Method for identifying behavior of lion-head goose based on EML-SlowFast

Abstract

The invention discloses an EML-SlowFast-based lion-head goose behavior identification method, and the method comprises the steps: replacing a 3DResNet module in a backbone network of a Slow path of a SlowFast model with an ECAbcheck module through an EML-SlowFast model, thereby improving the extraction capability of the model for the static features of lion-head goose behaviors; besides, a 3DResNet module in a backbone network of a Fast path of the SlowFast model is replaced by an LGLE module, the LGLE module can extract local features and global features of lion-head goose behaviors at the same time, the modeling capability of the model for long-range spatial-temporal features of the lion-head goose behaviors is enhanced, the model can more effectively capture the timeliness of the lion-head goose behaviors, and the robustness of the lion-head goose behaviors is improved. And the capability of extracting dynamic characteristics of lion-head goose behaviors by the model is enhanced. Through the mode, the lion-head goose behavior identification method can accurately and effectively identify lion-head goose behaviors in an actual breeding scene, has the characteristics of high identification precision, high universality and high speed, and provides technical support for intelligent management of a lion-head goose farm.

Description

Technical Field

[0001] The present invention relates to the technical field of computer vision, and particularly to a method for identifying the behaviors of Lionhead geese based on EML-SlowFast. Background Art

[0002] The Lionhead goose originated in Raoping County, Chaozhou City, Guangdong Province. It is the largest meat goose breed in China and one of the largest meat goose breeds in the world, and is a national key germplasm protection resource. The characteristics of the Lionhead goose are large size, tolerance to roughage, fast growth rate, strong stress resistance, low feed consumption, and excellent meat quality. In addition, the Lionhead goose is rich in various essential amino acids required for human growth and development, and its linolenic acid and linoleic acid contents are higher than those of chickens, which is deeply loved by consumers and provides huge economic benefits. In 2021, the output value of the Lionhead goose industrial chain exceeded 3.5 billion yuan, becoming a model of the Chinese goose industry; by May 2022, the annual slaughter volume of Lionhead geese in Shantou City reached approximately 8.1 million, and the stable breeding volume was approximately 3 million.

[0003] Although the Lionhead goose breeding industry shows a rapid development trend, the application of intensive breeding models and new technologies such as "fence spraying" also brings challenges to disease prevention and control. Moreover, due to problems such as the weak investment ability of farmers and poor feeding conditions, it is difficult to prevent and control diseases, and the problem of bio-security is prominent. And the behavior of animals can usually directly or indirectly serve as an indicator of their overall health and well-being. When animals are sick, feel unwell or are in a physiological stage, they will show unique behavioral characteristics. Therefore, the monitoring of animal behavior helps to detect and prevent animal diseases early. In recent years, the combination of computer vision and deep learning has attracted wide attention in the academic community for animal behavior monitoring and analysis. The method for monitoring animal behavior based on computer vision has the characteristics of non-contact, non-stress, high efficiency, etc., and thus plays an important role in precision animal husbandry.

[0004] The application of the 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 the behaviors of Lionhead geese. Therefore, it is necessary to provide a method for identifying the behaviors of Lionhead geese based on computer vision. Summary of the Invention

[0005] (I) Technical Problems to be Solved

[0006] Aiming at the deficiencies of the prior art, the present invention provides a method for identifying the behaviors of Lionhead geese based on EML-SlowFast, which can solve the above technical problems.

[0007] (II) Technical Solutions

[0008] To solve the above technical problems, the present invention provides the following technical solutions: A method for identifying the behavior of lionhead geese based on EML-SlowFast, comprising the following steps:

[0009] S1. Collection of lionhead goose video data: Use a video collection device to collect lionhead goose video data from a lionhead goose farm;

[0010] S2. Preprocessing of video data: Preprocess the lionhead goose video data;

[0011] S3. Creation of a lionhead goose behavior dataset: Make the preprocessed lionhead goose video data into a lionhead goose behavior dataset, where the lionhead goose behavior dataset is divided into a training set, a validation set, and a test set;

[0012] S4. Construction of the EML-SlowFast model: Construct the EML-SlowFast model based on the SlowFast model;

[0013] S5. Training of the EML-SlowFast model: Input the training set and the validation set into the EML-SlowFast model for model training and validation respectively to obtain a trained lionhead goose behavior recognition model;

[0014] S6: Lionhead goose behavior recognition: Input the test set into the trained lionhead goose behavior recognition model to obtain the behavior recognition result of the lionhead goose.

