Recognition and monitoring of the feeding behavior of dairy cows based on video and TCS-YOLO model

doi:10.21203/rs.3.rs-1743573/v1

Download PDF

Article

Recognition and monitoring of the feeding behavior of dairy cows based on video and TCS-YOLO model

https://doi.org/10.21203/rs.3.rs-1743573/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Animal feeding behavior is an important animal welfare evaluation index. For an automatic monitoring of dairy cow feeding behaviors based on the deep learning technology, this paper proposes a quantity model for the recognition of dairy cow feeding behaviors and head movements. This model can automatically identify the feeding, chewing, and grass-bending behaviors of multiple dairy cows and track and quantify the feeding process and head movement trajectory of dairy cows in real time. This paper proposes a Transformer-Convolution Block Attention Module-Squeeze and Excitation-You Only Look Once (TCS-YOLO) model, which is an improved deep learning model based on You Only Look Once v5 (YOLOv5). By adding the transformer, Convolution Block Attention Module (CBAM), and Squeeze and Excitation (SE) enhancement modules, the feature extraction ability of the YOLOv5 model and the detection accuracy of the model are improved. It is further combined with the Deep Sort algorithm to track the trajectory of the heads of dairy cows when feeding. A total of 10,288 images were extracted from the top-view and front-view videos during the feeding process to construct a training set and a test set, and the TCS-YOLO model was tested and verified. The test results showed that the accuracy in detecting dairy cow feeding behaviors was 77.73% and 76.32%, the recall rate was 82.57% and 86.33%, and the mean average precision was 83.7% and 76.81%. Compared with the YOLOv5 model, the overall performance of the proposed model improved by 6–8%. It combined well with the Deep Sort target tracking algorithm, which could accurately track the feeding behaviors of dairy cows and suppress the change in their identity document. The research results can effectively solve the problems of poor recognition accuracy of feeding behaviors and insufficient feature extraction in complex breeding environments while providing an important reference guide for intelligent animal husbandry and precise breeding.

Dairy cows

Deep Sort

Feeding behavior recognition

Optimized YOLOv5 model

The unbroken stretch development of the interests, money, goods work in society has caused reaction by pushing a greater request for milk and milk-place products, which has greatly increased in rate the development of the milk-place farming industry¹. At the same time, the balance shortage between producing request and special to some science or trade level takes new problems and questions to current-day milk-place business managers. Dairy cows have a big and frequent daily feed intake, a lengthy average feeding period, and a high metabolic rate. Dairy cows' feeding habit is linked to their milk production, according to studies. Expert in the diseases of animals technicians often make observation of and carefully look at ill dairy cows through the getting of food to behavior of dairy cows, among them, the daily getting of food to number of times every unit time and daily food in-take of healthy cows are much larger than those of lame cows, and the mean time of single getting of food to of dairy cows with mastitis will drop^2,3. Based on the experimental exploration of a large number of research samples, it can be basically determined that the disease has a significant impact on the feeding behavior of dairy cows⁴. Therefore, it is very important for the farm to monitor and track the feeding behavior of dairy cows.

Accurately identifying the feeding behavior of dairy cows is the basis of the entire monitoring and tracking technology system of dairy cattle feeding behavior. Current studies have shown that real-time location system (RTLS)⁵ and ultra-wideband (UWB) system^6,7 can establish a link between the time spent in the feeding area and feeding behavior of dairy cows. However, this method has limitations, because the stop process does not determine whether the cow is eating effectively, so the accuracy of the feeding time data obtained by RTLS and UWB needs to be further verified. There are also some studies monitoring the chewing motion of dairy cows through the data changes collected by pressure sensors^8,9.While ensuring good monitoring performance, it also brings a lot of inconvenience. First, the use process is complicated, and secondly, the wearing process takes time and energy, which will cause discomfort to the cows and affect normal eating. According to the type of sensor, some teams use the inertial measurement unit (IMU) and the ear-hook sensor of the radio frequency identification system to determine the position of the cow, so as to achieve the purpose of monitoring feeding behavior¹⁰. But this approach also runs the risk of uncertainty between the range of motion of the cow and the determination of the effective feeding of the cow. As can be seen from this type of research, although some research results have been achieved in the monitoring of feeding behavior of dairy cows. However, the defects are still very obvious. They usually face the dilemma of high cost or inaccurate judgment on feeding behavior, and their applicability and generalization in the actual breeding environment of farms are not high. Therefore, it is necessary to further find an effective solution.

Deep learning has become the most cutting-edge and most successful solution in the field of object detection. Deep learning is one of the technical branches of machine learning. Most of deep learning is supervised learning. Compared with traditional unsupervised learning, supervised learning uses deep artificial neural networks to learn knowledge, which is more reliable. In the target monitoring algorithm, it can be mainly divided into two-stage algorithm and one-stage algorithm. Of course, there are more-order algorithms such as Cascade RCNN¹¹. Two-stage algorithms usually focus on finding out where objects appear in the first stage, and then focus on classifying proposal boxes in the second stage to find more precise positions, typical algorithms such as Fater RCNN. Two-stage algorithms are generally more accurate, but slower. The first-order algorithm combines the two stages of the two-order algorithm into one, and completes the positioning and identification of the target in one stage. The method is usually simpler and relies on excellent network experience such as feature fusion and Focal Loss. The speed is generally faster than the two-stage network, and the typical algorithm is YOLO¹².

The advantages of deep learning provide a broader prospect for the development of animal husbandry. In recent years, the application of deep learning in animal husbandry is not uncommon. Achour et al.¹³ used four CNN networks to discover the getting of food to behavior of dairy cows and individual dairy cows, but the model is relatively complex, and it is very dependent on the coordination performance between different models. Porto et al.¹⁴ were based on the Violae-Jones algorithm and set up multiple image acquisition cameras to make out the getting of food to and positioned behavior of dairy cows. On the basis of modifying the original CNN model, Bezen et al.¹⁵ created a monitoring system suitable for collecting the daily food intake of individual dairy cows by relying on RGB-D cameras. Yang et al.¹⁶ used a fully convolutional network to get out the head features of female pigs, and used the partly covering area between the head and the getting of food to area as spatial features to make out the getting of food to behavior of female pigs. From the perspective of preprocessing, Lao et al.¹⁷ developed a new algorithm for studying the feeding behavior of sows by using depth image data. Shelley et al.¹⁸ designed a 3D image analysis method to evaluate the changes of available materials in the feeding area before and after the cows eat, so as to further study the feeding behavior of cows. The above studies let see that it is possible to use deep learning algorithms to learning process the getting of food to behavior of dairy cows, However, there is no research to distinguish the feeding action of dairy cows and track the feeding process of dairy cows, so the research on the recognition and tracking of feeding behavior of dairy cows is necessary.

