A lion-head goose behavior recognition method based on EML-SlowFast

Abstract

The application discloses a lion-headed goose behavior recognition method based on an EML-SlowFast, wherein the EML-SlowFast model of the application replaces the 3DResNet module in the trunk network of the Slow path of the SlowFast model with an ECAbneck module, thereby improving the extraction capability of the model for the static features of the lion-headed goose behavior; in addition, the application also replaces the 3DResNet module in the trunk network of the Fast path of the SlowFast model with an LGLE module, the LGLE module can extract the local features and the global features of the lion-headed goose behavior at the same time, and the modeling capability of the model for the long-range space-time features of the lion-headed goose behavior is enhanced, so that the model can more effectively capture the time of the lion-headed goose behavior, and the extraction capability of the model for the dynamic features of the lion-headed goose behavior is enhanced. In the above manner, the application can accurately and effectively recognize the lion-headed goose behavior in the actual breeding scene, has the characteristics of high recognition precision, strong generality and fast speed, and provides technical support for realizing intelligent management of the lion-headed goose breeding farm.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer vision, in particular to a lion-headed goose behavior recognition method based on EML-SlowFast. BACKGROUND

[0002] The lion-headed goose is originally from Raoping County, Chaozhou City, Guangdong Province, and is the largest meat goose breed in China and one of the largest meat goose breeds in the world. It is a national key germplasm protection resource. The lion-headed goose is characterized by large size, tolerance to coarse feed, fast growth rate, strong stress tolerance, low feed consumption, and excellent meat quality. In addition, the lion-headed goose is rich in various essential amino acids needed for human growth and development, and its linolenic acid and linoleic acid content is higher than that of chicken, which is deeply loved by consumers and provides huge economic benefits. In 2021, the lion-headed goose industry chain output value exceeded 3.5 billion yuan, becoming a model of China's goose industry; by May 2022, the annual output of lion-headed geese in Shantou City reached about 8.1 million, and the stable breeding capacity was about 3 million.

[0003] Although the lion-headed goose breeding industry is showing a rapid development trend, the application of intensive breeding mode and "fence spraying" and other new technologies has also brought challenges to disease prevention and control. Moreover, due to the weak investment capacity of breeders, poor feeding conditions and other problems, it is difficult to prevent and control diseases, and biological safety problems are prominent. Animal behavior can often be used as an indicator of overall health and well-being, directly or indirectly. When animals are sick, feel unwell, or are in a physiological stage, they will exhibit unique behavioral characteristics. Therefore, monitoring animal behavior can help to detect and prevent animal diseases early. In recent years, the combination of computer vision and deep learning has attracted widespread attention in the academic community in the field of animal behavior monitoring and analysis. Animal behavior monitoring methods based on computer vision have the characteristics of non-contact, no pressure, and high efficiency, and therefore play an important role in precision animal husbandry.

[0004] The application of automatic behavior recognition technology based on computer vision can improve the efficiency of poultry breeding production and management. However, there is currently a lack of an effective method for identifying lion-headed goose behavior. Therefore, it is necessary to provide a lion-headed goose behavior recognition method based on computer vision. SUMMARY

[0005] (I) Technical problems to be solved

[0006] In view of the deficiencies in the prior art, the present application provides a lion-headed goose behavior recognition method based on EML-SlowFast, which can solve the above technical problems.

[0007] (II) Technical solutions

[0008] To solve the above technical problems, the present application provides the following technical solutions: a lion-headed goose behavior recognition method based on EML-SlowFast, comprising the following steps:

[0009] S1, lion-headed goose video data acquisition: using a video acquisition device to acquire lion-headed goose video data from a lion-headed goose breeding farm;

[0010] S2, video data preprocessing: preprocessing the lion-headed goose video data;

[0011] S3, lion-headed goose behavior dataset making: making the preprocessed lion-headed goose video data into a lion-headed goose behavior dataset, wherein the lion-headed goose behavior dataset is divided into a training set, a validation set and a test set;

[0012] S4, constructing an EML-SlowFast model: constructing an EML-SlowFast model based on a SlowFast model;

[0013] S5, training the EML-SlowFast model: inputting the training set and the validation set into the EML-SlowFast model for model training and validation respectively, to obtain a trained lion-headed goose behavior recognition model;

[0014] S6: lion-headed goose behavior recognition: inputting the test set into the trained lion-headed goose behavior recognition model to obtain the behavior recognition result of the lion-headed goose.