[0015] Preferably, in step S1, the lionhead goose farm includes a fenced breeding scenario and an outdoor activity scenario.

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

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

[0018] S31. Input the preprocessed lionhead goose video data into a trained object detection model to detect lionhead goose targets;

[0019] S32. Track the lionhead goose targets detected in sub-step S31, crop and save the lionhead goose targets belonging to the same object ID in the image frame sequence to obtain the image frame sequence data of the lionhead goose;

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

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

[0022] S35. Perform data augmentation on the test set;

[0023] S36. Make the validation set, the test set, and the test set after data augmentation into a lionhead goose behavior dataset in the UCF101 format.

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

[0025] Preferably, step S4 includes: replacing the 3D ResNet module in the backbone network of the Slow path of the SlowFast model with an ECA bottleneck module, and replacing the 3D ResNet module in the backbone network of the Fast path of the SlowFast model with an LGLE module.

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

[0027]

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

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

[0030]

[0031] 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 convolution kernel of 1, σ 1 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 the convolution represents the size of the convolution kernel, ⊕ represents the element-wise addition operation, represents the element-wise multiplication operation.

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

[0033] F' = BN(σ2 (Conv 1×1×1 (F))) (5)

[0034] F' local = BN(σ 2 {DWConv 3×3×3 (F')}) (6)

[0035] F' global = BN(σ 2 {DWConv 7×7×7 (F')}) (7)

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

[0037] F” exp = BN(σ 2 {Conv 1×1×1 (F”)}) (9)

[0038] F” comp = BN(σ 2 {Conv 1×1×1 (F” exp )}) (10)

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

[0040] F out = σ 3 {BN(Conv 1×1×1 (F” d ))} (12)

[0041] Among them, σ 2 represents the GELU activation function, and σ 3 represents the ReLU activation function.

[0042] (III) Beneficial Effects

[0043] Compared with the prior art, the present invention provides a method for identifying the behaviors of Shitou geese based on EML-SlowFast, which has the following beneficial effects: The present invention conducts the identification of the behaviors of Shitou geese based on the EML-SlowFast model. In this 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 model's ability to extract the static features of the behaviors of Shitou geese. In addition, the present invention also replaces the 3DResNet module in the backbone network of the Fast path of the SlowFast model with the LGLE module. This LGLE module can extract both the local features and the global features of the behaviors of Shitou geese, enhancing the model's ability to model the long-range spatio-temporal features of the behavioral actions of Shitou geese, enabling the model to more effectively capture the temporality of the behavioral actions of Shitou geese, and enhancing the model's ability to extract the dynamic features of the behaviors of Shitou geese. Through the above method, the present invention can accurately and effectively identify the behaviors of Shitou geese in the actual breeding scenario, featuring high recognition accuracy, strong versatility, and fast speed, providing technical support for the intelligent management of Shitou goose farms. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 is the flowchart of the steps of a method for identifying the behaviors of Shitou geese based on EML-SlowFast according to the present invention;

[0045] Figure 2 are the images of different Shitou goose farms collected in the embodiment of the present invention;

[0046] Figure 3 is the flowchart for making the behavior dataset of Shitou geese according to the present invention;

[0047] Figure 4 is an example diagram in the UCF101 format according to the present invention;

[0048] Figure 5 is an example diagram of the content structure of the annotation file according to the present invention;

[0049] Figure 6 is the structure diagram of the EML-SlowFast model according to the present invention;

[0050] Figure 7 is the structure diagram of the ECAbneck module according to the present invention;

[0051] Figure 8 is the structure diagram of the LGLE module according to the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0052] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0053] The present invention provides a method for identifying the behavior of Shitou geese based on EML-SlowFast, including the following steps:

[0054] S1. Collection of Shitou goose video data: Use video collection devices such as digital cameras or monitors to collect Shitou goose video data from Shitou goose farms.

[0055] In step S1, the collected scenarios are preferably diversified. The Shitou goose farm preferably includes at least a fenced breeding scenario and an outdoor activity scenario, etc., so that the present invention can effectively detect Shitou geese in different complex Shitou goose farms such as fenced breeding scenarios and outdoor activity scenarios.