Due to the fast movements of dairy cows in the feeding process and the complex characteristics of feeding movements, it is easy to misdetect or miss the target. In view of the above problems, we design the YOLOv5 model based on the multi-head adaptive attention mechanism of the transformer module to strengthen the learning and extraction of the characteristics of dairy cow feeding behaviors, using the CBAM and SE modules to strengthen the characteristics of dairy cows and strengthen the classification. In this way, the accuracy of identification and detection of feeding behaviors can be enhanced.

2.1. Materials

2.1.1. Image acquisition

All animal handling procedures were approved by the Institutional Animal Care and Use Committee of Shandong Agricultural University (Approval Number: SADUA-2021-053), all methods were performed in accordance with the relevant guidelines and regulations. The dataset of dairy cow feeding behaviors in this study comes from Tai'an Jinlan Dairy Cows Breeding Company, Tai'an, Shandong, China. The company currently has more than 2,000 high-quality Holstein cattle. The No. 3 cowshed was selected as the experimental cowshed. There were 17 lactating dairy cows in the cowshed, aged between 1.5 and 2.5 years old and in good condition. The data collection time of the dairy cow feeding behavior experiment was from August 10, 2021, to September 5, 2021, a total of 27 days, from 8:00 to 12:00 and from 14:00 to 18:00 every day (feed supplementation started at 9:00–9:30 and 15:00–15:30 every day). The data collection was completed by a ZED2 binocular depth camera (the manufacturer: STEREOLABS company), and the maximum shooting angle of the ZED 2 binocular depth camera was 110° (H) × 70° (V) × 120° (D). However, the transmission frame rate under this pixel was only 15 frames; therefore, to ensure the stability and accuracy of the data, the monocular resolution was 1280×720 pixels, and the transmission frame rate under this pixel could reach a stable 30 frames. The height of the top of the cow pen was 1.35 m from the ground, and the width of the feed belt was approximately 0.8 m. To avoid interfering with normal feeding behaviors of dairy cows, in this paper, the camera was selected to be 1.75 m above the feeding area, 1.2 m in front, and 0.8 m high to collect feeding behaviors of dairy cows from different shooting directions.

In the process of image acquisition, a total of 20 groups of video data were collected, one group contained 4 videos, and two groups of videos were taken above and in front of the two time periods of 8:00–12:00 and 14:00–18:00. The ZED API was used to extract images frame by frame from the video to remove duplicates, blurs, ghosts and invalid images. A total of 10,288 images were obtained in the data set, and the feeding actions of 17 dairy cows were included in the data set. Among them, the number of front-facing images is 5,242, including 1,324 images containing one dairy cow, and 3,918 images containing multiple dairy cows; The number of images taken above is 5046, including 1263 images containing one dairy cow and 3783 images containing multiple dairy cows.

2.1.2. Dataset Labeling

In this paper, the open-source labeling tool LabelImg was used to manually label 10,288 original images of dairy cow feeding behaviors. The dataset was made by labeling and dividing the feeding behaviors of 17 different dairy cows, dividing the images into the top-view dataset and the front-view dataset. The upper-shooting dataset was divided into a training dataset with 4,320 images and a test dataset with 726 images according to the ratio of 1:6. The front-view dataset was divided into a training dataset with 4,484 images and a testing dataset with 758 images according to the ratio of 1:6. In normal feeding, multiple dairy cows may compete for a piece of feed, which would cause their heads to overlap when feeding. Therefore, the following labeling rules were formulated: 1) Do not label the occluded and incomplete dairy cow heads; 2) Label the adjacent dairy cow heads. LabelImg automatically generated the corresponding XML file every time a manual labeling of a dairy cow head was completed. The XML file recorded the coordinates of the upper-left corner and the lower-right corner of the labeling rectangle on the head of the dairy cow, that is, the feeding behavior, the length and width of the labeling rectangle, and the labeling category information. To improve the training efficiency and speed of the model, the annotation frame of a dairy cow head was normalized, as shown in Formulas (1)–(4).

$$x=\frac{{{x_{\hbox{max} }}+{x_{\hbox{min} }}}}{{2{R_w}}}$$

$$y=\frac{{{y_{\hbox{max} }}+{y_{\hbox{min} }}}}{{2{R_h}}}$$

$$w=\frac{{{x_{\hbox{max} }} - {x_{\hbox{min} }}}}{{{R_w}}}$$

$$h=\frac{{{y_{\hbox{max} }} - {y_{\hbox{min} }}}}{{{R_h}}}$$

where, x, y, w and h represent the coordinates of the center point of the rectangular box for the labeling of the feeding behavior of dairy cows after normalization and the width and height of the labeling box; x_max, x_min, y_max and y_min are the coordinates of the upper left corner and the lower right corner of the rectangle frame of the manually marked feeding behavior of dairy cows, respectively; R_w and R_h are the width and height of the dairy cows’ feeding behavior image, respectively.

In the breeding environment of a farm, dairy cows can eat only when they are in the feeding area and the head of the dairy cow is in contact with the feed. When the head of the feeding area is raised, they are basically in the process of chewing the feed and preparing to continue to eat. The dairy cows leave the feeding area when they have had enough feed. Therefore, the feeding behavior of dairy cows is divided into two parts: feeding and chewing. An example is shown in Fig. 1(a) - (f), which show the chewing, feeding, and grass arching behaviors of the dairy cows photographed from the front and the chewing, feeding, and grass arching behaviors of the dairy cows photographed from above.

Among them, the proportion of each behavior in the data set is shown in Table 1:

Table 1

Number of dairy cows' feeding behaviors marked in the dataset
Shooting direction of the dataset	Number of training datasets	Number of feeding behaviors	Number of chewing behaviors	Number of grass arching behaviors
Front	4484	5684	792	1613
Above	4320	6958	960	1946

2.2. Methods

The YOLO series model is a single-stage detection model, and the detection speed is faster than that of the two-stage model, but the corresponding accuracy is worse than that of the two-stage detection model. In the identification and tracking of the feeding behavior of dairy cows, due to the rapid feeding action of dairy cows, it is necessary to ensure not only fast detection speed, but also high detection accuracy. Therefore, based on the YOLOv5 model, we used the three model enhancement modules of Transformer、CBAM and SE to enhance the feature extraction ability of the model, and increased the accuracy of the recognition of the feeding behavior of dairy cows on the basis of ensuring the identification speed of 17 dairy cows and feeding behavior.

