Red-crowned crane behavior pattern recognition method and system based on artificial intelligence vision technology

Abstract

The application discloses a red-crowned crane behavior mode recognition method and system based on artificial intelligence vision technology, relates to the technical field of intelligent identification, and comprises the following steps: S1, collecting original video data containing red-crowned crane activities through high-resolution camera equipment arranged in a nature reserve; S2, pre-processing and labeling the original video data: frame splitting at a fixed time interval, removing redundant frames by using a perceptual hashing algorithm, and labeling the boundary box and behavior label of the red-crowned crane individual in the key frame. The red-crowned crane behavior mode automatic recognition method provided by the application significantly improves the efficiency and accuracy of red-crowned crane behavior monitoring by integrating advanced artificial intelligence and computer vision technology, overcomes the shortcomings of traditional manual observation, such as being limited by the environment and being low in efficiency, realizes all-weather and all-around automatic monitoring, and improves the fineness and adaptability of behavior recognition.

Description

Technical Field

[0001] This application belongs to the field of intelligent recognition technology, and specifically relates to a method and system for recognizing the behavior patterns of red-crowned cranes based on artificial intelligence vision technology. Background Technology

[0002] As a rare and endangered species, monitoring the behavioral patterns of the red-crowned crane is a crucial aspect of biodiversity conservation and behavioral research, playing a key role in understanding the species' survival status, ecological adaptability, and habitat quality. Traditional monitoring primarily relies on manual observation and recording, requiring staff to maintain continuous presence in complex environments such as wetlands. However, limitations imposed by weather, terrain, and vegetation obstruction make continuous, all-weather, and comprehensive monitoring difficult, and data collection efficiency and accuracy are easily influenced by the observer's experience. In recent years, the rapid development of computer vision and artificial intelligence technologies has provided new technological pathways for animal behavior recognition. Deep learning models have shown potential in areas such as species classification and posture estimation. However, for specific species like the red-crowned crane with unique behavioral characteristics, specialized behavior recognition technologies are still in the exploratory stage.

[0003] Existing technologies have certain limitations in practical applications: general-purpose behavior recognition models, trained on universal datasets, struggle to accurately capture the unique and complex behaviors of red-crowned cranes, are easily affected by interference in complex wetland environments, and require improved recognition accuracy; most technologies focus on species classification from static images, lacking the ability to process dynamic video data in real time, thus failing to meet the needs of continuous behavior sequence analysis; furthermore, the identification and early warning mechanisms for abnormal behaviors of red-crowned cranes are inadequate, making it difficult to quickly respond to potential risks in the species' survival process. These limitations stem from the scarcity of behavioral data specific to the species and the complexity of application scenarios, rather than inherent flaws in the technologies themselves. To address these limitations, we propose a red-crowned crane behavior pattern recognition method based on artificial intelligence vision technology. Summary of the Invention

[0004] To address the aforementioned technical issues, this paper provides a method for recognizing the behavioral patterns of red-crowned cranes based on artificial intelligence vision technology. This technical solution solves the problems of the general behavior recognition model, which is based on a universal dataset and is difficult to accurately capture the special characteristics of red-crowned crane behavior, and has insufficient recognition accuracy in complex wetland backgrounds; it focuses more on static image species classification and has weak ability to process dynamic videos in real time and analyze continuous behavior sequences; and the identification and early warning mechanism for abnormal behavior of red-crowned cranes is imperfect and it is difficult to quickly respond to potential risks to their survival.

[0005] To achieve the above objectives, the technical solution adopted in this application is as follows: The method for recognizing the behavior patterns of red-crowned cranes based on artificial intelligence vision technology includes the following steps: S1. Collect raw video data containing the activities of red-crowned cranes by deploying high-resolution camera equipment in nature reserves; S2. Preprocess and label the original video data: split the frames at fixed time intervals, use the perceptual hash algorithm to remove redundant frames, and label the red-crowned crane individuals in the key frames with bounding boxes and behavior labels to construct a red-crowned crane behavior dataset with spatiotemporal labels. S3. Based on the red-crowned crane behavior dataset, execute a multi-level data augmentation strategy, including applying spatial transformation to image data and applying temporal velocity change to video sequences, to generate an augmented training dataset. S4. Construct a lightweight spatiotemporal decoupled behavior recognition model and train the model end-to-end using the enhanced training dataset; wherein, in the spatial feature extraction stage of the model, an optimized MobileNetV3 network is used to extract high-level semantic features frame by frame, and in the temporal modeling stage, a multi-layer 1D convolutional network is used to perform multi-scale temporal dynamic modeling of inter-frame feature sequences. S5. Input the video of the red-crowned crane activity to be identified into the trained lightweight spatiotemporal decoupled behavior recognition model, output the recognition result of its behavior pattern, and provide real-time warning of abnormal behavior based on the recognition result.