[0056] S2. Preprocessing of video data: Preprocess the Shitou goose video data.

[0057] Specifically, in step S2, the preprocessing includes manually selecting the Shitou goose video data. The selected video data preferably meets the following requirements: (1) Each video contains 5-20 Shitou goose targets; (2) The Shitou goose video data with clear and definite Shitou goose behavior actions; (3) There is preferably no occlusion between Shitou goose individuals; (4) The video data preferably covers different scenarios and includes different Shitou goose behavior actions. Through the above video data preprocessing, the present invention can more effectively and accurately identify different behavior actions of Shitou geese in different scenarios.

[0058] A total of 191 segments of Shitou goose video data are selected in this embodiment, and the duration of each segment of video data ranges from 5 to 20 minutes.

[0059] S3. Making of Shitou goose behavior data set: Make the preprocessed Shitou goose video data into a Shitou goose behavior data set, where the Shitou goose behavior data set is divided into a training set, a validation set and a test set.

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

[0061] S31. Input the preprocessed Shitou goose video data into a trained object detection model to detect Shitou goose targets; preferably, the above object detection model can adopt the YOLOv8 model of the prior art. In addition, this sub-step S31 can also adopt other object detection models of the prior art to detect Shitou goose targets in the Shitou goose video data, and no excessive limitation is made here.

[0062] S32. Preferably, use the ByteTrack algorithm to track the lionhead goose targets detected in sub-step S31, crop the lionhead goose targets belonging to the same object ID in the image frame sequence and save them locally to obtain the image frame sequence data of the lionhead goose. ByteTrack is an object tracking algorithm in the prior art and belongs to the TBD (Tracking By Detection) method. The main idea of ByteTrack is to predict the position of each trajectory in the current frame through Kalman filtering, and then associate all the trajectories with the detection boxes to adjust the mean and variance of the Kalman predictor.

[0063] S33. Manually classify the image frame sequence data obtained in sub-step S32 by behavior. The lionhead goose can specifically include the following five types of 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 this behavior to obtain the original dataset of the 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. In this embodiment, they respectively contain 1,610, 632, and 724 pieces of data, and each piece of data is an image sequence ranging from 16 to 128 frames.

[0065] S35. There may be an imbalance in the data distribution in the original dataset, that is, the data corresponding to some lionhead goose behaviors is less, such as the Resting behavior. Therefore, data augmentation is performed on the data in the test set. Preferably, the data augmentation methods include horizontal flipping, vertical flipping, color inversion, Gaussian blur, and / or pepper noise of the image; after data augmentation, the number of data in the training set changes from the original 1,610 pieces of data to 2,019 pieces of data.

[0066] S36. Make the validation set, the test set, and the test set after data augmentation into a lionhead goose behavior dataset in the UCF101 format.

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

[0068] S4. Construct the EML-SlowFast model: Construct the EML-SlowFast model based on the SlowFast model.

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

[0070] And the present invention constructs the EML-SlowFast model based on the SlowFast model. Specifically, as Figure 6As shown, step S4 includes: replacing the 3D ResNet module in the backbone network of the Slow pathway of the SlowFast model with the ECAbneck module, and replacing the 3D ResNet module in the backbone network of the Fast pathway of the SlowFast model with the LGLE module, while keeping other modules of the SlowFast model unchanged to construct the EML-SlowFast model.

[0071] Specifically, the structure of the above ECAbneck module is as Figure 7 shown. The ECAbneck module includes the following processing procedures: (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 in terms of the number of channels through this convolutional layer, and the channel information is compressed and fused to obtain the feature map F squ . (2) Then, the feature map F squ passes through a 3×3×3 DWConv convolutional layer and a 1×1×1 3D convolutional layer to extract features from the feature map F squ and adjust the number of channels of the feature map to obtain the feature map F extr , where the 3×3×3 DWConv convolutional layer consists of a 3×3×3 depthwise separable convolution, a BatchNorm layer, and an hSwish activation function, and the 1×1×1 3D convolutional layer consists of a 1×1×1 3D convolution and a BatchNorm layer; further, the feature map F extr is processed through the channel attention module ECA3D to learn the channel information. The structure of the ECA3D module is as Figure 7 (a) shown, and the calculation formula of the ECA3D module is as shown in the following formula (1). (3) Finally, a 1×1×1 3D convolution and a BatchNorm layer are used to adjust the number of channels of the input feature map F, and then an element-wise addition operation is performed with the feature map obtained in the previous step (2) to obtain the output feature map F out .