2.2.1. YOLOv5 model

Figure 2 shows the YOLOv5 model, which uses CSPDarknet as the backbone network and continues to use in the neck network. Compared with the YOLOv4 model, the Mosaic data enhancement and adaptive anchor frame structure were used at the input, the Focus structure was added to the backbone network, and the BottleneckCSP structure was applied to the neck network. This change allowed the YOLOv5 model to reduce the number of parameters used under the premise that the detection accuracy was consistent with the YOLOv4 model. In addition, YOLOv5 used the adaptive image scaling function, which could make the scaled image aspect ratio the same as the original image, ensuring that the target tracking would not affect the detection position pixel judgment due to the change in the image aspect ratio.

In testing, the detection speed of the YOLOv5 model was very fast, as it only took approximately 0.01 s to detect one frame of the image. However, the detection accuracy was not high, and the feature extraction of the model was not comprehensive. Therefore, based on YOLOv5, we used the model enhancement module to improve the detection performance of the model.

2.2.2. Transformer module

The transformer module was proposed by Ashish Vaswani et al.¹⁹. The internal structure is shown in Fig. 3, which is mainly composed of encoding components and decoding components. The encoding component consists of multiple encoders, the decoding component consists of multiple decoders, and the number of encoders corresponds to the number of decoders. The encoder is composed of a multi-head adaptive attention layer and a feed-forward neural network. Each sublayer has a residual connection around it and follows a "layer-normalization" structure. The decoder is composed of a masked multi-head adaptive attention layer, a multi-head adaptive attention layer and a fully connected feed-forward neural network. The adaptive attention layer can help the current node obtain the current key content that needs to be considered. Through the multi-head adaptive attention layer, the ability of the model to focus on different positions is expanded, and multiple subspaces are formed to allow the model to focus on different aspects of information, which can effectively combine the identity characteristics and the different behaviors of dairy cows when feeding to achieve the integration of high-dimensional global features.

The transformer module can parallelize the amount of computation to obtain the job done with the minimum sequence needed, reducing the distance between any two positions in the sequence to a constant due to the use of an adaptive attention mechanism. The adaptive attention layer can directly capture the global connection, while the number of operations required by the CNN network to calculate the association between two locations through convolution increases with the distance between the two locations. As a result, it is easier for the transformer module to learn more feature information.

2.2.3. CBAM module

The CBAM module²⁰ included two independent sub-modules, the channel attention module and the spatial attention module, which performed channel and spatial attention work, respectively, as shown in Fig. 4.

Channel attention needed to compress the spatial dimension of the input feature map. After average pooling and maximum pooling, the transformation result was obtained through the multilayer perceptron layer, and finally applied to the two channels and the channel-to-channel attention result through the sigmoid function. Spatial attention was a supplement to channel attention. The feature map output by channel attention was used as the input of spatial attention, and the dimension of the channel itself was reduced to obtain the results of average pooling and maximum pooling respectively. Then spliced into a feature map, and finally used a convolutional layer for learning.

By combining the channel attention module and the spatial attention module, the CBAM module saves parameters and computing power, and ensures the lightweight of the module, which can be directly integrated into the existing network model framework.

After the introduction of the CBAM module, the features of network identification cover more head features when feeding dairy cows, and the features representing head movements of dairy cows have become the key information of the model.

2.2.4. SE module

The full name of the SE module is Squeeze and Excitation²¹, and its workflow is shown in Fig. 5. The SE module first performed compression, compressed the feature map into a 1×1×C vector through global pooling, and then performed the excitation operation. Through the fully connected layer and activation function, channel reduction was performed to reduce the amount of calculation. Finally, the scaling operation was performed, and the feature map was scaled through a 1×1×C vector.

After passing through the SE module, the channel attention can be enhanced so that it has a global receptive field through the feature map of the SE module, which enhances the feature map extraction of the network and improves the network's classification of the target.

2.2.5. Establishment of TCS-YOLO model

Since the premise of target tracking required an accurate extraction of model features, the transformer module was added on the basis of the high-efficiency network YOLOv5, and the adaptive attention mechanism of the transformer module was used to achieve an accurate grasp and judgment of the characteristics of the feeding behaviors of dairy cows. The CBAM module was added after the focus layer so that the model's attention was focused on the characteristics of dairy cow feeding behaviors, and the SE module was used at the end of the backbone network to improve the global receptive field of the feature map and the classification ability of the model. For the model prediction module, CIOU²² was used instead of GIOU (the ratio of the intersection and union of the predicted bounding box and the ground truth bounding box). Considering the influence of scale, distance, penalty term and the overlap rate between anchor boxes on the loss function, the target frame regression became more stable. The formula is as follows:

$$CIOU=IOU - \frac{{{\rho ^2}(b,{b^{gt}})}}{{{c^2}}} - \alpha v$$

$$\alpha =\frac{v}{{1 - IOU+v}}$$

$$v=\frac{4}{{{\pi ^2}}}{(\arctan \frac{{{w^{gt}}}}{{{h^{gt}}}} - \arctan \frac{w}{h})^2}$$

$$Los{s_{CIOU}}=1 - IOU+\frac{{{\rho ^2}(b,{b^{gt}})}}{{{c^2}}}+\alpha v$$

where c represents the diagonal distance of the smallest enclosing region of the bounding box that contains both the predicted bounding box and the classification accuracy of the supervised learning training set; w, h, w^gt, h^gt respectively represent the width and height of the current detection frame and the width and height of the prediction frame of the next frame;αis a weight function; v represents the similarity of aspect ratio; ${\rho ^2}(b,{b^{gt}})$ represents the Euclidean distance between the prediction center point and the real frame; and finally, 1-CIOU is used to obtain the corresponding loss.

The workflow of the TCS-YOLO model is shown in Fig. 6. When the input image entered the model, the focus module was first used to perform the slicing operation, and the W and H information was concentrated into the channel space. Next, through the CBAM module, the CSPDarknet network was entered to perform a series of convolution operations to obtain a high-dimensional feature map, and then the transformer module was entered through the SPP structure. The multi-head adaptive attention structure was used to perform parallel learning of features, and the results were input to the SE module to enhance the global receptive field of the model. After that, the Feature Pyramid Network²³ and the Path Enhancement Network²⁴ were entered, connections with the backbone network were established through upsampling and tensor splicing, etc., and the feature information of the feature map was enriched. Finally, the feature map was sent to the YOLO detection module to obtain the detection result.

