50 of the 81 counting points have an error of 0, accounting for 61.73%, and 31 have an error of 1, accounting for 38.27%, indicating a high accuracy rate.

Table 4. Detection accuracy of respiration rate for different cows.

Cow Number	1386	1809	2855	1563	1568
Accuracy ω	92.57%	95.24%	97.22%	93.89%	93.96%

shows the accuracy ω for the detection of RR.

Figure 8 and Table 4 show that the average accuracy ω of breath count detection of the test data is 94.58%, which demonstrates strong resistance to interference factors, such as temporary foreign object intrusion, video capture quality, light, and shade changes, and good robustness for RR detecting. Additionally, the average detection rate is high, which meets the practical requirements.

3.2. Discussion

3.2.1. Analysis of Nose–Nostril Detection

This paper also tests the direct instance segmentation model, which cannot fully adapt to the experimental data. Mask2Former is used to train the whole group of videos directly. Because the cow’s nostril area is small and contains fewer features, there are more interference features, leading to the training results AP₅₀ being only 55.0, so the information of nostril temperature change could not be accurately obtained. After changing to the cascade segmentation method, the cow nose was detected first, and the detection area was reduced to the cow nose area with less interference. Then, segmentation of the nostrils is carried out. The segmented AP₅₀ of the nostril can reach 75.71, which is 37.7% higher than that of the direct segmentation model.

The results show that the cascaded model can effectively reduce the interference features outside the nose and reduce the error. The direct segmentation model will falsely detect small objects such as moles or ear holes as nostrils, as shown in Figure 9a. After changing to the cascaded segmentation model, the false detection phenomenon completely disappears, as shown in Figure 9b. This effectively improves the accuracy of the next step of cow breath detection, shows a strong anti-interference capability, and can be used in combination with cow identification technology to achieve multiobject cow breath detection.

3.2.2. Analysis of the Influence of Head Swing on the Results of RR Detecting

Cows, as living bodies, often produce head swings and other interference factors during the data collection process. To verify the robustness of this algorithm more accurately, in this study, the head swing movement of cows is analysed to clarify the effectiveness of this method.

There are three fundamental reasons for the influence of the cow’s head swing on the detection results. First, because the cow’s respiratory behaviour is weak, for the video-based respiratory behaviour monitoring method, the camera needs to be placed close to the cow, and the cow’s action (interference) will affect the accuracy of the monitoring results. Second, the temperature of the nostril area is unstable along with the swing movement of the cow head. Third, the head of the cow shakes violently, which leads to a reduction in the confidence in object extraction.

The average accuracy of breath detection of the 20-segment thermal imaging video of the No. 1386 cow was 92.57%, and the main reason for the low accuracy was that the head of the cow shook violently. This reduced the confidence of object extraction and the temperature data used to detect the RR, resulting in a significant decline in the accuracy of RR detection.

In the test data of five cows with 81 segments, except for No. 1386, the other four cows had motion interference, and the average accuracy of respiration detection was 95.08%, which was higher than the 92.57% accuracy of No. 1386. This was because the interference caused by head swing movement was not intense, global, or continuous, and the movement time was short. This did not affect the detection of breathing behaviour using nostril temperature changes.

In summary, in the case of cow head swing interference, the final detection accuracy is less affected, which proves that the algorithm in this study has good robustness and can accurately complete the detection of breathing behaviour. In a follow-up study, we can increase the proportion of head swing images in the training set so that the algorithm has better robustness when the cow’s head swings.

3.2.3. Analysis of the Impact of Waveform Noise on the Results of RR Detecting

The temperature extraction result of this algorithm is shown in Figure 6. By analysing the noise phenomenon, it is found that there are two main reasons for the noise: one is generated in the process of data acquisition; and the second, is due to nose detection missed, false alarms, nostril segmentation missed, and false detection.

To eliminate noise interference in this paper, a Fourier transform is used to analyse the signal spectrum, and then a Butterworth filter is used to smooth the filter.

The experimental results show that the noise analysis in this paper is correct and the noise suppression effect of object extraction is obvious. In contrast, the noise generated by data acquisition is a systematic error, which is inevitable and has little impact on the experimental results. In summary, the final detection accuracy is less affected by noise, which proves that the algorithm has good robustness.