[0006] Preferably, S1 includes: By relying on fixed camera arrays and portable shooting equipment, multi-view video data of red-crowned crane activity scenes are acquired, raw video streams are collected in real time, and video sequences under multiple time periods and weather conditions are obtained. Preliminary timestamp synchronization and equipment information recording are performed on them. Set the main monitoring viewpoint and obtain the shooting parameters and geographical coordinates of the main monitoring viewpoint; Based on the wetland topography and the activity patterns of red-crowned cranes, the shooting direction and focal length of the auxiliary monitoring perspective were adjusted to cover the blind spots of the main perspective. Calculate the overlap between the secondary monitoring view and the primary monitoring view, and combine the geographic coordinates of the primary monitoring view to convert the overlap into a spatial coverage optimization parameter. Integrate multi-view video streams with spatial coverage parameters to generate synchronous and complementary raw video datasets.

[0007] Preferably, S2 includes: Based on the original video dataset, the original high frame rate video is split into frames at fixed intervals to obtain an initial image sequence. The image fingerprints of adjacent frames in the initial image sequence are calculated using a perceptual hashing algorithm, and their Hamming distances are also calculated. A similarity threshold is set, and redundant image frames with a Hamming distance less than the similarity threshold are removed, leaving key frames that represent individual red-crowned cranes to form a set of key frames after redundancy removal.

[0008] Preferably, S2 further includes: Based on the redundancy-removed keyframe set, the bounding boxes of individual red-crowned cranes in each frame image are annotated using the VIA annotation script to generate a target detection annotation file. Based on the original video dataset, the video was segmented segment by segment according to behavioral semantics using the VIA annotation script, and a labeling system for five typical behaviors including foraging, standing, walking, flying, and preening was established. By integrating the object detection annotation files and behavior labeling system, each video segment and its corresponding keyframe are assigned behavior category and timestamp information; By combining all the labeled video segments and keyframes, a dataset of red-crowned crane behavior with spatiotemporal labels is constructed.

[0009] Preferably, S3 includes: Spatial transformation is applied to the image data in the red-crowned crane behavior dataset. By performing random rotation transformation, motion blur processing, scale transformation and horizontal flipping in a random combination, the multi-angle posture, motion blur and visual morphology of red-crowned cranes at different distances and orientations in the natural environment are simulated to generate spatially enhanced image data. Temporal velocity variations are applied to video sequences in the red-crowned crane behavior dataset. By linearly adjusting the video playback speed within the range of 0.8 to 1.2 times while maintaining the integrity of the behavioral semantics, the differences in behavioral rhythms among different red-crowned crane individuals are simulated, and temporally enhanced video sequences are generated. All the spatially enhanced image data and the temporally enhanced video sequences are integrated to generate the enhanced training dataset.

[0010] Preferably, S4 includes: The model input is set to a normalized video data tensor with batch size of B, number of channels C of 3, temporal length T of 16 frames, and spatial resolution H×W of 224×224; In the spatial feature extraction stage, the MobileNetV3 network is used to process the input video tensor frame by frame, mapping each frame image to a 960-dimensional feature vector, and the output temporal feature is represented as (B,T,960). In the temporal modeling stage, a multi-layer 1D convolutional network is used to process the temporal feature representation and output the fused temporal feature (B,512). In the classification decision-making stage, a fully connected layer and a Softmax activation function are used to process the fused temporal features, and the output probability distributions corresponding to five behaviors: foraging, standing, walking, flying, and preening are generated.

[0011] Preferably, S4 further includes: Depthwise separable convolution is used; the h-swish activation function is defined as:

[0012] In the formula, s is the input tensor, and ReLU6 is the ReLU activation function; using the SE attention mechanism, its calculation formula is:

[0013] In the formula, For the Sigmoid function, h-swish function, GAP is global average pooling. , This is the weight matrix of the fully connected layer. This is the input feature map.

[0014] Preferably, S4 further includes: Parallel 1D convolutional layers with kernel sizes of 3, 5, and 7 are used, and the output features of each layer are concatenated and fused by channel concatenation. Using a temporal attention mechanism, for features Calculate temporal weights using 1D convolution:

[0015] In the formula, F represents the input temporal feature with dimensions T×C. Let be the feature vector of frame t, Conv1D be a one-dimensional convolution operation, and softmax be a normalized exponential function. Let be the attention weight for frame t; Using a global temporal pooling layer, the calculation formula is as follows:

[0016] In the formula, λ is the weighting coefficient. For one-dimensional max pooling, For one-dimensional average pooling, The fused global temporal features; Residual connections and batch normalization are used between the 1D convolutional layers.

[0017] Preferably, S5 includes: The video of the red-crowned crane activity to be identified is preprocessed and adjusted to a standardized input tensor with a batch size of 1, a temporal length of 16 frames T, and a spatial resolution of 224×224 H×W. The standardized input tensor is input into the trained lightweight spatiotemporal decoupled behavior recognition model, and forward propagation is performed sequentially through the spatial feature extraction stage, the temporal modeling stage, and the classification decision stage. Obtain the probability distribution of five types of behaviors output in the classification decision stage, wherein the five types of behaviors include foraging, standing, walking, flying, and preening; The behavior category corresponding to the highest probability value in the probability distribution is selected as the behavior pattern recognition result of the current input video segment.