2.2.6. Deep Sort Algorithm

The Deep Sort algorithm²⁵ is often used in multitarget tracking. Its workflow is shown in Fig. 7. The individual identity of a dairy cow and the corresponding feeding action were calculated through TCS-YOLO, and then the position of the object in the next frame was predicted by the Kalman filter. The predicted position was compared with the actual detection position of the next frame, and the IOU was calculated to obtain the similarity of two adjacent targets. Finally, the corresponding identity documents of adjacent frames were obtained through the Hungarian algorithm, which tracked the identities and feeding actions of dairy cows.

The Deep Sort algorithm used the comparison and confirmation of the front and back features and added cascade matching and new trajectory confirmation. As shown in Fig. 8, after the target detection network obtained the unique feature of a dairy cow, it strengthened the feature information of the dairy cow through prediction, observation, update, etc. Therefore, the dairy cow identity document detected by the Deep Sort algorithm was not easily changed, so dairy cow feeding behaviors could be tracked and detected more accurately.

3.1. Evaluation indicators

In order to accurately evaluate the performance of the model, we use four performance evaluation indicators commonly used in object detection algorithms: precision, recall, mean average precision (mAP) and F1-score to test the performance of the model proposed in this paper.

In terms of precision and recall, there are 4 states after predicting a test sample: true positive (T_p), false positive (F_p), true negative (T_N), and false negative (F_N). Defined as follows:

$$precision=\frac{{{T_p}}}{{{T_P}+{F_P}}}$$

$$recall=\frac{{{T_p}}}{{{T_P}+{F_N}}}$$

Both mAP value and F1-score can be used to evaluate the performance of the object detection model. mAP and F1-score are defined as follows:

$$AP=\int_{0}^{1} {precision \cdot recall{\text{ }}drecall}$$

$$mAP=\frac{1}{m}\sum\limits_{m} {AP}$$

$${F_1}=2 \times \frac{{precision \cdot recall}}{{precision+recall}}$$

In addition, the change trend of the loss curve of the model can also be used to evaluate the pros and cons of the model. The loss curve fitting speed is fast, the fitting degree is good, and the final loss value is low, which represents the stronger the performance of the model.

3.2 Experimental configuration and hyperparameter settings

The hardware platform configuration for training and testing was as follows: Intel(R) Xeon(R) W-2135 [email protected] GHz processor, 32 GB memory, 11 GB GeForce RTX 2080 Ti graphics card, and 500 G solid-state drive. The platform operating system was a 64-bit Ubuntu18.04 operating system, accessed through an Ubuntu18.04 terminal. The acceleration environment CUDA, Cudnn and third-party libraries OpenCV and PyTorch were called during the program run. The CUDA version was 10.2.89, the Cudnn version was one based on 7.6.5 under CUDA10.2, the OpenCV version was 4.4.0, and the PyTorch version was 1.7.1.

In the experimental parameters, the size of the batch size affects the training time. The batch size set in the parameters is the total batch size of all GPUs. The larger the batch size is, the shorter the training time for one iteration, as shown in Fig. 9. After the pretraining test, it was found that when the batch size was set greater than 16, the test terminal would show a CUDAError: out of memory. Therefore, a reasonable setting value of the batch size was 16. For input images, commonly used sizes for YOLOv5 are 416×416, 520×520, are 608×608. The input image size will affect the memory occupied by training; that is, the larger the input, the more the memory is occupied. Therefore, the input image was selected to be 416×416 pixels.

The learning rate and weight decay rate are two hyperparameters in deep learning. The learning rate is one of the important factors in whether and when the loss curve converges to the minimum. When the learning rate is set too large, the loss curve cannot converge; when the learning rate is set too small, the training time will increase evidently. By adjusting different learning rates for pretraining, it was found that when the learning rates were 0.1 and 0.01. As shown in the loss curve in Fig. 10, the loss value decreased and then rose and failed to converge. When the learning rate was 0.001, the loss curve converged well; there was no explosive growth, which met the expected convergence requirements. Therefore, the initial learning rate should be set to 0.001. The weight decay rate is equivalent to the L2 norm regularization. By adding a penalty term to the loss function, the purpose is to attenuate the weight to a smaller value to reduce the problem of model over fitting. The value should be close to 0. The model defaulted to 0.0005 to meet the requirements; therefore, the weight decay rate was set to 0.0005.

The maximum epoch of training is related to the training time and training effect, and one epoch is the time it takes for all the training sets to be trained once. The effect of iterative algebra was tested under the condition of an initial value of 0.001 and a weight decay rate of 0.0005. As shown in Fig. 11, when the number of iterations was 10, the loss value was 0.035, and the loss curve under fitted. When the number of iterations was 50, the loss value was 0.027, and the loss curve tended to be flat but still under fitted. When the number of iterations was 80, the loss value was 0.026, and the loss curve was almost flat with no obvious change, which showed that the loss curve reached the perfect fitting state. To ensure that the loss curve reached the perfect fitting state, the number of test iterations was set to 100 times. The loss curve had almost no change, and the loss value was 0.026, indicating that the model had reached the perfect fitting. We continued to test the case where the iteration algebra was 110 times. The program automatically terminated when the iteration reached 103 times, indicating that the loss curve had been perfectly fitted. Therefore, the iteration algebra should be set to 100 times.

3.3 Object tracking algorithm detection effect

To verify the reliability of the dairy cow foraging behavior object tracking algorithm, this paper used the dairy cow foraging behavior dataset shot from the above and from the front to test the algorithm.

Figure 12 shows the object tracking effect from the dairy cow foraging behavior dataset shot from the front. As seen in Fig. 12(a) and (b), the identity document of the dairy cow did not change when the foraging behavior of No. 1 changed from foraging to chewing, and the head movement trajectory of the dairy cow continued to be tracked and plotted. As seen in Fig. 12(c) and (d), the identity document of dairy cow No. 2 disappeared when dairy cow No. 1 covered dairy cow No. 2, but when dairy cow No. 2 returned to the detection state, the identity document of dairy cow No. 2 did not change. The experimental results above showed that the object detection network extracted sufficient features for dairy cows so that tracking according to the features of the object would not cause the identity document to change during object tracking.

Figure 13 shows the object tracking effect from the dairy cow foraging behavior dataset shot from above. Similar to the dataset shot from the front, there was no change in the identity document of the dairy cows due to shading and behavioral changes during foraging.

