Method for analyzing video behavior of non-cage broiler chickens under target detection perspective

Abstract

This disclosure relates to a video behavior analysis method for cold stress in cage-free broiler chickens from the perspective of target detection, aiming to improve the accuracy of cold stress state determination. The method includes: segmenting the video to be detected to obtain multiple frames of images to be detected; processing each frame of images to be detected using a target detection model to obtain detection results corresponding to each frame; determining the quantity and frequency of each non-specific behavior of the broiler chickens based on the non-specific behaviors in the detection results corresponding to each frame of images; determining the crowding level of the broiler chickens based on the broiler chicken detection boxes in the detection results corresponding to each frame of images; determining the location distribution of the broiler chickens based on the location information of the broiler chicken detection boxes in the detection results corresponding to each frame of images; and determining the cold stress state of the broiler chickens in the cage-free environment based on the quantity and frequency of each non-specific behavior, the crowding level, and the location distribution of the broiler chickens.

Description

Technical Field

[0001] This disclosure relates to the field of data analysis technology, specifically to a method for analyzing the cold stress video behavior of cage-free broiler chickens from the perspective of target detection. Background Technology

[0002] As consumers become increasingly demanding in their requirements for chicken quality, cage-free farming, which is more effective at improving meat quality compared to traditional intensive farming, is gradually becoming the mainstream. Cage-free farming is a new type of farming model that means animals are not confined to traditional cages during their growth process.

[0003] In addition, broilers are highly susceptible to cold stress in cold weather, which can negatively impact their production. Therefore, it is crucial to assess the cold stress status of broilers. Consequently, it is essential to analyze the cold stress status of cage-free broilers in conjunction with the requirements of cage-free rearing. Summary of the Invention

[0004] The purpose of this disclosure is to provide a method, apparatus, storage medium, and electronic device for analyzing cold stress video behavior of cage-free broiler chickens from the perspective of target detection, so as to at least partially solve the above-mentioned problems existing in the related technologies.

[0005] To achieve the above objectives, in a first aspect, this disclosure provides a method for analyzing the cold stress video behavior of cage-free broiler chickens from the perspective of target detection, the method comprising:

[0006] The video to be detected is segmented to obtain multiple frames of images to be detected. The video to be detected is a video collected from the target detection perspective of the non-cage-raised broiler chickens.

[0007] The target detection model is used to process each frame of the image to be detected to obtain the detection results corresponding to each frame of the image to be detected. The detection results include the botnet detection box, the non-specific behavior of the botnet selected by the botnet detection box, the position information of the botnet detection box, and the size information of the botnet detection box.

[0008] Based on the non-specific behaviors of chickens in the detection results corresponding to each frame of the image to be detected, the number and frequency of each non-specific behavior of chickens during the shooting time of the video to be detected are determined. Based on the chicken detection boxes in the detection results corresponding to each frame of the image to be detected, the crowding degree of chickens during the shooting time of the video to be detected is determined. Based on the position information of the chicken detection boxes in the detection results corresponding to each frame of the image to be detected, the position distribution of chickens during the shooting time of the video to be detected is determined.

[0009] Based on the number and frequency of various non-specific behaviors of broilers during the video recording period, the crowding degree of the broilers, and the location distribution of the broilers, the cold stress of broilers in the non-cage environment is determined.

[0010] Optionally, determining the crowding level of chickens within the time frame of the video to be detected based on the chicken detection boxes in the detection results corresponding to each frame of the image to be detected includes:

[0011] For any frame of the image to be detected, determine the number of times the boundaries of the chicken detection boxes in that frame of the image to be detected intersect.

[0012] Based on the correspondence between the number of boundary intersections and the density, the density corresponding to the number of boundary intersections of the chicken detection box in the image to be detected in that frame is determined;

[0013] Based on the density of each frame of the image to be detected, the crowding level of the broilers during the video recording time is determined.

[0014] Optionally, the position information of the broiler detection box is the coordinates of the geometric center point of the broiler detection box. Determining the position distribution of broilers within the time frame of the video to be detected based on the position information of the broiler detection boxes in the detection results corresponding to each frame of the image to be detected includes:

[0015] Based on the geometric center coordinates of the botnet detection boxes included in each frame of the image to be detected, the distribution ratio of the botnet detection boxes in the preset area of ​​the image during the shooting time of the video to be detected is statistically analyzed.

[0016] Optionally, the target detection model is trained through the following steps:

[0017] Obtain a sample dataset, wherein each sample data in the sample dataset includes an image of broiler chickens in a cage-free environment collected from the perspective of target detection and the corresponding annotation information of the image. The annotation information includes non-specific behavior labels of broiler chickens in the image and a rectangular box used to select broiler chickens.

[0018] The single-stage detection model is trained using the sample dataset to obtain the target detection model.

[0019] Optionally, the single-stage detection model includes an input network, a backbone network, a neck network, and a head network. The backbone network includes a first convolutional module, a second convolutional module, a first C2F module, a first simAM attention mechanism module, a third convolutional module, a second C2F module, a second simAM attention mechanism module, a fourth convolutional module, a third simAM attention mechanism module, a fifth convolutional module, a fourth simAM attention mechanism module, and an SPPF module connected in sequence. The outputs of the second simAM attention mechanism module, the third simAM attention mechanism module, and the fourth simAM attention mechanism module serve as the three inputs of the neck network, respectively.

[0020] Optionally, the intermediate results output by the head network include detection boxes and corresponding confidence scores. The single-stage detection model further includes a post-processing network connected to the output of the head network. The post-processing network is used to perform the following processing on the intermediate results output by the head network:

[0021] The detection boxes are sorted according to their confidence scores.