[0018] Preferably, S5 further includes: The behavioral pattern recognition results are continuously acquired and recorded to form a real-time behavioral sequence of red-crowned crane activities; Define rules for identifying abnormal behavior, including: the continuous occurrence of specific high-risk behaviors, significant disruption of normal behavior rhythms, or significant deviation of preset normal behavior patterns from their temporal distribution. When the real-time behavior sequence triggers any of the abnormal behavior discrimination rules, an abnormal behavior warning signal is automatically generated and sent. The warning signal triggers an audio-visual alarm on the protected area management platform and includes information on the type of abnormal behavior, the time of occurrence, and the geographical location.

[0019] Based on the same inventive concept, this application also provides a red-crowned crane behavior pattern recognition system based on artificial intelligence vision technology, including: The acquisition unit is used to acquire raw video data containing the activities of red-crowned cranes using high-resolution camera equipment deployed in nature reserves; The preprocessing unit is used to preprocess and label the original video data: split the frames at fixed time intervals, use the perceptual hash algorithm to remove redundant frames, and label the red-crowned crane individuals in the key frames with bounding boxes and behavior labels to construct a red-crowned crane behavior dataset with spatiotemporal labels. The data augmentation unit is used to execute a multi-level data augmentation strategy based on the red-crowned crane behavior dataset, including applying spatial transformation to image data and applying temporal velocity change to video sequences to generate an augmented training dataset. The recognition model training unit is used to construct a lightweight spatiotemporally decoupled behavior recognition model and to train the model end-to-end using the enhanced training dataset. In the spatial feature extraction stage of the model, an optimized MobileNetV3 network is used to extract high-level semantic features frame by frame, and in the temporal modeling stage, a multi-layer 1D convolutional network is used to perform multi-scale temporal dynamic modeling of the inter-frame feature sequence. The identification and early warning unit is used to input the video of the red-crowned crane activity to be identified into the trained lightweight spatiotemporal decoupled behavior recognition model, output the recognition result of its behavior pattern, and provide real-time early warning of abnormal behavior based on the recognition result.

[0020] Preferably, the acquisition unit includes: By relying on fixed camera arrays and portable shooting equipment, multi-view video data of red-crowned crane activity scenes are acquired, raw video streams are collected in real time, and video sequences under multiple time periods and weather conditions are obtained. Preliminary timestamp synchronization and equipment information recording are performed on them. Set the main monitoring viewpoint and obtain the shooting parameters and geographical coordinates of the main monitoring viewpoint; Based on the wetland topography and the activity patterns of red-crowned cranes, the shooting direction and focal length of the auxiliary monitoring perspective were adjusted to cover the blind spots of the main perspective. Calculate the overlap between the secondary monitoring view and the primary monitoring view, and combine the geographic coordinates of the primary monitoring view to convert the overlap into a spatial coverage optimization parameter. Integrate multi-view video streams with spatial coverage parameters to generate synchronous and complementary raw video datasets.

[0021] Preferably, the preprocessing and annotation of the original video data includes: splitting the video into frames at fixed time intervals and removing redundant frames using a perceptual hashing algorithm, comprising: Based on the original video dataset, the original high frame rate video is split into frames at fixed intervals to obtain an initial image sequence. The image fingerprints of adjacent frames in the initial image sequence are calculated using a perceptual hashing algorithm, and their Hamming distances are also calculated. A similarity threshold is set, and redundant image frames with a Hamming distance less than the similarity threshold are removed, leaving key frames that represent individual red-crowned cranes to form a set of key frames after redundancy removal.

[0022] Preferably, the step of annotating the bounding boxes and behavior labels of individual red-crowned cranes in the keyframes to construct a red-crowned crane behavior dataset with spatiotemporal labels includes: Based on the redundancy-removed keyframe set, the bounding boxes of individual red-crowned cranes in each frame image are annotated using the VIA annotation script to generate a target detection annotation file. Based on the original video dataset, the video was segmented segment by segment according to behavioral semantics using the VIA annotation script, and a labeling system for five typical behaviors including foraging, standing, walking, flying, and preening was established. By integrating the object detection annotation files and behavior labeling system, each video segment and its corresponding keyframe are assigned behavior category and timestamp information; By combining all the labeled video segments and keyframes, a dataset of red-crowned crane behavior with spatiotemporal labels is constructed.

[0023] Preferably, the enhanced data unit includes: Spatial transformation is applied to the image data in the red-crowned crane behavior dataset. By performing random rotation transformation, motion blur processing, scale transformation and horizontal flipping in a random combination, the multi-angle posture, motion blur and visual morphology of red-crowned cranes at different distances and orientations in the natural environment are simulated to generate spatially enhanced image data. Temporal velocity variations are applied to video sequences in the red-crowned crane behavior dataset. By linearly adjusting the video playback speed within the range of 0.8 to 1.2 times while maintaining the integrity of the behavioral semantics, the differences in behavioral rhythms among different red-crowned crane individuals are simulated, and temporally enhanced video sequences are generated. All the spatially enhanced image data and the temporally enhanced video sequences are integrated to generate the enhanced training dataset.