[0072] Specifically, the calculation formula of the ECAbneck module is as shown in the following formulas (1)-(4):

[0073]

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

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

[0076]

[0077] Among them, 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 kernel size of 1, σ 1 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 the convolution represents the size of the convolution kernel, ⊕ represents the element-wise addition operation, represents the element-wise multiplication operation.

[0078] In addition, the structure of the above LGLE module is as Figure 8 shown. The LGLE module includes the following processing procedures: (1) First, the input feature map F passes through a 3D convolution layer with a kernel size of 1×1×1, which consists of a 3D convolution with a kernel size of 1×1×1, the GELU activation function, and the BatchNorm layer. Through this convolution layer, the number of channels of the feature map F is compressed, and the channel information is fused to obtain the feature map F'. (2) Second, the feature map F' is processed in parallel through a 3×3×3 DWConv layer and a 7×7×7 DWConv layer. The 3×3×3 DWConv layer is used to extract local features to obtain the local feature map, focusing on the detailed information of the behavior of the lionhead goose; while the 7×7×7 DWConv layer is used to extract global features to obtain the global feature map, thereby enhancing the long-range modeling ability of the module; in this way, the LGLE module can capture the correlation between information in a longer spatio-temporal channel, which is beneficial for the model to learn the spatio-temporality of the lionhead goose behavior. (3) Further, as shown in the following formula (8), the above local feature map and global feature map are fused, and then a residual connection is made with the input feature map F to obtain the feature map F". Then, the feature map F" expands the number of channels of the feature map four times through a 1×1×1 convolution layer, and then restores the number of channels of the feature map through another 1×1×1 convolution layer; this "expansion-compression" operation enhances the interaction between channels, thereby promoting the in-depth integration of features in the feature map. (4) Finally, a 3×3×3 DWConv layer and a 1×1×1 convolution layer are applied to the feature map to further learn the features and obtain the output feature map F out .

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

[0080] F' = BN(σ 2 (Conv 1×1×1 (F))) (5)

[0081] F' local = BN(σ 2 {DWConv 3×3×3 (F')}) (6)

[0082] F' global = BN(σ 2 {DWConv 7×7×7 (F')}) (7)

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

[0084] F” exp = BN(σ 2 {Conv 1×1×1 (F”)}) (9)

[0085] F” comp = BN(σ 2 {Conv 1×1×1 (F” exp )}) (10)

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

[0087] F out = σ 3 {BN(Conv 1×1×1 (F” d ))} (12)

[0088] Among them, σ 2 represents the GELU activation function, and σ 3 represents the ReLU activation function. The meanings of other parameters can be referred to the relevant descriptions of the above formulas (1)-(4), which will not be elaborated here.

[0089] S5. Train the EML-SlowFast model: Input the training set and the validation set into the EML-SlowFast model respectively for model training and validation to obtain a trained lionhead goose behavior recognition model, that is, a trained EML-SlowFast model for lionhead goose behavior recognition.

[0090] Specifically, in this step S5, model training parameters are set, preferably including: the resolution of the input image sequence is 224×224, the batch size is 16, the number of training iterations is 300 times, the optimizer adopts 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 cosine annealing strategy is used to update the learning rate. Then, the training set and validation set in the above-mentioned lionhead goose behavior dataset are input into the constructed EML-SlowFast model for model training and validation, and a trained lionhead goose behavior recognition model is obtained. Further, the weight parameters of the lionhead goose behavior recognition model can be saved as a file in the.pt format.