The analysis above showed that the combined object tracking algorithm of TCS-YOLO and Deep Sort could accurately track the dairy cow head movements during foraging and plot the dairy cow head movement trajectory to accurately determine the dairy cow's foraging behavior. By exporting the data of the movement points of the dairy cow, the time of the entire foraging process of the dairy cow and the total amount of movement of the dairy cow during the foraging process could be calculated. As shown in Fig. 14, the top-left vertex of the dairy cow foraging behavior tracking video was taken as the origin, and the X-axis and Y-axis were led in the 0° and − 90° directions. The value on the coordinate axis indicates the distance of the pixel point from the origin. By detecting the video with TCS-YOLO, the top-left vertex position of the dairy cows’ Bounding BOX and the width and height of Bounding BOX were obtained, the center point position of the dairy cow head movement was calculated, and the dairy cow head movement trajectory was drawn. Since the model tracked the video frame by frame and the video frame rate was 30 f/s, the total number of frames detected by the model was used to calculate the total time the dairy cow spent feeding. The total amount of movement could be calculated from the distance between the current frame and the center of the next frame. The final total amount of dairy cow movement was obtained by converting the ratio of the distance between pixels in the video to the distance of the actual object.

3.4 Experimental analysis of object detection model

3.4.1 The effect of shot angle on the detection results

This study was conducted to collect the dairy cow foraging dataset from two angles, from the front and the above, to find the best angle to perform edge computing. The comparison between models was conducted in terms of precision, recall, F1-score, loss curve, and mAP.

Figure 15(a) shows the training precision curve for the dairy cow foraging behavior dataset shot from the front (hereafter referred to as the front view), and Fig. 15(B) shows the training precision curve for the dairy cow foraging behavior dataset shot from the above (hereafter referred to as the top view). The precision in the front view was 0.9% higher than the precision in the top view. The precision was calculated as T_P/(T_P+F_P). The features for the top view were not as abundant as the features for the front view. As a result, the F_P for the top view was higher than the F_P for the front view, therefore the precision for the top view less than the precision for the front view.

The recall curves are shown in Fig. 16,. Figure 16 (a) shows the recall curve for the front-view dataset, and Fig. 16 (b) shows the recall curve for the top-view dataset. The front-view recall curves showed similar trends and similar curve values for the foraging category curve and the chewing category curve, which made the foraging category recall and chewing category recall essentially equivalent to the average recall. The top-view recall curve showed that recall for the foraging category was 10%-20% higher than recall for the chewing category. In the analysis that followed, it was found that the feature distinction between foraging and chewing for the top view was not as obvious as the feature distinction for the front view, which made the number of F_Ns in chewing category higher for the top-view dataset, resulting in low recall for the top-view chewing category.

The F1-score curves are shown in Fig. 17, Fig. 17 (a) shows the F1-score curve for the front-view dataset, and Fig. 17 (b) shows the F1-score curve for the top-view dataset. The consideration of both precision and recall in the F1-score made the final value of the front-view F1-score curve 4% higher than the final value of the top-view F1-score curve.

The loss curves and the mAP curves are shown in Fig. 18. Figure 18 (a) shows the loss curves and mAP curves for the front-view dataset, and Fig. 18 (b) shows the loss curves and mAP curves for the top-view dataset. From Fig. 18, both the front-view and the top-view training dataset had perfectly fit loss curves. For the validation dataset loss curves, the front-view cls_loss was fitted slower in the early stage compared to the top view, while the final cls_loss had a perfect fit. The reason was that more features in the front view required a longer learning time, which made the fit slower. From the results, box_loss and cls_loss in the front view and the top view were the same, while obj_loss in the front view was 0.002 lower than obj_loss in the top view. The factor affecting the obj_loss was the confidence level, which could judge the detection accuracy in the front view to be higher than that in the top view. From the mAP curve, it can be seen that the final mAP_0.5 and mAP_0.5:0.95 in the front view were 5% and 12% higher than the top view, while the trend of the mAP curve in the front view changed more slowly compared to the top view, and the same trend as the loss curve and the final value of mAP0.5 in the front view was in accordance with the test results.

From the above analysis, the dairy cow head features collected from the front view were more abundant, which made the training results of the front-view dataset better than the training results of the top-view dataset. Therefore, the dairy cow foraging behavior dataset shot from the front was better suited to edge computing than the dataset shot from the above.

3.4.2 Ablation test

To verify the effectiveness of the improved TCS-YOLO method proposed in this paper on model performance, this paper conducted model ablation tests by using different modules based on the YOLOv5 model. This was tested on the dairy cow foraging behavior dataset shot from the front and the dataset shot from above, as shown in Table 2 and Table 3. The experimental data were analyzed to reflect the effects of different improvement methods in the TCS-YOLO model on the performance in the detection.

Table 2

Effect of each module on the model from the front shot dairy cows foraging dataset
YOLOv5	Transformer	CBAM	SE	Precision (%)	Recall (%)	mAP (%)
√				69.76	75.72	75.9
√	√			70.56	77.56	77.53
√		√		67.14	81.76	77.82
√			√	69.75	72.23	73.93
√	√	√		69.40	85.65	76.64
√	√		√	67.54	75.80	75.76
√		√	√	64.32	86.77	77.66
√	√	√	√	77.73	82.57	83.7

Table 3

Effect of each module on the model from the above shot dairy cows foraging dataset
YOLOv5	Transformer	CBAM	SE	Precision (%)	Recall (%)	mAP (%)
√				68.29	77.16	67.23
√	√			70.93	82.35	72.36
√		√		73.05	80.5	73.0
√			√	69.5	80.0	67.63
√	√	√		73.54	85.92	73.82
√	√		√	73.24	84.41	72.59
√		√	√	71.45	82.60	69.47
√	√	√	√	76.32	86.33	76.81