[0022] The confidence S of all detection boxes except the candidate detection box with the highest confidence is calculated using the following formula. x :

[0023]

[0024]

[0025] Where M represents the candidate detection box, B x Represents the xth other detection box, IOU(M,B) x R represents the intersection-union ratio (IoU) between a candidate bounding box and the xth other bounding box. DIOU (M,B x ) represents the first intermediate parameter, and b represents the center point of other detection boxes. gt ρ represents the center point of the candidate detection box. 2 (b,b gt ) represents the square of the distance between the center point of the x-th other detection box and the center point of the candidate detection box, c 2 It represents the square of the diagonal length of the smallest bounding rectangle of the x-th other detection box and the candidate detection box;

[0026] Delete other detection boxes whose confidence Sx is greater than the preset confidence threshold, and determine the remaining other detection boxes as the detection boxes. Return to the execution step: sort the detection boxes according to the confidence of each detection box until there are no more remaining detection boxes.

[0027] The candidate detection boxes determined in each loop are selected as the detection boxes output by the single-stage detection model.

[0028] Optionally, the cross-union ratio between the candidate detection box and the xth other detection box is calculated using the third version of the WIOU strategy.

[0029] To achieve the above objectives, in a second aspect, this disclosure provides a device for analyzing the cold stress video behavior of cage-free broiler chickens from the perspective of target detection, the device comprising:

[0030] The image segmentation module is used to segment the video to be detected to obtain multiple frames of images to be detected. The video to be detected is a video collected from the non-cage-raised environment of broilers from the target detection perspective.

[0031] The image processing module is used to process each frame of the image to be detected using the target detection model to obtain the detection results corresponding to each frame of the image to be detected. The detection results include the broiler detection box, the non-specific behavior of the broiler selected by the broiler detection box, the position information of the broiler detection box, and the size information of the broiler detection box.

[0032] The first determining module is used to determine the number and frequency of each non-specific behavior of chickens during the shooting time of the video under test based on the non-specific behaviors of chickens in the detection results corresponding to each frame of the image under test; to determine the crowding degree of chickens during the shooting time of the video under test based on the chicken detection boxes in the detection results corresponding to each frame of the image under test; and to determine the position distribution of chickens during the shooting time of the video under test based on the position information of the chicken detection boxes in the detection results corresponding to each frame of the image under test.

[0033] The second determining module is used to determine the cold stress status of broilers in the non-cage environment based on the number and frequency of various non-specific behaviors of broilers during the video recording time, the crowding degree of the broilers, and the location distribution of the broilers.

[0034] Thirdly, this disclosure provides a non-transitory computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the steps of the method described in any of the first aspects.

[0035] Fourthly, this disclosure provides an electronic device, comprising:

[0036] A memory on which computer programs are stored;

[0037] A processor for executing the computer program in the memory to implement the steps of the method of any one of the first aspects.

[0038] By incorporating the target detection model into the video analysis method for cold stress in cage-free broilers from the perspective of target detection, the analysis efficiency is greatly improved compared to manual video analysis. In addition, the embodiments of this disclosure further analyze the quantity and frequency changes of non-specific behaviors of cage-free broilers in the video based on the detection results of the target detection model. At the same time, the crowding degree and location distribution of cage-free broilers are also statistically analyzed. By combining multiple aspects to comprehensively analyze the cold stress state of cage-free broilers, the accuracy of the cold stress state analysis of cage-free broilers is improved.

[0039] Other features and advantages of this disclosure will be described in detail in the following detailed description section. Attached Figure Description

[0040] The accompanying drawings are provided to further illustrate the present disclosure and form part of the specification. They are used together with the following detailed description to explain the present disclosure, but do not constitute a limitation thereof. In the drawings:

[0041] Figure 1 This is a flowchart illustrating an exemplary embodiment of the present disclosure of a method for analyzing cold stress video behavior of cage-free broiler chickens from the perspective of target detection;

[0042] Figure 2 This is a schematic diagram of the structure of a single-stage detection model according to an exemplary embodiment of the present disclosure;

[0043] Figure 3 This is a block diagram illustrating a video behavior analysis device for cold stress in cage-free broiler chickens from the perspective of target detection, according to an exemplary embodiment of this disclosure.

[0044] Figure 4 This is a block diagram illustrating an electronic device according to an exemplary embodiment of the present disclosure. Detailed Implementation

[0045] The specific embodiments of this disclosure will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit this disclosure.

[0046] The applicant discovered through long-term research that, in related technologies, conventional video analysis methods can be used—that is, manually identifying non-specific behaviors in videos frame by frame, and then further analyzing the stress state of cage-free broilers based on these non-specific behaviors. However, this method is inefficient and its accuracy cannot be guaranteed.

[0047] Furthermore, some related technologies utilize computer vision (CV)-based object detection models to detect animal behavior and further analyze the animal's state based on the behavior. However, the applicant discovered that due to the unique characteristics of non-cage-raised broiler chicken scenarios, for example:

[0048] 1. The specific features of different behaviors become smaller under specific detection perspectives, resulting in lower distinguishability. For example, under an overhead view, the differences between behaviors are mostly reflected in smaller parts such as the feet and head, which can easily lead to inaccurate behavior differentiation;

[0049] 2. In low-temperature environments, chickens exhibit more pronounced aggregation, resulting in significant occlusion in images and increasing the likelihood of false positives in individual identification. Parts of some broiler chickens' bodies may be obscured by other chickens, making accurate identification of specific individuals more difficult and increasing the likelihood of false positives.

[0050] 3. In images from non-cage environments, the ground background is often complex, and some parts of the background are very similar in color and texture to the broiler chickens. This can easily lead to a lack of clear distinction between the broiler chicken image and the background, posing a challenge to the detection and identification of broiler chickens.

[0051] 4. Complex lighting conditions in the scene may lead to variations in brightness, color, and messy textures in the background. These variations and noises can interfere with the target detector's judgment, causing it to mistakenly identify the variations and noises in the background as targets.

[0052] Therefore, the method of using the target detection model commonly used in related technologies to detect the behavior of cage-free broilers and further analyzing the cold stress behavior of cage-free broilers based on the behavior also suffers from inaccurate detection.

[0053] In response to the above findings, the applicant proposes a method for analyzing cold stress video behavior of cage-free broilers from the perspective of target detection, in order to improve the accuracy of cold stress video behavior analysis of cage-free broilers from the perspective of target detection.