[0024] Preferably, the construction of the lightweight spatiotemporal decoupled behavior recognition model includes: The model input is set to a normalized video data tensor with batch size of B, number of channels C of 3, temporal length T of 16 frames, and spatial resolution H×W of 224×224; In the spatial feature extraction stage, the MobileNetV3 network is used to process the input video tensor frame by frame, mapping each frame image to a 960-dimensional feature vector, and the output temporal feature is represented as (B,T,960). In the temporal modeling stage, a multi-layer 1D convolutional network is used to process the temporal feature representation and output the fused temporal feature (B,512). In the classification decision-making stage, a fully connected layer and a Softmax activation function are used to process the fused temporal features, and the output probability distributions corresponding to five behaviors: foraging, standing, walking, flying, and preening are generated.

[0025] Preferably, the spatial feature extraction stage of the model uses an optimized MobileNetV3 network to extract high-level semantic features frame by frame, including: Depthwise separable convolution is used; the h-swish activation function is defined as:

[0026] In the formula, s is the input tensor, and ReLU6 is the ReLU activation function; using the SE attention mechanism, its calculation formula is:

[0027] In the formula, For the Sigmoid function, h-swish function, GAP is global average pooling. , This is the weight matrix of the fully connected layer. This is the input feature map.

[0028] Preferably, the temporal modeling stage uses a multi-layer 1D convolutional network to perform multi-scale temporal dynamic modeling of inter-frame feature sequences, including: Parallel 1D convolutional layers with kernel sizes of 3, 5, and 7 are used, and the output features of each layer are concatenated and fused by channel concatenation. Using a temporal attention mechanism, for features Calculate temporal weights using 1D convolution:

[0029] In the formula, F represents the input temporal feature with dimensions T×C. Let be the feature vector of frame t, Conv1D be a one-dimensional convolution operation, and softmax be a normalized exponential function. Let be the attention weight for frame t; Using a global temporal pooling layer, the calculation formula is as follows:

[0030] In the formula, λ is the weighting coefficient. For one-dimensional max pooling, For one-dimensional average pooling, The fused global temporal features; Residual connections and batch normalization are used between the 1D convolutional layers.

[0031] Preferably, the step of inputting the video of the red-crowned crane activity to be identified into the trained lightweight spatiotemporally decoupled behavior recognition model and outputting the recognition result of its behavior pattern includes: The video of the red-crowned crane activity to be identified is preprocessed and adjusted to a standardized input tensor with a batch size of 1, a temporal length of 16 frames T, and a spatial resolution of 224×224 H×W. The standardized input tensor is input into the trained lightweight spatiotemporal decoupled behavior recognition model, and forward propagation is performed sequentially through the spatial feature extraction stage, the temporal modeling stage, and the classification decision stage. Obtain the probability distribution of five types of behaviors output in the classification decision stage, wherein the five types of behaviors include foraging, standing, walking, flying, and preening; The behavior category corresponding to the highest probability value in the probability distribution is selected as the behavior pattern recognition result of the current input video segment.

[0032] Preferably, the step of providing real-time early warning for abnormal behavior based on the identification results includes: The behavioral pattern recognition results are continuously acquired and recorded to form a real-time behavioral sequence of red-crowned crane activities; Define rules for identifying abnormal behavior, including: the continuous occurrence of specific high-risk behaviors, significant disruption of normal behavior rhythms, or significant deviation of preset normal behavior patterns from their temporal distribution. When the real-time behavior sequence triggers any of the abnormal behavior discrimination rules, an abnormal behavior warning signal is automatically generated and sent. The warning signal triggers an audio-visual alarm on the protected area management platform and includes information on the type of abnormal behavior, the time of occurrence, and the geographical location.

[0033] Compared with the prior art, the beneficial effects of the present invention are as follows: The red-crowned crane behavior pattern recognition method and system proposed in this application significantly improves the efficiency and accuracy of red-crowned crane behavior monitoring by integrating advanced artificial intelligence and computer vision technologies. It overcomes the shortcomings of traditional manual observation, which is subject to large environmental limitations and is inefficient, and realizes all-weather, all-round automatic monitoring. Its personalized recognition model is optimized for the unique behavioral characteristics of red-crowned cranes, improving the precision and adaptability of behavior recognition. It can analyze dynamic video data in real time, capture subtle changes in behavior patterns, and provide timely warnings of abnormal behavior, providing strong technical support for protected area management. The fusion of deep learning algorithms and bird behavioral data not only enhances the scientific nature of behavior recognition but also provides new perspectives and tools for behavioral research, promoting the advancement of wildlife conservation technology. Attached Figure Description