From the tables, it can be seen that the transformer module improved the performance of YOLOv5 the most in the single module improvement YOLOv5 network, both in the front-view dataset and the top-view dataset, and the precision, the recall, and the mAP improved. In the front-view dataset, the precision improved by 0.8%, recall improved by 1.84%, and mAP improved by 1.63%; in the top-view dataset, the precision improved by 2.64%, recall improved by 5.91%, and mAP improved by 5.13%. When the CBAM module was used alone, the recall and mAP of the model improved significantly. The recall improved by 6.04% and the mAP improved by 1.92% in the front-view dataset, and the recall improved by 3.36% and the mAP improved by 5.77% of the top-view dataset. The precision of the front-view dataset decreased by 2.62% after using the CBAM module, and the same of the top-view dataset improved by 4.67%. This paper considered that the dairy cow foraging behavior dataset shot from the front was more abundant in dairy cow facial features, and the extraction of features by a single attention layer was not sufficient, which made the lifting effect worse. In contrast, the dairy cow foraging behavior dataset shot from the above had relatively few dairy cow facial features, and it was relatively easy for the model to extract features. The same situation occurred in the SE module. In the front-view dataset, the recall decreased by 2.69% and the mAP decreased by 1.97% after adding the SE module, while the precision and recall increased in the top-view dataset, with the precision increasing by 1.21%, the recall increasing by 2.84%, and the mAP remaining essentially invariant. In the double-module improved YOLOv5, the network including the transformer module improved compared to both of the original single-module improved networks. When increasing the double-module CBAM module and SE module, the precision of the front-view dataset was reduced by 5.44% compared to that of the YOLOv5 model, and the improvement of the top-view dataset was relatively small. Increasing the transformer module could have a large improvement on the model. When the three modules were used to improve YOLOv5, the precision, the recall, and the mAP values improved significantly. In the front-view dataset, the precision, the recall, and the mAP improved by 7.97%, 6.85%, and 7.80%, respectively, while in the top-view dataset, the precision, the recall, and the mAP improved by 8.03%, 9.17%, and 9.58%, respectively.

This analysis showed that the transformer module contributed the most to the model among the improved modules, the TCS-YOLO model with the addition of three modules had a greater performance improvement compared to the YOLOv5 model, and the feature extraction of dairy cow foraging behaviors was more adequate.

4.1 Object detection model

4.1.1 Module contribution

As object detection is usually based on edge computing or cloud computing for animal detection²⁶, the size of the model is not demanding. This paper was based on the YOLOv5 object detection model, which adds three model enhancement modules: transformer, CBAM, and SE. From Tables 2 and 3, the transformer module has been shown to have the largest contribution to model enhancement. When the transformer module was used for the YOLOv5 model, the precision, recall, and mAP improved by 1–2% when training with the dataset shot from the front, and the precision improved by 2.64%, recall by 5.91%, and mAP by 5.13% when training with the dataset shot from above. When the CBAM module or the SE module was used for YOLOv5, some parameters of the model were decreased. The analysis showed that the multi-head adaptive attention mechanism in the transformer module could help the model extract the object features effectively and improve the detection performance. The SE module had the smallest performance improvement of the three enhancement modules. When the SE module was used for the YOLOv5 model, the recall and mAP values of the model decreased by 2.69% and 1.97%, respectively, when training with the dataset shot from the front. Since the main function of the SE module was to improve the classification performance of the model, the SE module did not work well when the model did not extract the object features sufficiently. When the transformer module, CBAM module, and SE module were used simultaneously for the YOLOv5 model, through the multi-head adaptive attention mechanism of the transformer module and the spatial and channel attention mechanism of the CBAM module to fully extract the object features, the classification function of the SE module could be effectively exercised, significantly improving the performance of the model. After the object features were fully extracted by the multi-head adaptive attention mechanism of the Transformer module and the spatial and channel attention mechanism of the CBAM module, the classification function of the SE module could be effectively performed, which led to a significant improvement in the performance of the model. For the front-view dataset, the precision, recall, and mAP improved by 7.97%, 6.85%, and 7.80%, respectively, while for the top-view dataset, the precision, recall, and mAP improved by 8.03%, 9.17%, and 9.58%, respectively.

4.1.2 Effect of object detection

To visualize the improvement of the TCS-YOLO model compared to YOLOv5, TCS-YOLO and YOLOv5 were tested using a test dataset of dairy cow foraging behaviors shot from the front and from the above. The detection result of TCS-YOLO is shown in Fig. 19 (a), and the detection result of YOLOv5 is shown in Fig. 19 (b). From the detection results in Fig. 19, the confidence level of chewing and foraging actions improved by 6–10% in the dataset shot from the front, which was a significant improvement. The confidence level of chewing actions improved by 9% and the same of foraging actions improved by 1% in the dataset shot from the above, but YOLOv5 missed a detection. It was evident that the detection was better for the shot from the front than for the shot from above. The reason was that dairy cow head features in the dataset shot from the front were abundant, and the feature extraction ability of TCS-YOLO was enhanced, which improved the detection in the dataset shot from the front. In contrast, there were fewer dairy cow head features in the dataset shot from the above, and the feature extraction effect was completely exerted, which resulted in worse detection than in the dataset shot from the front. In the action category, the confidence level of chewing was higher than the confidence level of foraging. The results of the precision images obtained in Fig. 15 were consistent. The reason was that the head of the dairy cow was raised during the chewing action, which made the features of the dairy cow head more prominent. While the head of the dairy cow was in direct contact with the feed during foraging, some of the feed features were also extracted, and thus, the confidence level was reduced.

4.1.3 Noise Effect

Because of the feeding environment of dairy farms, the head of a dairy cow keeps moving when foraging, unavoidably raising dust or part of the feed, which may affect the effectiveness of the lens video capture. To simulate the dusty environment of the farm, this paper imposed different levels of noise points on the images to detect the effect of noise on the model. This was tested using datasets shot from the front and the above with noise imposed. Figure 20 (a) shows the detection without noise points imposed, Fig. 20 (b) shows the detection with 1,000 noise points imposed, Fig. 20(c) shows the detection with 3,000 noise points imposed, and Fig. 20 (d) shows the detection with 5,000 noise points imposed. From the results, it could be seen that after imposing 1,000, 3,000, and 5,000 noise points, the detection results did not change significantly, and the confidence fluctuations were all in the range of 1–2%, within the range of normal fluctuation. It could be inferred from the test results that dust from the farm feeding environment would have no effect on the model.

4.2 Analysis of classification detection of foraging behavior

This paper divided dairy cow foraging behaviors into two parts, foraging and chewing, while the foraging behavior could be divided into two parts, foraging and arching grass. The following analysis showed the effect of the dairy cow foraging behavior classification on the detection results in terms of precision, recall, and PR curves for each category. As the dataset shot from the front and the dataset shot from the above had the same trend of data growth during training, the dataset shot from the above was used only for training.