[0054] Figure 1 This is a flowchart illustrating a method for analyzing cold stress video behavior of cage-free broiler chickens from the perspective of target detection, according to an exemplary embodiment of this disclosure. (Refer to...) Figure 1 The video behavior analysis method for cold stress in non-cage-raised broiler chickens from the perspective of target detection includes:

[0055] S110, the video to be detected is segmented to obtain multiple frames of images to be detected. The video to be detected is a video collected from the non-cage-raised environment of broilers from the target detection perspective.

[0056] In some implementations, the target detection perspective can be an overhead view to more accurately capture the behavior of cage-free broiler chickens.

[0057] In this embodiment of the disclosure, the video to be detected can be a video over a period of time. After the video to be detected is obtained, the video can be segmented frame by frame to obtain multiple frames of images, i.e., multiple frames of images to be detected.

[0058] S120, the target detection model is used to process each frame of the image to be detected to obtain the detection results corresponding to each frame of the image to be detected. The detection results include the botnet detection box, the non-specific behavior of the botnet selected by the botnet detection box, the position information of the botnet detection box, and the size information of the botnet detection box.

[0059] In this embodiment, each frame of a multi-frame image to be detected can be fed into a target detection model for processing, resulting in a processing result for each frame. The processing result for each frame can include two parts: a schematic diagram with a bounding box added to the image and a document containing detection information. In the document, each line represents a detected target, i.e., a botnet. The first set of numbers in each line represents the botnet's non-specific behavior, and the following four sets of numbers represent the normalized x and y coordinates of the center point of the botnet's bounding box, as well as the length and width of the bounding box. These five pieces of information collectively represent the non-specific behavior of a botnet in an image, the location of the botnet's bounding box, and the size of the bounding box.

[0060] In some implementations, nonspecific behaviors may include standing, lying down, pecking feathers, flapping wings, and feeding.

[0061] S130, based on the non-specific behaviors of chickens in the detection results corresponding to each frame of the image to be detected, determine the number and frequency of each non-specific behavior of chickens during the shooting time of the video to be detected; based on the chicken detection boxes in the detection results corresponding to each frame of the image to be detected, determine the crowding degree of chickens during the shooting time of the video to be detected; based on the position information of the chicken detection boxes in the detection results corresponding to each frame of the image to be detected, determine the position distribution of chickens during the shooting time of the video to be detected.

[0062] In the embodiments of the present disclosure, after obtaining the detection results corresponding to each frame of the to-be-detected images, analysis can be performed based on the detection results. That is, further based on the non-specific behaviors of the broilers in the detection results corresponding to each frame of the to-be-detected images, the quantity and frequency of each non-specific behavior of the broilers within the shooting time of the to-be-detected video are determined. Based on the broiler detection frames in the detection results corresponding to each frame of the to-be-detected images, the crowding degree of the broilers within the shooting time of the to-be-detected video is determined. Based on the position information of the broiler detection frames in the detection results corresponding to each frame of the to-be-detected images, the position distribution of the broilers within the shooting time of the to-be-detected video is determined.

[0063] In some embodiments, after determining the quantity and frequency of each non-specific behavior of the broilers within the shooting time of the to-be-detected video, the quantity and frequency of each non-specific behavior can be represented by drawing a heat map to show more intuitively.

[0064] In some embodiments, determining the crowding degree of the broilers within the shooting time of the to-be-detected video based on the broiler detection frames in the detection results corresponding to each frame of the to-be-detected images may include the following steps:

[0065] For any frame of the to-be-detected image, determine the number of boundary intersections of the broiler detection frame in this frame of the to-be-detected image;

[0066] Based on the corresponding relationship between the number of boundary intersections and the density degree, determine the density degree corresponding to the number of boundary intersections of the broiler detection frame in this frame of the to-be-detected image;

[0067] Based on the density degrees corresponding to each frame of the to-be-detected images, determine the crowding degree of the broilers within the shooting time of the to-be-detected video.

[0068] Considering that in a non-cage environment, the density of broilers is small, and the distribution of broilers is relatively scattered under normal circumstances. In object detection, the case of target dispersion is approximately the case where the number of intersections of broiler detection frames is small. Therefore, in the embodiments of the present disclosure, borrowing the idea of IOU (Intersection over Union), the Boundary crossing frequency (BRF) is selected to describe the crowded aggregation situation of the chicken flock, that is, the density degree.

[0069] In some embodiments, the range of the number of boundary intersections corresponding to the density degree can be determined according to the number of broilers in the non-cage environment. Exemplarily, assuming there are 8 non-cage broilers, if the density degree is divided into three levels according to the BRF value, the BRF value ranges corresponding to each density degree can be determined: non-aggregated (0 < BRF < 8), moderate (9 < BRF < 15), aggregated (BRF ≥ 16).

[0070] In this embodiment of the disclosure, for any frame of the image to be detected, after determining the number of boundary intersections of the chicken detection boxes in the frame of the image to be detected, the density corresponding to the number of boundary intersections of the chicken detection boxes in the frame of the image to be detected can be determined according to the pre-determined correspondence between the number of boundary intersections and the density. Then, the crowding degree of chickens during the shooting time of the video to be detected can be determined according to the density corresponding to each frame of the image to be detected in the video.

[0071] In some implementations, the crowding level of broilers during the video shooting time is determined based on the density of each frame of the image to be detected in the video. For example, the density in the video can be determined as the ratio of the total number of clustered image frames to the total number of video frames, thus determining the crowding level of broilers during the video shooting time.

[0072] Furthermore, considering that moderate density can reflect density to some extent, in other embodiments, a first ratio of the total number of image frames with moderate density to the total number of video frames can be determined, and a second ratio of the total number of image frames with clustered density to the total number of video frames can be determined. The first ratio and the second ratio are then weighted and summed to obtain the crowding level of the broilers during the video recording time. The weighting coefficient corresponding to the first ratio is smaller than the weighting coefficient corresponding to the second ratio.