[0015] Preferably, in step S1, the lion-headed goose breeding farm includes a fence breeding scene and an outdoor activity scene.

[0016] Preferably, in step S2, the preprocessing includes manually selecting the lion-headed goose video data to retain the lion-headed goose video data containing 5-20 lion-headed goose targets and clear and explicit lion-headed goose behavior actions.

[0017] Preferably, step S3 includes the following sub-steps:

[0018] S31, inputting the preprocessed lion-headed goose video data into a trained target detection model to detect the lion-headed goose target;

[0019] S32, tracking the lion-headed goose target detected in sub-step S31, cropping and saving the lion-headed goose target belonging to the same object ID in the image frame sequence to obtain the image frame sequence data of the lion-headed goose;

[0020] S33, classifying the image frame sequence data obtained in sub-step S32 by behavior, and saving the image frame sequence corresponding to each behavior as a data instance to the folder corresponding to the behavior to obtain the original dataset of the lion-headed goose behavior;

[0021] S34, dividing the original data set obtained in sub-step S33 into a training set, a validation set and a test set;

[0022] S35, performing data augmentation on the test set;

[0023] S36, preparing the validation set, the test set and the data-augmented test set into a lion-headed goose behavior data set in UCF101 format.

[0024] Preferably, in sub-step S35, the data augmentation includes image horizontal flipping, vertical flipping, color inversion, Gaussian blur and / or pepper noise.

[0025] Preferably, step S4 comprises: replacing the 3DResNet module in the backbone network in the Slow path of the SlowFast model with an ECAbneck module, and replacing the 3DResNet module in the backbone network in the Fast path of the SlowFast model with an LGLE module.

[0026] Preferably, the calculation formula of the ECAbneck module is as shown in the following formulas (1)-(4):

[0027]

[0028] F squ =σ1(BN(Conv 1×1×1 (F))) (2)

[0029] F extr =BN(Conv 1×1×1 (σ1(BN(DWConv 3×3×3 (F squ ))))) (3)

[0030]

[0031] wherein F is a feature map, σ represents a Sigmoid activation function, GAP represents a global average pooling operation, Conv1d represents a 1D convolution with a convolution kernel of 1, σ1 represents an hSwish activation function, BN represents a BatchNorm layer, Conv represents a 3D convolution, DWConv represents a 3D depth separable convolution, the subscript of convolution represents the size of the convolution kernel, and represents an element-wise addition operation, represents an element-wise multiplication operation.

[0032] Preferably, the calculation formula of the LGLE module is as shown in the following formulas (5)-(12):

[0033] F'=BN(σ2(Conv 1×1×1 (F))) (5)

[0034] F' local = BN(σ2{DWConv 3×3×3 (F'))} (6)

[0035] F' global = BN(σ2{DWConv 7×7×7 (F'))} (7)

[0036] F'' = Concat(F' local ,F' global ) + F (8)

[0037] F' exp = BN(σ2{Conv 1×1×1 (F'')} ) (9)

[0038] F' comp = BN(σ2{Conv 1×1×1 (F' exp )} ) (10)

[0039] F' d = σ3{BN(DWConv 3×3×3 (F' comp )} ) (11)

[0040] F out = σ3{BN(Conv 1×1×1 (F d )} ) (12)

[0041] wherein σ2 represents a GELU activation function, and σ3 represents a ReLU activation function.