The precision curves are shown in Fig. 21. Figure 21 (a) shows the precision curves for classifying dairy cow foraging behaviors into foraging and chewing. DTCS is the precision curve for the foraging category. TTJJ is the precision curve for the chewing category, and all classes are the average precision curve for each category. Figure 21 (b) shows the precision curves for classifying dairy cow foraging behaviors into foraging, arching grass, and chewing. DTCS is the precision curve for the foraging category. PTGC is the precision curve for the arching grass category. TTJJ is the precision curve for the chewing category, and all classes are the average precision curve for each category. The experimental results showed that when the arching grass category was added to the classification, the precision curve of the chewing category was not affected. When the confidence level was between 50% and 70%, the precision of the chewing category was stable between 85% and 90%. However, the foraging category precision curves showed a decrease, with values decreasing by 5% on average, and the relatively low precision of only 20% for the arching grass category caused the average precision curve values to decrease by 6%. The reason was the high similarity between the characteristics of both the arching grass behavior and the foraging behavior, and the arching grass behavior was generally completed in an extremely short time. The model easily determined the arching grass behavior and the foraging behavior as the same behavior; therefore, the precision of the model in detecting the arching grass behavior was not high, and the precision of the foraging behavior was reduced.

The recall curves are shown in Fig. 22. Figure 22 (a) shows the recall curves for classifying dairy cow foraging behaviors into foraging and chewing. DTCS is the recall curve for the foraging category. TTJJ is the recall curve for the chewing category, and all classes are the average recall curve for each category. Figure 22 (b) shows the recall curves for classifying dairy cow foraging behaviors into foraging, arching grass, and chewing. DTCS is the recall curve for the foraging category. PTGC is the recall curve for the arching grass category. TTJJ is the recall curve for the chewing category, and all classes are the average recall curve for each category. The experimental results showed that when the arching grass category was added, the average recall curve showed a significant decrease. When the confidence level was between 6% and 40%, the value of the average recall decreased by 20%, and the arching grass category had a very low recall of 6%. However, the foraging and chewing categories were not changed significantly, as shown in Fig. 22 (b). The reason was that the recall calculation formula was T_P/(T_P+F_N), and F_N was the relevant but undetected type. For the fixed dataset, F_N was certain, so the increase in the arching grass category had no effect on the foraging category and chewing category. The arching grass category had low detection precision, which led to a higher F_N of the arching grass category and made the recall of the arching grass category low.

The PR curves are shown in Fig. 23. Figure 23 (a) shows the PR curves for classifying dairy cow foraging behaviors into foraging and chewing. DTCS is the PR curve for the foraging category. TTJJ is the PR curve for the chewing category, and all classes are the average PR curve for each category. Figure 22 (b) shows the PR curves for classifying dairy cow foraging behaviors into foraging, arching grass, and chewing. DTCS is the PR curve for the foraging category. PTGC is the PR curve for the arching grass category. TTJJ is the PR curve for the chewing category, and all classes are the average PR curve for each category. In Fig. 23, the PR curve has the recall in the vertical coordinate and the precision in the horizontal coordinate. When the arching grass category was added, there was no change in the PR curves of the foraging and chewing categories. The foraging category PR curve decreased by 2%, the chewing category PR curve improved by 2%, and the average PR curve decreased by 25%, which was a serious decrease. Moreover, the value of the PR curve for the arching grass category was very low at only 16%. Since the PR curve was directly related to precision and recall, both precision and recall were low for the arching grass category, which made the value of the PR curve for the arching grass category very low.

From the above analysis, it was found that when dairy cow foraging behaviors were divided into two categories, foraging and chewing, the detection of each category was good. The additional category of arching grass decreased the precision of the foraging category and made the average detection of each category lower. As a result of sampling the collected video data of dairy cow foraging behaviors, the duration of the arching behaviors was observed to account for approximately 1/15 − 1/20 of the total duration of the foraging process. Through the above analysis, this paper chose not to use the arching grass category for testing.

4.3 Analysis of limitations

In this paper, the proposed model algorithm is limited by the ablation experiments for each detection module only, without comparison with other object detection and object tracking models, which makes the model lack comparison data. The model detects arching grass behaviors in dairy cow foraging behaviors poorly, easily misclassifying it as foraging action. The reason is the high similarity between the characteristics of the arching grass behavior and the foraging behavior, and an arching grass behavior is generally completed in an extremely short time. The model easily determines arching grass and foraging as the same behavior; therefore, the model has not produced the expected results in the detection of arching grass behaviors.

This paper investigates the foraging behavior of dairy cows in the foraging area in the farm feeding environment. In this research, we proposed the TCS-YOLO deep learning algorithm based on the YOLOv5 algorithm. For the detection results of the dairy cows foraging behavior dataset taken from the previous side, the precision, recall and mAP of TCS-YOLO were 77.73%, 82.57% and 83.7%. Compared with the YOLOv5 algorithm, it has been greatly improved. It not only has good anti-noise ability, but also improves the accuracy of the judgment of the feeding behavior of dairy cows. This paper proposes the TCS-YOLO model and combines the Deep Sort algorithm to perform simultaneous object tracking of foraging behaviors of multiple dairy cows and to map the trajectory of the dairy cow head, achieving the expected goal of object detection and tracking dairy cow foraging behaviors, which makes up for the lack of research on the detection and tracking of feeding behavior of dairy cows.

From the detection results of TCS-YOLO, it can be seen that the mAP value of the cow feeding behavior data set photographed in the front is 6.89% higher than that of the cow feeding behavior data set photographed above, indicating that this algorithm has a better detection effect. In the future research, the data is directly collected from the front camera angle, so as to avoid the trouble of repeatedly selecting the camera angle.

In the following research, the feed depth image information collected by ZED binocular camera can be used to analyze the feed consumption of dairy cows in a certain period of time, and then realize the calculation of feed intake of dairy cows, and then combine the research results of this paper to finally realize the design and development of an automatic monitoring system for dairy cows' foraging behavior.

Data Availability

The data used to support the findings of this study are included in the article.

Acknowledgments

This study was partly supported by NATIONAL NATURAL SCIENCE FOUNDATION OF CHINA, grant number 32072786, and CHINESE NATIONAL KEY RESEARCH & DEVELOPMENT PLAN, grant number 2016YFD0700205.

Conflicts of Interest:

The authors declare no conflict of interest.