[0034] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0035] Figure 1 The flowchart of the method for recognizing the behavior patterns of red-crowned cranes based on artificial intelligence vision technology provided in this application is shown. Figure 2 The duration distribution of the behavior recognition samples provided in this application is shown; Figure 3 An organizational diagram of the red-crowned crane behavior pattern recognition system based on artificial intelligence vision technology provided in this application is shown. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0037] Example 1 This application provides a method for recognizing the behavioral patterns of red-crowned cranes based on artificial intelligence vision technology, such as... Figure 1 This includes the following steps: S1. Collect raw video data containing the activities of red-crowned cranes by deploying high-resolution camera equipment in nature reserves; S1 includes: By relying on fixed camera arrays and portable shooting equipment, multi-view video data of red-crowned crane activity scenes are acquired, raw video streams are collected in real time, and video sequences under multiple time periods and weather conditions are obtained. Preliminary timestamp synchronization and equipment information recording are performed on them. Set the main monitoring viewpoint and obtain the shooting parameters and geographical coordinates of the main monitoring viewpoint; Based on the wetland topography and the activity patterns of red-crowned cranes, the shooting direction and focal length of the auxiliary monitoring perspective were adjusted to cover the blind spots of the main perspective. Calculate the overlap between the secondary monitoring view and the primary monitoring view, and combine the geographic coordinates of the primary monitoring view to convert the overlap into a spatial coverage optimization parameter. Integrate multi-view video streams with spatial coverage parameters to generate synchronous and complementary raw video datasets.

[0038] When using it, please refer to the above: In existing technologies, monitoring of red-crowned crane behavior often relies on a single fixed camera or manual patrols, which is insufficient to fully cover the activity areas in the complex terrain of wetlands, easily leading to blind spots. Furthermore, video data integration is difficult due to heterogeneous equipment and asynchronous timing, affecting the accuracy and efficiency of behavior analysis. This step, by deploying a fixed camera array and portable equipment, achieves multi-view, multi-time period data acquisition. By combining primary and secondary viewpoint coordination and field-of-view overlap optimization, a synchronous and spatially complementary original video dataset is generated. This provides high-quality, high-coverage input for subsequent artificial intelligence visual analysis, improving the reliability and comprehensiveness of red-crowned crane behavior pattern recognition.

[0039] S2. Preprocess and label the original video data: split the frames at fixed time intervals, use the perceptual hash algorithm to remove redundant frames, and label the red-crowned crane individuals in the key frames with bounding boxes and behavior labels to construct a red-crowned crane behavior dataset with spatiotemporal labels. S2 includes: Based on the original video dataset, the original high frame rate video is split into frames at fixed intervals to obtain an initial image sequence. The image fingerprints of adjacent frames in the initial image sequence are calculated using a perceptual hashing algorithm, and their Hamming distances are also calculated. A similarity threshold is set, and redundant image frames with a Hamming distance less than the similarity threshold are removed, leaving key frames that represent individual red-crowned cranes to form a set of key frames after redundancy removal.

[0040] S2 further includes: Based on the redundancy-removed keyframe set, the bounding boxes of individual red-crowned cranes in each frame image are annotated using the VIA annotation script to generate a target detection annotation file. Based on the original video dataset, the video was segmented segment by segment according to behavioral semantics using the VIA annotation script, and a labeling system for five typical behaviors including foraging, standing, walking, flying, and preening was established. By integrating the object detection annotation files and behavior labeling system, each video segment and its corresponding keyframe are assigned behavior category and timestamp information; By combining all the labeled video segments and keyframes, a dataset of red-crowned crane behavior with spatiotemporal labels is constructed.

[0041] When using it, please refer to the above: In existing technologies, preprocessing of red-crowned crane behavior data largely relies on manual selection and annotation of keyframes, which suffers from low efficiency, strong subjectivity, and inconsistent annotation standards. Furthermore, the large volume of video data and numerous redundant frames lead to wasted storage and computing resources, and the incomplete behavior labeling system also affects the training effect of subsequent recognition models. This step utilizes fixed-interval frame splitting and perceptual hashing for redundancy removal to automatically extract red-crowned crane keyframes with significant variations, greatly improving data quality and processing efficiency. Combined with the VIA annotation tool, the bounding boxes and five typical behavior labels are systematically annotated, constructing a behavior dataset with unified spatiotemporal labels. This lays a solid foundation for training a high-precision, highly generalized artificial intelligence behavior recognition model.

[0042] S3. Based on the red-crowned crane behavior dataset, execute a multi-level data augmentation strategy, including applying spatial transformation to image data and applying temporal velocity change to video sequences, to generate an augmented training dataset. S3 includes: Spatial transformation is applied to the image data in the red-crowned crane behavior dataset. By performing random rotation transformation, motion blur processing, scale transformation and horizontal flipping in a random combination, the multi-angle posture, motion blur and visual morphology of red-crowned cranes at different distances and orientations in the natural environment are simulated to generate spatially enhanced image data. Temporal velocity variations are applied to video sequences in the red-crowned crane behavior dataset. By linearly adjusting the video playback speed within the range of 0.8 to 1.2 times while maintaining the integrity of the behavioral semantics, the differences in behavioral rhythms among different red-crowned crane individuals are simulated, and temporally enhanced video sequences are generated. All the spatially enhanced image data and the temporally enhanced video sequences are integrated to generate the enhanced training dataset.