[0073] As can be seen from the foregoing, in some implementations, the location information of the botnet detection box can be the coordinates of the geometric center point of the botnet detection box. In this case, determining the location distribution of botnets within the time frame of the video to be detected based on the location information of the botnet detection boxes in the detection results corresponding to each frame of the image to be detected can include the following steps:

[0074] Based on the geometric center coordinates of the broiler detection boxes included in each frame of the image to be detected, the distribution ratio of the broiler detection boxes in the preset area of ​​the image during the video shooting time is statistically analyzed.

[0075] In this embodiment of the disclosure, the location distribution of broilers in the non-cage-raising area can be represented by the position of the geometric center point of the broiler detection frame in the image. Therefore, based on the coordinates of the geometric center point of the broiler detection frame included in each frame of the image to be detected, the distribution ratio of the broiler detection frame in the preset area of ​​the image during the video shooting time can be statistically analyzed.

[0076] In some implementations, the image can be divided into 9 regions, and the distribution ratio of botnets in different regions can be counted by using the geometric center point of the botnet detection box. The region with the highest proportion in each image is the representative region of that image. Finally, the proportion of the representative region of each frame of the video to be detected in each region is counted. The proportion of the representative region of each frame of the video to be detected in each region is determined as the distribution ratio of the botnet detection box in the preset region of the image.

[0077] S140. Based on the number and frequency of various non-specific behaviors of broilers, the crowding degree of broilers, and the location distribution of broilers during the video recording time, determine the cold stress of broilers in the non-cage environment.

[0078] In some implementations, after determining the number and frequency of various non-specific behaviors of broilers, the degree of crowding of broilers, and the location distribution of broilers during the video recording period, a comprehensive analysis can be conducted based on the number and frequency of various non-specific behaviors of broilers, the degree of crowding of broilers, and the location distribution of broilers during the video recording period to determine the cold stress status of broilers in the cageless environment, that is, whether broilers in the cageless environment are in a state of cold stress.

[0079] For example, when one or more of the following characteristics are present, it can be preliminarily determined that broilers in a non-cage environment are under cold stress: the number of one or more behaviors such as feeding, flapping wings, pecking feathers, and lying down is reduced compared to the normal environment; the standing behavior is increased compared to the normal environment; the crowding of broilers during the video recording time is increased compared to the normal environment; and the location distribution of broilers during the video recording time indicates that a large number of broilers are distributed in areas with higher temperatures.

[0080] By incorporating the target detection model into the video analysis of cold stress in cage-free broilers from the perspective of target detection, the analysis efficiency is greatly improved compared to manual video analysis. Furthermore, this embodiment of the present disclosure further analyzes the quantity and frequency changes of non-specific behaviors of cage-free broilers in the video based on the detection results of the target detection model. It also statistically analyzes the crowding degree and location distribution of cage-free broilers. By combining multiple aspects to comprehensively analyze the cold stress state of cage-free broilers, the accuracy of the cold stress state analysis of cage-free broilers is improved.

[0081] Furthermore, this disclosure also provides a training method for the target detection model, as detailed in subsequent embodiments.

[0082] In some implementations, the target detection model used in the target detection perspective of the cold stress video behavior analysis method for cage-free broiler chickens according to the embodiments of this disclosure can be trained through the following steps:

[0083] Obtain a sample dataset. Each sample dataset includes an image of broiler chickens in a cage-free environment collected from the perspective of target detection, as well as the corresponding annotation information. The annotation information includes non-specific behavioral labels of broiler chickens in the image and a rectangular box used to select broiler chickens.

[0084] The single-stage detection model is trained using the sample dataset to obtain the target detection model.

[0085] In this embodiment of the disclosure, in order to improve the efficiency of video behavior analysis of non-cage-raised broiler chickens under cold stress from the perspective of target detection, a single-stage detection model is selected. A single-stage detection model is a model that completes the localization and classification tasks simultaneously.

[0086] In some implementations, the single-stage detection model of this disclosure may be a YOLO (You Only Look Once) series model.

[0087] In some implementations, raw video data of broiler chickens in an uncage-raised environment can be pre-collected from the target detection perspective, and 15,000 JPG images can be extracted from the collected raw video data. The image data is then manually filtered, and 1,200 clear images with rich behavioral representations are used to construct a sample dataset. Image annotation tools are then used to annotate the selected broiler chicken images. During the annotation process, chickens are marked with bounding boxes, and non-specific behaviors of each chicken within the bounding box are also labeled.

[0088] After the annotation is completed, data augmentation techniques such as rotation, dropout, and affine transformation can be used to expand the sample dataset to 2400 images, making the candidate bounding boxes of botnets more uniform from the perspective of object detection.

[0089] The sample dataset was randomly divided in a 6:2:2 ratio, with 1440 images in the training set, and 480 images each in the test and validation sets.

[0090] Next, the single-stage detection model can be trained using the sample dataset to obtain the target detection model.

[0091] Furthermore, considering the specific characteristics of the non-cage-raised broiler chicken scenario in this disclosure, on the one hand, the smaller key features require the model to have extremely strong feature extraction capabilities, meaning the feature extraction module should focus more on subtle features that can distinguish behavior categories; on the other hand, dense occlusion can significantly interfere with behavior detection, and the model should be able to better distinguish targets in dense scenes. Therefore, the current YOLO series models cannot adequately meet the requirements. To improve the detection accuracy of the target detection model, this disclosure improves upon the YOLOv8 model by replacing some of the original C2F (Convolutional To Full Connected) modules with the SimAM attention mechanism module. This gives the target detection model of this disclosure a better ability to focus on key features, making it more suitable for broiler chicken detection in non-cage-raised broiler chicken scenarios. Simultaneously, this disclosure also proposes a DIOU-based NMS, which quickly and accurately distinguishes dense broiler chickens in the post-processing stage of the algorithm. Furthermore, replacing the traditional CIOU in the target detection model with WIOU brings a dual improvement in IOU calculation speed and accuracy. In addition, the FasterNet module, introduced to improve computational speed, makes the object detection model more lightweight.