Bareille, N., Beaudeau, F., Billon, S., Robert, A. & Faverdin, P. Effects of health disorders on feed intake and milk production in dairy cows. Livestock Production Science 83, 53-62, doi:10.1016/s0301-6226(03)00040-x (2003).
Fogsgaard, K. K., Rontved, C. M., Sorensen, P. & Herskin, M. S. Sickness behavior in dairy cows during Escherichia coli mastitis. J Dairy Sci 95, 630-638, doi:10.3168/jds.2011-4350 (2012).
Fogsgaard, K. K., Bennedsgaard, T. W. & Herskin, M. S. Behavioral changes in freestall-housed dairy cows with naturally occurring clinical mastitis. J Dairy Sci 98, 1730-1738, doi:10.3168/jds.2014-8347 (2015).
Thorup, V. M. et al. Lameness Affects Cow Feeding But Not Rumination Behavior as Characterized from Sensor Data. Front Vet Sci 3, 37, doi:10.3389/fvets.2016.00037 (2016).
Shane, D. D., White, B. J., Larson, R. L., Amrine, D. E. & Kramer, J. L. Probabilities of cattle participating in eating and drinking behavior when located at feeding and watering locations by a real time location system. Computers and Electronics in Agriculture 127, 460-466, doi:10.1016/j.compag.2016.07.005 (2016).
Pastell, M. & Frondelius, L. A hidden Markov model to estimate the time dairy cows spend in feeder based on indoor positioning data. Computers and Electronics in Agriculture 152, 182-185, doi:10.1016/j.compag.2018.07.005 (2018).
Porto, S. M. C., Arcidiacono, C., Giummarra, A., Anguzza, U. & Cascone, G. Localisation and identification performances of a real-time location system based on ultra wide band technology for monitoring and tracking dairy cow behaviour in a semi-open free-stall barn. Computers and Electronics in Agriculture 108, 221-229, doi:10.1016/j.compag.2014.08.001 (2014).
Li, W., Ji, Z., Wang, L., Sun, C. & Yang, X. Automatic individual identification of Holstein dairy cows using tailhead images. Computers and Electronics in Agriculture 142, 622-631, doi:10.1016/j.compag.2017.10.029 (2017).
Zehner, N., Umstätter, C., Niederhauser, J. J. & Schick, M. System specification and validation of a noseband pressure sensor for measurement of ruminating and eating behavior in stable-fed cows. Computers and Electronics in Agriculture 136, 31-41, doi:10.1016/j.compag.2017.02.021 (2017).
Aoughlis, S., Saddaoui, R., Achour, B. & Laghrouche, M. Dairy cows' localisation and feeding behaviour monitoring using a combination of IMU and RFID network. International Journal of Sensor Networks 37, doi:10.1504/ijsnet.2021.117962 (2021).
Ren, S., He, K. & Girshick, R. Faster R-CNN: towards real-time object detection with egion proposal networks. International Conference on Neural Information rocessing Systems. (2015).
Redmon, J., Divvala, S., Girshick, R. & Farhadi, A. You Only Look Once: Unified, Real-Time Object Detection. . 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). doi:doi:10.1109/cvpr.2016.91 (2016).
Achour, B., Belkadi, M., Filali, I., Laghrouche, M. & Lahdir, M. Image analysis for individual identification and feeding behaviour monitoring of dairy cows based on Convolutional Neural Networks (CNN). Biosystems Engineering 198, 31-49, doi:10.1016/j.biosystemseng.2020.07.019 (2020).
Porto, S. M. C., Arcidiacono, C., Anguzza, U. & Cascone, G. The automatic detection of dairy cow feeding and standing behaviours in free-stall barns by a computer vision-based system. Biosystems Engineering 133, 46-55, doi:10.1016/j.biosystemseng.2015.02.012 (2015).
Bezen, R., Edan, Y. & Halachmi, I. Computer vision system for measuring individual cow feed intake using RGB-D camera and deep learning algorithms. Computers and Electronics in Agriculture 172, doi:10.1016/j.compag.2020.105345 (2020).
Yang, A. et al. An automatic recognition framework for sow daily behaviours based on motion and image analyses. Biosystems Engineering 192, 56-71, doi:10.1016/j.biosystemseng.2020.01.016 (2020).
Lao, F. et al. Automatic recognition of lactating sow behaviors through depth image processing. Computers and Electronics in Agriculture 125, 56-62, doi:10.1016/j.compag.2016.04.026 (2016).
Shelley, A. N., Lau, D. L., Stone, A. E. & Bewley, J. M. Short communication: Measuring feed volume and weight by machine vision. J Dairy Sci 99, 386-391, doi:10.3168/jds.2014-8964 (2016).
Ashish Vaswani, N. S., Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, Illia Polosukhin. Attention Is All You Need. (2017).
Sanghyun Woo, J. P., Joon-Young Lee, In So Kweon. CBAM: Convolutional Block Attention Module. (2018).
Hu, J., Shen, L., Albanie, S., Sun, G. & Wu, E. Squeeze-and-Excitation Networks. IEEE Trans Pattern Anal Mach Intell 42, 2011-2023, doi:10.1109/TPAMI.2019.2913372 (2020).
Zheng, Z. et al. Distance-IoU Loss: Faster and Better Learning for Bounding Box Regression. Proceedings of the AAAI Conference on Artificial Intelligence 34, 12993-13000, doi:10.1609/aaai.v34i07.6999 (2020).
Lin, T.-Y., Dollar, P., Girshick, R., He, K., Hariharan, B., Belongie, S. Feature Pyramid Networks for Object Detection. doi:10.1109/CVPR.2017.106 (2017).
Liu, S., Qi, L., Qin, H., Shi, J., Jia, J. Path Aggregation Network for Instance Segmentation. doi:doi:10.1109/cvpr.2018.00913 (2018).
Wojke, N., Bewley, A., Paulus, D. Simple Online and Realtime Tracking with a Deep Association Metric. (2017).
Shen, W. et al. Automatic recognition method of cow ruminating behaviour based on edge computing. Computers and Electronics in Agriculture 191, doi:10.1016/j.compag.2021.106495 (2021).

No competing interests reported.

Download PDF

Version 1

posted

You are reading this latest preprint version

Recognition and monitoring of the feeding behavior of dairy cows based on video and TCS-YOLO model

Status:

Version 1

Abstract

Figures

1. Introduction

2. Materials And Methods

2.1. Materials

2.1.1. Image acquisition

2.1.2. Dataset Labeling

2.2. Methods

2.2.1. YOLOv5 model

2.2.2. Transformer module

2.2.3. CBAM module

2.2.4. SE module

2.2.5. Establishment of TCS-YOLO model

2.2.6. Deep Sort Algorithm

3. Results And Analysis

3.1. Evaluation indicators

3.2 Experimental configuration and hyperparameter settings

3.3 Object tracking algorithm detection effect

3.4 Experimental analysis of object detection model

3.4.1 The effect of shot angle on the detection results

3.4.2 Ablation test

4 Discussion And Analysis

4.1 Object detection model

4.1.1 Module contribution

4.1.2 Effect of object detection

4.1.3 Noise Effect

4.2 Analysis of classification detection of foraging behavior

4.3 Analysis of limitations

5. Conclusion

Declarations

References

Additional Declarations

Status:

Version 1