[0043] When using it, please refer to the above: In existing technologies, training deep learning-based red-crowned crane behavior recognition models is often limited by the insufficient size and diversity of the original dataset. Video data from a single perspective and at a fixed speed cannot cover the varying postures, distances, lighting conditions, and individual behavioral rhythms of red-crowned cranes in their natural state, resulting in weak model generalization ability and a tendency to overfit. This step implements a multi-level data augmentation strategy to simulate various visual morphological changes of red-crowned cranes in the spatial dimension and reproduce the behavioral rhythm characteristics of different individuals in the temporal dimension. This significantly expands the diversity and realism of the training data and effectively improves the adaptability and robustness of the subsequent behavior recognition model to complex natural environments.

[0044] S4. Construct a lightweight spatiotemporal decoupled behavior recognition model and train the model end-to-end using the enhanced training dataset; wherein, in the spatial feature extraction stage of the model, an optimized MobileNetV3 network is used to extract high-level semantic features frame by frame, and in the temporal modeling stage, a multi-layer 1D convolutional network is used to perform multi-scale temporal dynamic modeling of inter-frame feature sequences. S4 includes: The model input is set to a normalized video data tensor with batch size of B, number of channels C of 3, temporal length T of 16 frames, and spatial resolution H×W of 224×224; In the spatial feature extraction stage, the MobileNetV3 network is used to process the input video tensor frame by frame, mapping each frame image to a 960-dimensional feature vector, and the output temporal feature is represented as (B,T,960). In the temporal modeling stage, a multi-layer 1D convolutional network is used to process the temporal feature representation and output the fused temporal feature (B,512). In the classification decision-making stage, a fully connected layer and a Softmax activation function are used to process the fused temporal features, and the output probability distributions corresponding to five behaviors: foraging, standing, walking, flying, and preening are generated.

[0045] S4 further includes: Depthwise separable convolution is used; the h-swish activation function is defined as:

[0046] In the formula, s is the input tensor, and ReLU6 is the ReLU activation function; using the SE attention mechanism, its calculation formula is:

[0047] In the formula, For the Sigmoid function, h-swish function, GAP is global average pooling. , This is the weight matrix of the fully connected layer. This is the input feature map.

[0048] S4 further includes: Parallel 1D convolutional layers with kernel sizes of 3, 5, and 7 are used, and the output features of each layer are concatenated and fused by channel concatenation. Using a temporal attention mechanism, for features Calculate temporal weights using 1D convolution:

[0049] In the formula, F represents the input temporal feature with dimensions T×C. Let be the feature vector of frame t, Conv1D be a one-dimensional convolution operation, and softmax be a normalized exponential function. Let be the attention weight for frame t; Using a global temporal pooling layer, the calculation formula is as follows:

[0050] In the formula, λ is the weighting coefficient. For one-dimensional max pooling, For one-dimensional average pooling, The fused global temporal features; Residual connections and batch normalization are used between the 1D convolutional layers.

[0051] When using it, please refer to the above: In existing technologies, red-crowned crane behavior recognition models often face the contradiction between high computational complexity and poor real-time performance. Traditional 3D convolutional networks have a large number of parameters and slow inference speed, making them difficult to deploy on edge devices. At the same time, single-scale temporal modeling is unable to capture the complex dynamic features of alternating fast and slow behaviors in red-crowned cranes, resulting in limited recognition accuracy. This step constructs a lightweight spatiotemporal decoupling model, uses an optimized MobileNetV3 network to efficiently extract spatial features, and combines multi-scale 1D convolution with temporal attention mechanisms to accurately model dynamic changes in behavior. While significantly reducing the computational load, it enhances the ability to represent behaviors with different rhythms, achieving a balance between accuracy and efficiency. This provides a feasible technical solution for real-time and accurate behavior recognition at the edge of nature reserves.

[0052] S5. Input the video of the red-crowned crane activity to be identified into the trained lightweight spatiotemporal decoupled behavior recognition model, output the recognition result of its behavior pattern, and provide real-time warning of abnormal behavior based on the recognition result.

[0053] S5 includes: The video of the red-crowned crane activity to be identified is preprocessed and adjusted to a standardized input tensor with a batch size of 1, a temporal length of 16 frames T, and a spatial resolution of 224×224 H×W. The standardized input tensor is input into the trained lightweight spatiotemporal decoupled behavior recognition model, and forward propagation is performed sequentially through the spatial feature extraction stage, the temporal modeling stage, and the classification decision stage. Obtain the probability distribution of five types of behaviors output in the classification decision stage, wherein the five types of behaviors include foraging, standing, walking, flying, and preening; The behavior category corresponding to the highest probability value in the probability distribution is selected as the behavior pattern recognition result of the current input video segment.

[0054] The S5 also includes: The behavioral pattern recognition results are continuously acquired and recorded to form a real-time behavioral sequence of red-crowned crane activities; Define rules for identifying abnormal behavior, including: the continuous occurrence of specific high-risk behaviors, significant disruption of normal behavior rhythms, or significant deviation of preset normal behavior patterns from their temporal distribution. When the real-time behavior sequence triggers any of the abnormal behavior discrimination rules, an abnormal behavior warning signal is automatically generated and sent. The warning signal triggers an audio-visual alarm on the protected area management platform and includes information on the type of abnormal behavior, the time of occurrence, and the geographical location.