[0042] (Three) beneficial effects

[0043] Compared with the prior art, the application provides a lion-headed goose behavior recognition method based on EML-SlowFast, which has the following beneficial effects: the lion-headed goose behavior recognition is based on the EML-SlowFast model, the 3DResNet module in the backbone network of the Slow path of the SlowFast model is replaced by the ECAbneck module, thereby improving the extraction capability of the model for the static features of the lion-headed goose behavior; in addition, the 3DResNet module in the backbone network of the Fast path of the SlowFast model is replaced by the LGLE module, the LGLE module can extract the local features and global features of the lion-headed goose behavior at the same time, enhances the modeling capability of the model for the long-range spatiotemporal features of the lion-headed goose behavior action, and makes the model more effectively capture the time of the lion-headed goose behavior action, thereby enhancing the extraction capability of the model for the dynamic features of the lion-headed goose behavior. Through the above-mentioned mode, the lion-headed goose behavior in the actual breeding scene can be accurately and effectively recognized, and the lion-headed goose behavior recognition method based on EML-SlowFast has the characteristics of high recognition precision, strong generality and fast speed, and provides technical support for realizing intelligent management of the lion-headed goose breeding farm. BRIEF DESCRIPTION OF DRAWINGS

[0044] Figure 1 The step flowchart of the lion-headed goose behavior recognition method based on EML-SlowFast of the application;

[0045] Figure 2 The different lion-headed goose breeding farm images collected in the embodiment of the application;

[0046] Figure 3 The lion-headed goose behavior dataset production flowchart of the application;

[0047] Figure 4 The example image in the UCF101 format of the application;

[0048] Figure 5 The content structure example image of the annotation file of the application;

[0049] Figure 6 The EML-SlowFast model structure diagram of the application;

[0050] Figure 7 The ECAbneck module structure diagram of the application;

[0051] Figure 8 The LGLE module structure diagram of the application. DETAILED DESCRIPTION

[0052] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of the present application.

[0053] The present application provides a lion-headed goose behavior recognition method based on EML-SlowFast, comprising the following steps:

[0054] S1, lion-headed goose video data acquisition: using a digital camera or a video acquisition device such as a monitor to acquire lion-headed goose video data from a lion-headed goose farm.

[0055] In step S1, the collected scenes are preferably diversified, and the lion-headed goose farm preferably includes at least a fence breeding scene and an outdoor activity scene, etc., so that the present application can effectively detect lion-headed geese in different complex lion-headed goose breeding farms such as fence breeding scenes and outdoor activity scenes.

[0056] S2, video data preprocessing: preprocessing the lion-headed goose video data.

[0057] Specifically, in step S2, the preprocessing includes manually selecting lion-headed goose video data, and the selected video data preferably meets the following requirements: (1) each video contains 5-20 lion-headed goose targets; (2) lion-headed goose video data with clear and explicit lion-headed goose behavior actions; (3) there is no occlusion between lion-headed goose individuals as much as possible; (4) video data covers different scenes and contains different lion-headed goose behavior actions as much as possible. Through the above video data preprocessing, the present application can more effectively and accurately recognize different behavior actions of lion-headed geese in different scenes.

[0058] In this embodiment, 191 pieces of lion-headed goose video data were selected, and each piece of video data had a duration of 5-20 minutes.

[0059] S3, lion-headed goose behavior dataset making: making the preprocessed lion-headed goose video data into a lion-headed goose behavior dataset, wherein the lion-headed goose behavior dataset is divided into a training set, a validation set and a test set.

[0060] Preferably, step S3 includes the following sub-steps:

[0061] S31, inputting the preprocessed lion-headed goose video data into a trained target detection model to detect lion-headed goose targets; preferably, the above target detection model can adopt a YOLOv8 model of the prior art. In addition, this sub-step S31 can also adopt other target detection models of the prior art to detect lion-headed goose targets in the lion-headed goose video data, which is not limited here.