[0092] In some implementations, such as Figure 2 As shown, the single-stage detection model includes an input network 21, a backbone network 22, a neck network 23, and a head network 24. The backbone network 22 includes a first convolutional module 2201, a second convolutional module 2202, a first C2F module 2203, a first simAM attention mechanism module 2204, a third convolutional module 2205, a second C2F module 2206, a second simAM attention mechanism module 2207, a fourth convolutional module 2208, a third simAM attention mechanism module 2209, a fifth convolutional module 2210, a fourth simAM attention mechanism module 2211, and an SPPF module 2212, which are connected in sequence. The outputs of the second simAM attention mechanism module 2207, the third simAM attention mechanism module 2209, and the fourth simAM attention mechanism module 2211 serve as the three inputs to the neck network 23.

[0093] In some implementations, the simAM attention mechanism module includes an energy function for neurons and a scaling module. After estimating the energy of each neuron, the scaling module can be used for feature refinement.

[0094] The energy function can be expressed as follows:

[0095]

[0096] Among them, and yes and The linear transformation, and and The input feature X∈Γ C×H×W The target neuron and other neurons in a single channel, where i is the index in the spatial dimension, M = H × W is the number of neurons in that channel, and ω t t and b t It refers to the transformation weights and biases, and the energy e. t The lower the value, the greater the difference between this neuron and surrounding neurons, and the more important it is for visual processing.

[0097] The scaling term module can be represented as follows:

[0098]

[0099] Where E represents all e t Grouping is done in both spatial and channel dimensions; the sigmoid function is added to limit e. t Neurons with excessive energy.

[0100] In this embodiment of the disclosure, the applicant's long-term research has revealed that in broiler detection in non-cage-raised broiler scenarios, the original C2F module of YOLOv8 focuses too much on the background and irrelevant pixels of the image for feature extraction. This results in the model being unable to focus on the detailed features of the broiler, thus preventing the overall accuracy of the model from improving. By adding the SimAM module to the corresponding layer, the three-dimensional attention weights of the feature mapping in the layer can be inferred without adding parameters to the original network. This allows the SimAM module to focus more on the key neurons containing feature information, thereby focusing on the key features in the image details and significantly improving the overall classification accuracy of the target detection model.

[0101] In some implementations, the specific structures of the input network, neck network, and head network can be referred to in related technologies, and will not be elaborated here.

[0102] Furthermore, considering that the intermediate results output by the head network include bounding boxes and corresponding confidence scores, in order to filter out the bounding boxes corresponding to the botnet from all bounding boxes, in some implementations, the following steps are continued... Figure 2 As shown, the single-stage detection model also includes a post-processing network 25 connected to the output of the head network. The post-processing network 25 is used to perform the following processing on the intermediate results output by the head network:

[0103] The detection boxes are sorted according to their confidence scores.

[0104] The confidence S of all detection boxes except the candidate detection box with the highest confidence is calculated using the following formula. x :

[0105]

[0106]

[0107] Where M represents the candidate detection box, B x Represents the xth other detection box, IOU(M, B) x R represents the intersection-union ratio (IoU) between a candidate bounding box and the xth other bounding box. DIOU (M, B) x ) represents the first intermediate parameter, and b represents the center point of other detection boxes. gt ρ represents the center point of the candidate detection box. 2 (b,b gt ) represents the square of the distance between the center point of the x-th other detection box and the center point of the candidate detection box, c 2 It represents the square of the diagonal length of the smallest bounding rectangle of the x-th other detection box and the candidate detection box;

[0108] Delete other detection boxes whose confidence Sx is greater than the preset confidence threshold, and identify the remaining detection boxes as individual detection boxes. Return to the execution steps: sort the detection boxes according to their confidence scores until there are no more remaining detection boxes.

[0109] The candidate detection boxes determined in each iteration are selected as the detection boxes output by the single-stage detection model.

[0110] Considering that the intermediate results include multiple rectangular detection boxes of different sizes, positions and angles, these rectangular detection boxes may overlap and cover the same detection target. Redundant detection boxes can be eliminated by using a post-processing network, and the best rectangular detection box can be selected to describe the target.

[0111] In this embodiment of the disclosure, after receiving the intermediate results output by the output end of the head network, the post-processing network can sort the detection boxes according to the confidence level of each detection box, and then determine the detection box with the highest confidence level as the candidate detection box. Next, the confidence level of the other detection boxes excluding the candidate detection boxes can be calculated. After determining the confidence level of the other detection boxes excluding the candidate detection boxes, the other detection boxes with a calculated confidence level greater than a preset confidence threshold can be determined as redundant detection boxes and deleted. Then, the remaining other detection boxes are re-sorted according to the calculated confidence level, and the above process is repeated until there are no remaining other detection boxes. After this process, the candidate detection boxes determined in each loop can be determined as the detection boxes output by the single-stage detection model.

[0112] In this embodiment of the disclosure, when calculating the confidence of other detection boxes besides the candidate detection box, not only the intersection-union ratio (IU) between the candidate detection box and the xth other detection box is considered, but also the distance between the center point of the candidate detection box and the xth other detection box is considered. This makes the IU between the candidate detection box and the xth other detection box greater than the threshold, but it can still not be deleted when the distance between the two detection boxes is large, thereby improving the accuracy of detection box screening. This effectively overcomes the impact of dense occlusion on broiler detection in non-cage-raised broiler scenarios and improves the accuracy of broiler detection in non-cage-raised broiler scenarios.

[0113] In some implementations, to reduce the overhead of the training process, the cross-union ratio between the candidate detection box and the xth other detection box is calculated using a third version of the WIOU strategy.

[0114] WIOU provides three versions of the function: V1 (first version), V2 (second version), and V3 (third version). WIOU1 implements the construction of the boundary loss, while V2 and V3 add a focusing coefficient on top of that.

[0115] The formula for calculating WIOU V1 is as follows:

[0116] WIOU V1 =R WIOU L IOU

[0117]

[0118]

[0119] The formula for calculating WIOU V2 is as follows:

[0120]

[0121] in, It calculates the exponential running average with momentum m, and dynamically updates the normalization factor to improve the gradient gain. It maintained a high level and solved the problem of slow convergence in the later stages of training. This represents the average value of IOU.