[0055] When using it, please refer to the above: In existing technologies, red-crowned crane behavior monitoring systems are often limited to post-event analysis, lacking real-time identification and early warning capabilities. This results in a significant lag in response to abnormal behaviors, making it difficult to effectively support the proactive management needs of nature reserves. Furthermore, traditional methods are insufficient in analyzing the dynamic patterns of continuous behavioral sequences, failing to accurately capture subtle changes in behavioral rhythms. This step, by inputting pre-processed video data into a trained lightweight spatiotemporal decoupling model, achieves real-time and accurate identification of five typical red-crowned crane behaviors. Based on predefined anomaly detection rules, the system continuously performs intelligent analysis of behavioral sequences. Once high-risk behaviors or rhythmic disturbances are detected, the system immediately generates and sends an early warning signal containing specific type, time, and location information to the management platform. This transforms the system from passive observation to proactive early warning, significantly improving the real-time and precise management level of red-crowned crane conservation in nature reserves.

[0056] Based on the above steps, the specific implementation method is as follows: Relying on high-resolution fixed camera arrays and portable shooting equipment deployed within the nature reserve, continuous shooting of red-crowned cranes is carried out in the wetland environment under multiple weather conditions, at multiple times, and from multiple perspectives to ensure the capture of real and diverse behavioral samples. The equipment can achieve 24-hour real-time recording, and the camera installation positions are fully combined with the wetland terrain and the activity patterns of red-crowned cranes to completely cover their daily activity scenes.

[0057] After data acquisition, the original high frame rate video was split into 5-frame segments at fixed intervals. Redundant images with a similarity exceeding 95% were removed using a perceptual hashing algorithm, retaining high-quality keyframes that clearly show the red-crowned cranes. Individual bounding boxes were then labeled using VIA scripts, resulting in a dataset containing 5112 high-resolution images of red-crowned cranes. Simultaneously, VIA scripts were used to segment the video segments according to behavioral semantics, establishing a labeling system covering five typical behaviors: foraging, standing, walking, flying, and preening. 1778 valid video segments were retained, including 484 foraging segments, 351 standing segments, 582 walking segments, 187 flying segments, and 174 preening segments, meeting the requirements for temporal modeling.

[0058] To improve the model's generalization ability and robustness, random rotation transformation, motion blur, scale transformation, horizontal flipping, and random combination are used for image data, while speed changes of 0.8-1.2 times are used for video sequence data to simulate the differences in behavioral rhythms of different individuals and enrich the diversity of training data.

[0059] A lightweight behavior recognition framework based on the concept of "spatial-temporal decoupling" is constructed, adopting a "2D+1D" hybrid architecture. First, the original video sequence is standardized and resized to a batch-size-adaptive, 3-channel, 16-frame temporal length, and 224×224 spatial resolution format. Then, it is processed frame by frame through an optimized MobileNetV3 backbone network. Using depthwise separable convolution, h-swish activation function, and SE attention mechanism, each frame image is mapped to a 960-dimensional high-level semantic feature vector to form a temporal feature representation. In the temporal modeling stage, a multi-layer 1D convolutional network is used to capture short-term, medium-term, and long-term temporal dependencies in parallel through 3, 5, and 7 different sizes of convolutional kernels. Combined with temporal attention mechanism and global temporal pooling layer, fused features are extracted. Finally, a fully connected classifier and Softmax activation function are used to output the probability distribution of five types of behavior.

[0060] The experiment was conducted on a Windows 10 device configured with an Intel Core i5-13600KF processor and an NVIDIA GeForce RTX 4080 (16G) graphics card. A joint training framework for detection and behavior recognition was built based on Python and the PyTorch 1.13 framework. Feature cascade transfer, joint loss optimization, and real-time inference acceleration strategies were employed. The dataset was divided into training, validation, and test sets in an 8:1:1 ratio. The initial training parameters for YOLO detection were set as follows: initial learning rate 0.001, momentum 0.937, weight decay coefficient 0.0005, batch size 128, optimizer SGD, and input image size 640×640×3. The initial training parameters for behavior recognition were set as follows: initial learning rate 0.0003, sequence length 16 frames, weight decay coefficient 0.0001, batch size 32, optimizer Adam, and frame image size 224×224×3. Cosine decay was used to adjust the learning rate during training. The multi-task loss function was jointly optimized by weighted combination of detection loss and behavior recognition loss. The model hyperparameters are shown in the table below.

[0061] like Figure 2 The temporal characteristic analysis showed that the average duration of the samples was 1.27 seconds (standard deviation 0.82 seconds), the median duration was 1.0 seconds, and the distribution range was 0.5-3.0 seconds, which is consistent with the natural temporal pattern of bird behavior.