[0062] S32, preferably, the lionhead goose target detected in sub-step S31 is tracked using the ByteTrack algorithm, and the lionhead goose targets belonging to the same object ID in the image frame sequence are cropped and saved locally to obtain the image frame sequence data of the lionhead goose. ByteTrack is a target tracking algorithm in the prior art, which belongs to the TBD (Tracking By Detection) method. The main idea of ByteTrack is to predict the position of each track in the current frame through Kalman filtering, and then associate all tracks with the detection box to adjust the mean and variance of the Kalman predictor.

[0063] S33, manually classify the image frame sequence data obtained in sub-step S32 according to behavior, and the lionhead goose can specifically include the following five behaviors: feeding, resting, preening, standing and walking; and save the image frame sequence corresponding to each behavior as a data instance to the folder corresponding to the behavior to obtain the original dataset of lionhead goose behavior.

[0064] S34, divide the original dataset obtained in sub-step S33 into a training set, a validation set and a test set, and the embodiment respectively corresponds to 1610, 632 and 724 data containing data, each data being an image sequence of 16 to 128 frames.

[0065] S35, there may be data distribution imbalance in the original dataset, i.e. some lionhead goose behaviors correspond to less data, such as the resting behavior, so the data in the test set is data enhanced. Preferably, the data enhancement method includes image horizontal flip, vertical flip, color inversion, Gaussian blur and / or pepper noise; after data enhancement, the data of the training set is changed from the original 1610 data to 2019 data.

[0066] S36, the validation set, the test set and the data enhanced test set are made into a lionhead goose behavior dataset according to the UCF101 format.

[0067] The UCF101 format is as follows Figure 4As shown, the lion-headed goose behavior dataset contains a training set directory train, a validation set directory val and a test set directory test, and corresponding three annotation files. Taking the training set directory train as an example, the val and test directories are the same: there are five behavior directories under train, which are feeding behavior Feeding, preening behavior Preening, resting behavior Resting, standing behavior Standing and walking behavior Walking. Each behavior directory contains corresponding behavior data, and each data is named in order from "1". The data directory contains image sequences, and the image sequences are named in order in a format similar to "00001.jpg". The content structure of the annotation file is as shown Figure 5 As shown, each data is in a row, the first element of each row is the file path corresponding to the data, the second element is the number of pictures contained in the image sequence of the data, and the third element is the behavior classification of the data, 0, 1, 2, 3 and 4 correspond to feeding behavior Feeding, preening behavior Preening, resting behavior Resting, standing behavior Standing and walking behavior Walking respectively.

[0068] S4, constructing an EML-SlowFast model: constructing an EML-SlowFast model based on a SlowFast model.

[0069] The SlowFast model is a dual-channel model for video recognition tasks in the prior art, which is proposed by Facebook AI Research (FAIR). Its design concept is derived from the observation of the redundancy of information between different frames in a video. By using a dual-path network structure to process fast and slow frame sequences respectively, key information in the video can be effectively captured. The SlowFast model adopts a dual-channel structure to process fast and slow paths respectively. The fast path processes low-resolution, high-frame-rate video sequences to capture fast-changing motion information, and the slow path processes high-resolution, low-frame-rate video sequences to obtain more detailed information. The two paths interact through lateral connections to complete the understanding task of the video together.

[0070] And the EML-SlowFast model is constructed on the basis of the SlowFast model, specifically, as shown in Figure 6As shown, step S4 includes: replacing the 3DResNet module in the backbone network of the Slow pathway of the SlowFast model with the ECAbneck module, and replacing the 3DResNet module in the backbone network of the Fast pathway of the SlowFast model with the LGLE module, while leaving the other modules of the SlowFast model unchanged, to construct the EML-SlowFast model.