3.3. Comparison with Other Detection Methods

Jorquera-Chavez et al. (2019) collected nonradiative infrared videos of cows’ faces via an infrared thermal camera and calculated their RR according to the changes in pixel intensity values in the nose region caused by respiratory airflow during breathing [30]. The correlation coefficient R between the respiratory frequency measured by this method and the results obtained by manual observation was 0.87. Lowe et al. (2019) completed the detection of respiratory behaviour by recording thermal fluctuations around the nostrils with an infrared thermal camera, and the test results showed that the correlation between this method and manual counting reached 0.93 [31]. Wu et al. (2020) proposed a combination of Deeplab V3+ and the PBVM action amplification algorithm to detect respiratory behaviour in standing cows with an accuracy of 93.04% [32].

In this study, IRT videos of dairy cows were recorded by an infrared thermal camera, and a cascade method of dairy cow respiratory status monitoring was proposed by combining YOLO v8 and Mask2Former to detect the nose first and then the nostrils. The accuracy of respiratory number detection in 81 thermal imaging videos of five dairy cows was 94.58%, and the correlation coefficient R was 0.95. The results show that the algorithm can detect the cow’s breathing behaviour well, providing a basis for the follow-up study of the multiobject cow’s breathing behaviour detection method.

3.4. Outlook

Although the method in this paper has achieved a reliable effect on the change in temperature flow in the nostril area of dairy cows, there are still some shortcomings. The research basis of this model is nose detection based on YOLO v8 and nostril area segmentation based on deep learning. The cow nostril object occupies fewer effective pixels in the video picture, and the breathing amplitude is smaller and more difficult to observe. In addition, the environment is changeable, and there is considerable interference, which aggravate the difficulty of cow breathing detection. In future studies, cameras with higher resolution can be utilized, and nostril features can be added to improve the accuracy of RR detection.

In the data acquisition process in this study, the camera is fixed, and there is a situation in which the cow’s nostril area is not in the acquisition area due to the swing of the cow’s head, which results in the loss of breathing fragments. In future research, a tracking camera can be used to develop a real-time image display and processing platform to achieve remote real-time acquisition and processing of cow video.

As experiments progressed, the respiratory behaviour of multiobject dairy cows can be explored based on the study of single-object dairy cows’ respiratory behaviour. However, dairy cows are easy to gather in natural scenes, and there are some problems in the detection of multiobject dairy cows’ respiratory behaviour, such as complex and changeable movements, many irrelevant interferences, as well as difficulty in detecting the RR of different dairy cows. The identity of a cow can be determined by a cow face recognition algorithm, and the position area of the cow object can be detected by an object detection algorithm. Different cow objects can be marked according to the position information, and the RR detecting of each cow object can be completed by developing a respiratory status monitoring algorithm. Multiobjective monitoring of the respiratory status of dairy cows can be achieved, and the application value of new technology and new algorithms for the welfare of dairy cows and the economic benefits of dairy farms can be improved.

This method has more advantages in the application prospect of large dairy farms, and the infrared thermal camera can be installed in the intelligent inspection robot, which has a higher utilization rate than the fixed camera position acquisition method. High utilization of instruments means lower costs. At the same time, it is more convenient to collect the facial data of cows than the abdominal data, and the requirements for the cooperation of cows are lower. This method will provide a reference for large-scale intelligent dairy farming. In the future, it may provide reference for automatic monitoring of heat stress in dairy cows and remote diagnosis of other diseases related to respiratory behavior.

4. Conclusions

According to the characteristics of temperature change in the nostril area when cows breathe, a cow respiratory status monitoring system based on IRT is proposed. YOLO v8 is proposed to detect the cow’s nose from IRT images without contact; the results showed that the AP₅₀ of nose detection reached 98.6%. Three kinds of instance segmentation models were used to realize nostril area, and the final experimental results show that Mask2Former has obvious advantages over Mask R-CNN and SOLOv2, AP₅₀ achieved 75.71%. The maximum accuracy of RR detection is 97.22%, the minimum accuracy of RR detection is 92.57%, the average accuracy of RR detection is 94.58%, and the correlation coefficient R was 0.95, higher than in previous studies on RR. The recognition effect was better than that of the combination of Deeplab V3+ and the PBVM action amplification algorithm to detect respiratory behaviour in standing cows, which remarkably improved the accuracy and efficiency of the detection of RR. Therefore, the method presented in this paper could promote the popularization and application of automatic detection of RR systems in large-scale dairy farming.