[0122] WIOU V3, which has all the strategies, is represented as follows:

[0123] WIOU V3 =rWIOU V1

[0124]

[0125]

[0126] Where β represents the outlier degree of the anchor frame, and α and δ are hyperparameters.

[0127] In this case, a smaller outlier means a higher quality anchor frame, and the IOU is determined by gradient gain. This allows us to shift our computational focus away from high-quality anchor frames and instead concentrate on anchor frames of general quality. Furthermore, since the criteria for classifying anchor frame quality are dynamically changing, ... It is also constantly changing, which allows WIOU V3 to allocate the most reasonable gradient gain at any point during model training.

[0128] For example, in the mid-to-late stages of model training, when the model is about to converge, WIOU can allocate small gradient gains to low-quality anchor boxes to reduce harmful gradients, thereby further reducing the Boxloss.

[0129] In one example, in the single-stage detection model of this disclosure embodiment, the SGD optimizer is selected as the optimization algorithm, the model input is set to 640*640, the initial learning rate is fixed at 0.01, and the preset confidence threshold is 0.5. To ensure the reproducibility of experimental results, a seed is used in this disclosure embodiment and set to 100. The training epoch is set to automatic stop training. By automatically judging the convergence degree of the model and setting patience to 100, training automatically stops after 100 epochs of convergence.

[0130] The object detection model trained using the single-stage detection model of this disclosure considers not only the intersection-union ratio (IU) between the candidate detection box and the xth other detection box, but also the distance between the center point of the candidate detection box and the xth other detection box when calculating the confidence of other detection boxes besides the candidate detection box. This effectively overcomes the impact of dense occlusion on broiler detection in cage-free broiler scenes and improves the accuracy of broiler detection in cage-free broiler scenes. In addition, the third version of the WIOU strategy is used to calculate the IU between the candidate detection box and the xth other detection box, which not only maintains the recognition accuracy but also reduces the inference time. Furthermore, the addition of SimAM allows the model to pay more attention to the key details and features in cage-free broiler scene images. It can be seen that the object detection model trained by the single-stage detection model of this disclosure considers the impact of feature extraction, dense differentiation, and computational cost, balancing detection accuracy and detection speed, and is organically combined with cage-free broiler scenes.

[0131] Based on the same concept, this disclosure also provides a device for analyzing the cold stress video behavior of cage-free broiler chickens from the perspective of target detection. This device can be part or all of an electronic device through software, hardware, or a combination of both. (See also...) Figure 3 The target detection perspective of the non-cage-raised broiler cold stress video behavior analysis device 300 may include:

[0132] The image segmentation module 310 is used to segment the video to be detected to obtain multiple frames of images to be detected. The video to be detected is a video collected from the non-cage-raised environment of broilers from the target detection perspective.

[0133] Image processing module 320 is used to process each frame of the image to be detected using a target detection model to obtain the detection results corresponding to each frame of the image to be detected. The detection results include the broiler detection box, the non-specific behavior of the broiler selected by the broiler detection box, the position information of the broiler detection box, and the size information of the broiler detection box.

[0134] The first determining module 330 is used to determine the number and frequency of each non-specific behavior of chickens during the shooting time of the video under test based on the non-specific behaviors of chickens in the detection results corresponding to each frame of the image under test; to determine the crowding degree of chickens during the shooting time of the video under test based on the chicken detection boxes in the detection results corresponding to each frame of the image under test; and to determine the position distribution of chickens during the shooting time of the video under test based on the position information of the chicken detection boxes in the detection results corresponding to each frame of the image under test.

[0135] The second determining module 340 is used to determine the cold stress of broilers in the non-cage environment based on the number and frequency of various non-specific behaviors of broilers during the video recording time, the crowding degree of the broilers, and the location distribution of the broilers.

[0136] Optionally, the first determining module 330 is further configured to, for any frame of the image to be detected, determine the number of times the boundaries of the chicken detection boxes in the frame of the image to be detected intersect; based on the correspondence between the number of boundary intersections and the density, determine the density corresponding to the number of boundary intersections of the chicken detection boxes in the frame of the image to be detected; and based on the density corresponding to each frame of the image to be detected, determine the crowding degree of the chickens during the video shooting time.

[0137] Optionally, the position information of the botnet detection box is the coordinates of the geometric center point of the botnet detection box. The first determining module 330 is further used to calculate the distribution ratio of the botnet detection box in the preset area of ​​the image during the shooting time of the video to be detected based on the coordinates of the geometric center point of the botnet detection box included in each frame of the image to be detected.

[0138] Optionally, the target detection perspective video behavior analysis device 300 for non-cage-raised broiler chickens under cold stress also includes:

[0139] The model training module is used to acquire a sample dataset. Each sample data in the sample dataset includes an image of broiler chickens in a cage-free environment collected from the perspective of target detection, and the corresponding annotation information of the image. The annotation information includes non-specific behavioral labels of broiler chickens in the image and a rectangular box used to select broiler chickens. The single-stage detection model is trained using the sample dataset to obtain the target detection model.

[0140] Optionally, the single-stage detection model includes an input network, a backbone network, a neck network, and a head network. The backbone network includes a first convolutional module, a second convolutional module, a first C2F module, a first simAM attention mechanism module, a third convolutional module, a second C2F module, a second simAM attention mechanism module, a fourth convolutional module, a third simAM attention mechanism module, a fifth convolutional module, a fourth simAM attention mechanism module, and an SPPF module connected in sequence. The outputs of the second simAM attention mechanism module, the third simAM attention mechanism module, and the fourth simAM attention mechanism module serve as the three inputs of the neck network, respectively.

[0141] Optionally, the intermediate results output by the head network include detection boxes and corresponding confidence scores. The single-stage detection model further includes a post-processing network connected to the output of the head network. The post-processing network is used to perform the following processing on the intermediate results output by the head network:

[0142] The detection boxes are sorted according to their confidence scores.