[0071] Specifically, the structure of the ECAbneck module described above is as follows: Figure 7 As shown, the ECAbneck module includes the following processing steps: (1) First, the input feature map F (i.e., the image of the lion-headed goose) passes through a 1×1×1 3D convolutional layer, which consists of a 1×1×1 3D convolution, a BatchNorm layer, and an hSwish activation function. The feature map F is compressed by this convolutional layer to compress the number of channels and fuse the channel information to obtain the feature map F. squ (2) Next, feature map F squ The feature map F is processed by a 3×3×3 DWConv convolutional layer and a 1×1×1 3D convolutional layer. squ Feature extraction is performed, and the number of channels in the feature map is adjusted to obtain the feature map F. extr The 3×3×3 DWConv convolutional layer consists of a 3×3×3 depthwise separable convolution, a BatchNorm layer, and an hSwish activation function, while the 1×1×1 3D convolutional layer consists of a 1×1×1 3D convolution and a BatchNorm layer; furthermore, the feature map F extr The channel attention module (ECA3D) processes the information to learn the channel information. The structure of the ECA3D module is as follows: Figure 7 As shown in (a), the calculation formula of the ECA3D module is shown in equation (1). (3) Finally, the input feature map F is adjusted by a 1×1×1 3D convolution and a BatchNorm layer, and then it is added element-wise to the feature map obtained in the previous step (2) to obtain the output feature map F. out .

[0072] Specifically, the calculation formulas for the ECAbneck module are shown in equations (1)-(4) below:

[0073]

[0074] F squ =σ1(BN(Conv 1×1×1 (F))) (2)

[0075] F extr = BN(Conv 1×1×1 (σ1(BN(DWConv 3×3×3 (F squ ))))) (3)

[0076]

[0077] where F is the feature map, σ represents the Sigmoid activation function, GAP represents the global average pooling operation, Conv1d represents the 1D convolution with a 1 convolution kernel, σ1 represents the hSwish activation function, BN represents the BatchNorm layer, Conv represents the 3D convolution, DWConv represents the 3D depthwise separable convolution, the subscript of convolution represents the size of the convolution kernel, and represents the element-wise addition operation, represents the element-wise multiplication operation.

[0078] In addition, the structure of the LGLE module is as shown in Figure 8 the following processing procedures: (1) First, the input feature map F passes through a 1x1x1 3D convolution layer composed of a 1x1x1 3D convolution, a GELU activation function and a BatchNorm layer. The feature map F is compressed in the number of channels through this convolution layer, and the channel information is fused to obtain the feature map F'. (2) Second, the feature map F' is processed in parallel through a 3x3x3 DWConv layer and a 7x7x7 DWConv layer. The 3x3x3 DWConv layer is used to extract local features to obtain a local feature map, focusing on the detailed information of the lion-headed goose behavior; while the 7x7x7 DWConv layer is used to extract global features to obtain a global feature map, thereby enhancing the long-range modeling capability of the module; in this way, the LGLE module can capture the correlation between information in a longer spatiotemporal channel, which is conducive to the model learning the spatiotemporal nature of the lion-headed goose behavior. (3) Further, as shown in the following formula (8), the local feature map and the global feature map are fused, and then connected with the input feature map F in residual connection to obtain the feature map F". Then, the feature map F" passes through a 1x1x1 convolution layer to expand the number of feature map channels by four, and then passes through another 1x1x1 convolution layer to restore the number of feature map channels; this "expansion-compression" operation enhances the interaction between channels, thereby promoting the deep integration of features in the feature map. (4) Finally, a 3x3x3 DWConv layer and a 1x1x1 convolution layer are applied to the feature map for further learning of the features to obtain the output feature map F out .

[0079] Specifically, the calculation formula of the LGLE module is as shown in the following formulas (5)-(12):

[0080] F' = BN(σ2(Conv 1×1×1 (F))) (5)

[0081] F' local = BN(σ2{DWConv 3×3×3 (F')}) (6)

[0082] F' global = BN(σ2{DWConv 7×7×7 (F')}) (7)

[0083] F” = Concat(F' local ,F' global )+F (8)

[0084] F” exp = BN(σ2{Conv 1×1×1 (F”)}) (9)

[0085] F” comp = BN(σ2{Conv 1×1×1 (F” exp )}) (10)

[0086] F” d = σ3{BN(DWConv 3×3×3 (F” comp ))} (11)

[0087] F out = σ3{BN(Conv 1×1×1 (F” d ))} (12)

[0088] Wherein, σ2 represents a GELU activation function, σ3 represents a ReLU activation function, and other parameter meanings can refer to the related descriptions of the above formulas (1)-(4), which will not be repeated here.