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

[0092] Compared with the prior art, the present invention provides a lionhead goose behavior recognition method based on EML-SlowFast, which has the following beneficial effects: The present invention conducts lionhead goose behavior recognition based on the EML-SlowFast model. The EML-SlowFast model replaces the 3DResNet module in the backbone network of the Slow path of the SlowFast model with an ECAbneck module, thereby improving the model's ability to extract static features of lionhead goose behavior; in addition, the present invention also replaces the 3DResNet module in the backbone network of the Fast path of the SlowFast model with an LGLE module. This LGLE module can extract both local features and global features of lionhead goose behavior, enhancing the model's ability to model long-range spatio-temporal features of lionhead goose behavior actions, enabling the model to more effectively capture the temporality of lionhead goose behavior actions, and enhancing the model's ability to extract dynamic features of lionhead goose behavior. By the above method, the present invention improves the accuracy of lionhead goose behavior recognition in the breeding scenario. The present invention can accurately and effectively recognize the behavior of lionhead geese in the actual breeding scenario, and has the characteristics of high recognition accuracy, strong versatility, and fast speed, providing technical support for the intelligent management of lionhead goose farms.

[0093] It should be noted that the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, such that a process, method, article or apparatus comprising a series of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article or apparatus. Without further limitation, an element defined by the phrase "comprising an..." does not exclude the presence of additional identical elements in the process, method, article or apparatus comprising said element.

[0094] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for identifying lion-headed goose behavior based on EML-SlowFast, characterized in that: The following steps are involved: S1. Lionhead goose video data collection: using a video collection device to collect Lionhead goose video data from a Lionhead goose breeding farm; S2, video data preprocessing: preprocessing the lion-head goose video data; S3, lion-headed goose behavior data set preparation: the lion-headed goose video data after preprocessing is prepared into a lion-headed goose behavior data set, wherein the lion-headed goose behavior data set is divided into a training set, a validation set and a test set; S4. Construct EML-SlowFast model: Construct EML-SlowFast model based on SlowFast model; 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-head 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.

2. The method for identifying lion-headed goose behavior based on EML-SlowFast according to claim 1, characterized in that: In the step S1, the lionhead goose farm includes a fence breeding scene and an outdoor activity scene.

3. The method for identifying lion-headed goose behavior based on EML-SlowFast according to claim 1, characterized in that: In the 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 geese targets and with clear and definite lion-headed goose behaviors and movements.

4. The method for identifying lion-headed goose behavior based on EML-SlowFast according to claim 1, characterized in that: The step S3 comprises the following sub-steps: S31, inputting the pre-processed lion-head goose video data into a trained target detection model to detect the lion-head goose target; S32, tracking the lion-headed goose target detected in the sub-step S31, cropping and saving the lion-headed goose targets belonging to the same object ID in the image frame sequence, so as to obtain the image frame sequence data of the lion-headed goose; S33, classifying the image frame sequence data obtained in the sub-step S32 by behavior, and saving the image frame sequence corresponding to each behavior as a data instance in a folder corresponding to the behavior, so as to obtain an original data set of lion-headed goose behaviors; S34, dividing the original data set obtained in the sub-step S33 into the training set, the validation set and the test set; S35, performing data enhancement on the test set; S36. The validation set, the test set, and the test set after data enhancement are made into the lion-headed goose behavior dataset in UCF101 format.

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

6. The method for identifying lion-headed goose behavior based on EML-SlowFast according to claim 1, characterized in that: The step S4 includes: replacing the 3DResNet module in the backbone network in the Slow path of the SlowFast model with the ECAbneck module, and replacing the 3DResNet module in the backbone network in the Fast path of the SlowFast model with the LGLE module.

7. The method for identifying lion-headed goose behavior based on EML-SlowFast according to claim 6, characterized in that: The calculation formula of the ECAbneck module is shown in the following formulas (1)-(4): 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) Among them, F is the feature map, σ represents the Sigmo id activation function, GAP represents the global average pooling operation, Conv1d represents the 1D convolution with a convolution kernel of 1, σ1 represents the hSwish activation function, BN represents the BatchNorm layer, Conv represents the 3D convolution, DWConv represents the 3D depth-separable convolution, and the subscript of the convolution represents the size of the convolution kernel. represents the element-by-element addition operation, Represents an element-wise multiplication operation.

8. The method for identifying lion-headed goose behavior based on EML-SlowFast according to claim 7, characterized in that: The calculation formula of the LGLE module is shown in the following equations (5)-(12): 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) F out =σ3{BN(Conv 1×1×1 (F” d ))} (12) Among them, σ2 represents the GELU activation function and σ3 represents the ReLU activation function.

Images