[0143] The confidence S of all detection boxes except the candidate detection box with the highest confidence is calculated using the following formula. x :

[0144]

[0145]

[0146] Where M represents the candidate detection box, B x Represents the xth other detection box, IOU(M,B) x R represents the intersection-union ratio (IoU) between a candidate bounding box and the xth other bounding box. DIOU (M, B) x ) represents the first intermediate parameter, and b represents the center point of other detection boxes. gt ρ represents the center point of the candidate detection box. 2 (b,b gt ) represents the square of the distance between the center point of the x-th other detection box and the center point of the candidate detection box, c 2 It represents the square of the diagonal length of the smallest bounding rectangle of the x-th other detection box and the candidate detection box;

[0147] Delete other detection boxes whose confidence Sx is greater than the preset confidence threshold, and determine the remaining other detection boxes as the detection boxes. Return to the execution step: sort the detection boxes according to the confidence of each detection box until there are no more remaining detection boxes.

[0148] The candidate detection boxes determined in each loop are selected as the detection boxes output by the single-stage detection model.

[0149] Optionally, the cross-union ratio between the candidate detection box and the xth other detection box is calculated using the third version of the WIOU strategy.

[0150] Regarding the apparatus in the above embodiments, the specific manner in which each module performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated upon here.

[0151] Based on the same inventive concept, this disclosure also provides an electronic device, comprising:

[0152] A memory on which computer programs are stored;

[0153] A processor for executing the computer program in the memory to implement the steps of any of the above data determination methods.

[0154] In one possible manner, the block diagram of the electronic device can be as follows: Figure 4 As shown. (Refer to...) Figure 4 The electronic device 400 may include a processor 401 and a memory 402. The electronic device 400 may also include one or more of a multimedia component 403, an input / output (I / O) interface 404, and a communication component 405.

[0155] The processor 401 controls the overall operation of the electronic device 400 to complete all or part of the steps in the above-mentioned method for analyzing cold stress video behavior in cage-free broiler chickens from the perspective of target detection. The memory 402 stores various types of data to support the operation of the electronic device 400. This data may include, for example, instructions for any application or method operating on the electronic device 400, and application-related data such as contact data, sent and received messages, images, audio, video, etc. The memory 402 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. Multimedia component 403 may include a screen and an audio component. The screen may be, for example, a touchscreen, and the audio component is used to output and / or input audio signals. For example, the audio component may include a microphone for receiving external audio signals. The received audio signals may be further stored in memory 402 or transmitted via communication component 405. The audio component also includes at least one speaker for outputting audio signals. I / O interface 404 provides an interface between processor 401 and other interface modules, such as a keyboard, mouse, buttons, etc. These buttons may be virtual or physical buttons. Communication component 405 is used for wired or wireless communication between the electronic device 400 and other devices. Wireless communication, such as Wi-Fi, Bluetooth, Near Field Communication (NFC), 2G, 3G, 4G, NB-IoT, eMTC, or other 5G technologies, or combinations thereof, is not limited here. Therefore, the corresponding communication component 405 may include: a Wi-Fi module, a Bluetooth module, an NFC module, etc.

[0156] In an exemplary embodiment, the electronic device 400 may be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to perform the above-described method for analyzing cold stress video behavior of cage-free broiler chickens from the perspective of target detection.

[0157] In another exemplary embodiment, a computer-readable storage medium including program instructions is also provided. When executed by a processor, these program instructions implement the steps of the above-described method for analyzing the cold stress video behavior of cage-free broilers from the perspective of target detection. For example, the computer-readable storage medium may be the memory 402 including the program instructions, which may be executed by the processor 401 of the electronic device 400 to complete the above-described method for analyzing the cold stress video behavior of cage-free broilers from the perspective of target detection.

[0158] In another exemplary embodiment, a computer program product is also provided, comprising a computer program executable by a programmable device, the computer program having a code portion for performing the above-described method for analyzing cold stress video behavior of cage-free broiler chickens from the perspective of target detection when executed by the programmable device.

[0159] The preferred embodiments of this disclosure have been described in detail above with reference to the accompanying drawings. However, this disclosure is not limited to the specific details of the above embodiments. Within the scope of the technical concept of this disclosure, various simple modifications can be made to the technical solutions of this disclosure, and these simple modifications all fall within the protection scope of this disclosure.

[0160] It should also be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction. In order to avoid unnecessary repetition, this disclosure will not describe the various possible combinations separately.

[0161] Furthermore, various different embodiments of this disclosure can be combined in any way, as long as they do not violate the spirit of this disclosure, they should also be regarded as the content disclosed in this disclosure.

Claims

1. A method for analyzing the video behavior of non-caged broilers under cold stress in a target detection view angle, characterized in that, The method includes: The video to be detected is segmented to obtain multiple frames of images to be detected. The video to be detected is a video collected from the target detection perspective of the non-cage-raised broiler chickens. The target detection model is used to process each frame of the image to be detected to obtain the detection results corresponding to each frame of the image to be detected. The detection results include the botnet detection box, the non-specific behavior of the botnet selected by the botnet detection box, the position information of the botnet detection box, and the size information of the botnet detection box. Based on the non-specific behaviors of chickens in the detection results corresponding to each frame of the image to be detected, the number and frequency of each non-specific behavior of chickens during the shooting time of the video to be detected are determined. Based on the chicken detection boxes in the detection results corresponding to each frame of the image to be detected, the crowding degree of chickens during the shooting time of the video to be detected is determined. Based on the position information of the chicken detection boxes in the detection results corresponding to each frame of the image to be detected, the position distribution of chickens during the shooting time of the video to be detected is determined. Based on the number and frequency of various non-specific behaviors of broilers during the video recording time, the crowding degree of the broilers, and the location distribution of the broilers, the cold stress of broilers in the non-cage environment is determined. The target detection model is obtained by training a single-stage detection model, which includes an input network, a backbone network, a neck network, and a head network. The backbone network includes a first convolutional module, a second convolutional module, a first C2F module, a first simAM attention mechanism module, a third convolutional module, a second C2F module, a second simAM attention mechanism module, a fourth convolutional module, a third simAM attention mechanism module, a fifth convolutional module, a fourth simAM attention mechanism module, and an SPPF module, which are connected in sequence. The outputs of the second simAM attention mechanism module, the third simAM attention mechanism module, and the fourth simAM attention mechanism module serve as the three inputs of the neck network, respectively. The intermediate results output by the head network include detection boxes and corresponding confidence scores. The single-stage detection model also includes a post-processing network connected to the output of the head network. The post-processing network is used to perform the following processing on the intermediate results output by the head network: The detection boxes are sorted according to their confidence scores. The confidence of the other detection boxes except for the candidate detection box with the highest confidence is calculated by the following calculation formula : Where M represents the candidate detection box. This represents the xth other detection box. Let represent the intersection-union ratio (IoU) between the candidate detection box and the x-th other detection box. Represents the first intermediate parameter. Indicates the center point of other detection boxes. Indicates the center point of the candidate detection box. This represents the square of the distance between the center point of the x-th other detection box and the center point of the candidate detection box. It represents the square of the diagonal length of the smallest bounding rectangle of the x-th other detection box and the candidate detection box; Delete other detection boxes whose confidence Sx is greater than the preset confidence threshold, and determine the remaining other detection boxes as the detection boxes. Return to the execution step: sort the detection boxes according to the confidence of each detection box until there are no more remaining detection boxes. The candidate detection boxes determined in each loop are selected as the detection boxes output by the single-stage detection model.