[0089] S5, training the EML-SlowFast model: inputting the training set and the validation set into the EML-SlowFast model to respectively perform model training and validation, so as to obtain a trained lion-headed goose behavior recognition model, that is, to obtain a trained EML-SlowFast model for lion-headed goose behavior recognition.

[0090] Specifically, in this step S5, the model training parameters are set, preferably including: the resolution of the input image sequence is 224x224, the batch size is 16, the number of training iterations is 300 times, the optimizer uses the Stochastic Gradient Descent (SGD) algorithm, the initial learning rate is 0.05, the momentum is 0.9, the weight decay is 0.001, and the learning rate is updated using the cosine annealing strategy. Then the training set and the validation set in the above lion-headed goose behavior dataset are input into the EML-SlowFast model built to perform model training and validation, and a trained lion-headed goose behavior recognition model is obtained. Further, the weight parameters of the lion-headed goose behavior recognition model can be saved as a file in.pt format.

[0091] S6: Lion-headed goose behavior recognition: input the test set into the trained lion-headed goose behavior recognition model to obtain the behavior recognition result of the lion-headed goose. Specifically, the.pt format file saved in the above step S5 can be loaded into the program corresponding to the lion-headed goose behavior recognition model to open the lion-headed goose behavior recognition model; then the test set in the lion-headed goose behavior dataset is input into the trained lion-headed goose behavior recognition model to obtain the behavior recognition result of the lion-headed goose.

[0092] Compared with the prior art, the present application provides a lion-headed goose behavior recognition method based on EML-SlowFast, which has the following beneficial effects: the present application performs lion-headed goose behavior recognition based on the EML-SlowFast model, which replaces the 3DResNet module in the backbone network of the Slow path of the SlowFast model with the ECAbneck module, thereby improving the model's ability to extract the static features of the lion-headed goose behavior; in addition, the present application also replaces the 3DResNet module in the backbone network of the Fast path of the SlowFast model with the LGLE module, which can simultaneously extract the local features and global features of the lion-headed goose behavior, thereby enhancing the model's ability to model the long-range spatiotemporal features of the lion-headed goose behavior, making the model more effectively capture the temporal nature of the lion-headed goose behavior, and enhancing the model's ability to extract the dynamic features of the lion-headed goose behavior. In the above manner, the present application improves the accuracy of lion-headed goose behavior recognition in the breeding scene, and the present application can accurately and effectively recognize the lion-headed goose behavior in the actual breeding scene, has the characteristics of high recognition accuracy, strong generality, and fast speed, and provides technical support for realizing intelligent management of the lion-headed goose breeding farm.

[0093] It is to be understood that the terminology "including", "comprising", or any other variation thereof, is intended to cover a non-exclusive inclusion such that process, method, article, or apparatus that comprises a list of elements does not include only those elements but can also include other elements not expressly listed or inherent to such process, method, article, or apparatus. An element proceeded by "comprises a... " does not, without more constraints, exclude the existence of additional identical elements in the process, method, article, or apparatus that comprises the element.

[0094] While embodiments of the application have been shown and described, it is to be understood that the application is not limited to the details of the embodiments described, since the scope of the application is defined with respect to the appended claims.