2. The method according to claim 1, characterized in that, The step of determining the crowding level of chickens during the video recording time based on the chicken detection bounding boxes in the detection results corresponding to each frame of the image to be detected includes: For any frame of the image to be detected, determine the number of times the boundaries of the chicken detection boxes in that frame of the image to be detected intersect. Based on the correspondence between the number of boundary intersections and the density, the density corresponding to the number of boundary intersections of the chicken detection box in the image to be detected in that frame is determined; Based on the density of each frame of the image to be detected, the crowding level of the broilers during the video recording time is determined.

3. The method according to claim 1, characterized in that, The location information of the broiler detection box is the coordinates of the geometric center point of the broiler detection box. Determining the location distribution of broilers within the time frame of the video to be detected based on the location information of the broiler detection boxes in the detection results corresponding to each frame of the image to be detected includes: Based on the geometric center coordinates of the botnet detection boxes included in each frame of the image to be detected, the distribution ratio of the botnet detection boxes in the preset area of ​​the image during the shooting time of the video to be detected is statistically analyzed.

4. The method according to claim 1, characterized in that, The target detection model is trained through the following steps: Obtain a sample dataset, wherein each sample data in the sample dataset includes an image of broiler chickens in a cage-free environment collected from the perspective of target detection and the corresponding annotation information of the image. The annotation information includes non-specific behavior labels of broiler chickens in the image and a rectangular box used to select broiler chickens. The single-stage detection model is trained using the sample dataset to obtain the target detection model.

5. The method according to any one of claims 1-4, characterized in that, The intersection-union ratio (CUC) between the candidate detection box and the xth other detection box is obtained by... The third version of the strategy is calculated.

6. A device for analyzing the cold stress video behavior of cage-free broiler chickens from the perspective of target detection, characterized in that, The device includes: The image segmentation module is used to segment the video to be detected to obtain multiple frames of images to be detected. The video to be detected is a video collected from the non-cage-raised environment of broilers from the target detection perspective. The image processing module is used to process each frame of the image to be detected using the target detection model to obtain the detection results corresponding to each frame of the image to be detected. The detection results include the broiler detection box, the non-specific behavior of the broiler selected by the broiler detection box, the position information of the broiler detection box, and the size information of the broiler detection box. The first determining module is used to determine the number and frequency of each non-specific behavior of chickens during the shooting time of the video under test based on the non-specific behaviors of chickens in the detection results corresponding to each frame of the image under test; to determine the crowding degree of chickens during the shooting time of the video under test based on the chicken detection boxes in the detection results corresponding to each frame of the image under test; and to determine the position distribution of chickens during the shooting time of the video under test based on the position information of the chicken detection boxes in the detection results corresponding to each frame of the image under test. The second determining module is used to determine the cold stress of broilers in the non-cage environment based on the number and frequency of various non-specific behaviors of broilers during the video recording time, the crowding degree of the broilers, and the location distribution of the broilers. The target detection model is obtained by training a single-stage detection model, which includes an input network, a backbone network, a neck network, and a head network. The backbone network includes a first convolutional module, a second convolutional module, a first C2F module, a first simAM attention mechanism module, a third convolutional module, a second C2F module, a second simAM attention mechanism module, a fourth convolutional module, a third simAM attention mechanism module, a fifth convolutional module, a fourth simAM attention mechanism module, and an SPPF module, which are connected in sequence. The outputs of the second simAM attention mechanism module, the third simAM attention mechanism module, and the fourth simAM attention mechanism module serve as the three inputs of the neck network, respectively. The intermediate results output by the head network include detection boxes and corresponding confidence scores. The single-stage detection model also includes a post-processing network connected to the output of the head network. The post-processing network is used to perform the following processing on the intermediate results output by the head network: The detection boxes are sorted according to their confidence scores. The confidence scores of all bounding boxes except the candidate bounding box with the highest confidence score are calculated using the following formula. : Where M represents the candidate detection box. This represents the xth other detection box. Let represent the intersection-union ratio (IoU) between the candidate detection box and the x-th other detection box. Represents the first intermediate parameter. Indicates the center point of other detection boxes. Indicates the center point of the candidate detection box. This represents the square of the distance between the center point of the x-th other detection box and the center point of the candidate detection box. It represents the square of the diagonal length of the smallest bounding rectangle of the x-th other detection box and the candidate detection box; Delete other detection boxes whose confidence Sx is greater than the preset confidence threshold, and determine the remaining other detection boxes as the detection boxes. Return to the execution step: sort the detection boxes according to the confidence of each detection box until there are no more remaining detection boxes. The candidate detection boxes determined in each loop are selected as the detection boxes output by the single-stage detection model.

7. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the method described in any one of claims 1-5.

8. An electronic device, characterized in that, include: A memory on which computer programs are stored; A processor for executing the computer program in the memory to implement the steps of the method according to any one of claims 1-5.

Images