Claims

1. A method for recognizing lion-headed geese behavior based on EML-SlowFast, characterized in that, Includes the following steps: S1. Lion Head Goose Video Data Acquisition: Collect Lion Head Goose video data from the Lion Head Goose farm using video acquisition devices; S2. Video data preprocessing: Preprocess the Lion Head Goose video data; S3. Lion Head Goose Behavior Dataset Creation: The preprocessed lion head goose video data is used to create a lion head goose behavior dataset, wherein the lion head goose behavior dataset is divided into a training set, a validation set, and a test set. S4. Constructing the EML-SlowFast model: Constructing the EML-SlowFast model based on the SlowFast model; S5. Training the EML-SlowFast model: Input the training set and validation set into the EML-SlowFast model for training and validation respectively, so as to obtain a trained lion-headed goose behavior recognition model. S6: Lion-headed goose behavior recognition: Input the test set into the trained lion-headed goose behavior recognition model to obtain the lion-headed goose behavior recognition result; Step S4 includes: replacing the 3DResNet module in the backbone network of the Slow path of the SlowFast model with the ECAbneck module, and replacing the 3DResNet module in the backbone network of the Fast path of the SlowFast model with the LGLE module. The calculation formulas for the ECAbneck module are shown in equations (1)-(4) below: F squ =σ1(BN(Conv 1×1×1 (F))) (2) F extr =BN(Conv 1×1×1 (σ1(BN(DWConv 3×3×3 (F squ ))))) (3) Where F represents the feature map, σ represents the Sigmoid activation function, GAP represents global average pooling, Conv1d represents a 1D convolution with a kernel of 1, σ1 represents the hSwish activation function, BN represents a BatchNorm layer, Conv represents a 3D convolution, DWConv represents a 3D depthwise separable convolution, and the subscript of the convolution indicates the size of the convolution kernel. This indicates an element-wise addition operation. F represents the element-wise multiplication operation. out The output feature map of the ECAbneck module; The calculation formulas for the LGLE module are shown in equations (5)-(12) below: F'=BN(σ2(Conv 1×1×1 (F))) (5) F' local =BN(σ2{DWConv 3×3×3 (F')}) (6) F' global =BN(σ2{DWConv 7×7×7 (F')}) (7) F”=Concat(F' local ,F' global )+F (8) F” exp =BN(σ2{Conv 1×1×1 (F”)}) (9) F” comp =BN(σ2{Conv 1×1×1 (F” exp )}) (10) F” d =σ3{BN(DWConv 3×3×3 (F” comp ))} (11) Where σ² represents the GELU activation function and σ³ represents the ReLU activation function. This is the output feature map of the LGLE module.

2. The lion-headed goose behavior recognition method based on EML-SlowFast according to claim 1, characterized in that: In step S1, the Lionhead Goose Farm includes both an enclosed breeding area and an outdoor activity area.

3. The lion-headed goose behavior recognition method based on EML-SlowFast according to claim 1, characterized in that: In step S2, the preprocessing includes manually selecting the lion-headed goose video data to retain video data containing 5-20 lion-headed geese with clear and distinct lion-headed goose behaviors.

4. The lion-headed goose behavior recognition method based on EML-SlowFast according to claim 1, characterized in that, Step S3 includes the following sub-steps: S31. Input the preprocessed lion-headed goose video data into the trained target detection model to detect the lion-headed goose target; S32. Track the lion-headed goose targets detected in sub-step S31, crop and save the lion-headed goose targets belonging to the same object ID in the image frame sequence to obtain the image frame sequence data of the lion-headed goose. S33. Classify the image frame sequence data obtained in sub-step S32 according to behavior, and save the image frame sequence corresponding to each behavior as a data instance into the folder corresponding to that behavior to obtain the original dataset of the Lion Head Goose behavior. S34. Divide the original dataset obtained in sub-step S33 into the training set, the validation set, and the test set; S35. Perform data augmentation on the test set; S36. The validation set, the test set, and the data-enhanced test set are combined into the Lionhead Goose Behavior Dataset according to the UCF101 format.

5. The lion-headed goose behavior recognition method based on EML-SlowFast according to claim 4, characterized in that: In sub-step S35, the data augmentation includes horizontal image flipping, vertical image flipping, color inversion, Gaussian blurring, and / or pepper noise.

Images