[0062] The model was comprehensively evaluated based on metrics such as precision, recall, F1-Score, accuracy, mAP, and FPS. The final trained model size was 19.5MB, as shown in the table below: Behavioral category data percentage, accuracy, precision, recall, F1 score

[0063] The system achieved an overall accuracy of 94.96% on the test set, with the highest F1 score of 0.954 for flight behavior, 0.938 for preening behavior, and 0.954, 0.958, and 0.942 for walking, foraging, and standing behaviors, respectively. It can accurately identify various behavioral patterns of red-crowned cranes and has real-time processing capabilities, enabling it to promptly capture abnormal behaviors and issue early warnings, thus providing technical support for the protection and behavioral research of red-crowned cranes in nature reserves.

[0064] Example 2 Based on the same inventive concept, this application also provides an automatic recognition system for red-crowned crane behavior patterns based on intelligent vision technology, such as... Figure 3 ,include: The acquisition unit is used to acquire raw video data containing the activities of red-crowned cranes using high-resolution camera equipment deployed in nature reserves; The preprocessing unit is used to preprocess and label the original video data: split the frames at fixed time intervals, use the perceptual hash algorithm to remove redundant frames, and label the red-crowned crane individuals in the key frames with bounding boxes and behavior labels to construct a red-crowned crane behavior dataset with spatiotemporal labels. The data augmentation unit is used to execute a multi-level data augmentation strategy based on the red-crowned crane behavior dataset, including applying spatial transformation to image data and applying temporal velocity change to video sequences to generate an augmented training dataset. The recognition model training unit is used to construct a lightweight spatiotemporally decoupled behavior recognition model and to train the model end-to-end using the enhanced training dataset. In the spatial feature extraction stage of the model, an optimized MobileNetV3 network is used to extract high-level semantic features frame by frame, and in the temporal modeling stage, a multi-layer 1D convolutional network is used to perform multi-scale temporal dynamic modeling of the inter-frame feature sequence. The identification and early warning unit is used to input the video of the red-crowned crane activity to be identified into the trained lightweight spatiotemporal decoupled behavior recognition model, output the recognition result of its behavior pattern, and provide real-time early warning of abnormal behavior based on the recognition result.

[0065] Preferably, the acquisition unit includes: By relying on fixed camera arrays and portable shooting equipment, multi-view video data of red-crowned crane activity scenes are acquired, raw video streams are collected in real time, and video sequences under multiple time periods and weather conditions are obtained. Preliminary timestamp synchronization and equipment information recording are performed on them. Set the main monitoring viewpoint and obtain the shooting parameters and geographical coordinates of the main monitoring viewpoint; Based on the wetland topography and the activity patterns of red-crowned cranes, the shooting direction and focal length of the auxiliary monitoring perspective were adjusted to cover the blind spots of the main perspective. Calculate the overlap between the secondary monitoring view and the primary monitoring view, and combine the geographic coordinates of the primary monitoring view to convert the overlap into a spatial coverage optimization parameter. Integrate multi-view video streams with spatial coverage parameters to generate synchronous and complementary raw video datasets.

[0066] Preferably, the preprocessing and annotation of the original video data includes: splitting the video into frames at fixed time intervals and removing redundant frames using a perceptual hashing algorithm, comprising: Based on the original video dataset, the original high frame rate video is split into frames at fixed intervals to obtain an initial image sequence. The image fingerprints of adjacent frames in the initial image sequence are calculated using a perceptual hashing algorithm, and their Hamming distances are also calculated. A similarity threshold is set, and redundant image frames with a Hamming distance less than the similarity threshold are removed, leaving key frames that represent individual red-crowned cranes to form a set of key frames after redundancy removal.

[0067] Preferably, the step of annotating the bounding boxes and behavior labels of individual red-crowned cranes in the keyframes to construct a red-crowned crane behavior dataset with spatiotemporal labels includes: Based on the redundancy-removed keyframe set, the bounding boxes of individual red-crowned cranes in each frame image are annotated using the VIA annotation script to generate a target detection annotation file. Based on the original video dataset, the video was segmented segment by segment according to behavioral semantics using the VIA annotation script, and a labeling system for five typical behaviors including foraging, standing, walking, flying, and preening was established. By integrating the object detection annotation files and behavior labeling system, each video segment and its corresponding keyframe are assigned behavior category and timestamp information; By combining all the labeled video segments and keyframes, a dataset of red-crowned crane behavior with spatiotemporal labels is constructed.

[0068] Preferably, the enhanced data unit includes: Spatial transformation is applied to the image data in the red-crowned crane behavior dataset. By performing random rotation transformation, motion blur processing, scale transformation and horizontal flipping in a random combination, the multi-angle posture, motion blur and visual morphology of red-crowned cranes at different distances and orientations in the natural environment are simulated to generate spatially enhanced image data. Temporal velocity variations are applied to video sequences in the red-crowned crane behavior dataset. By linearly adjusting the video playback speed within the range of 0.8 to 1.2 times while maintaining the integrity of the behavioral semantics, the differences in behavioral rhythms among different red-crowned crane individuals are simulated, and temporally enhanced video sequences are generated. All the spatially enhanced image data and the temporally enhanced video sequences are integrated to generate the enhanced training dataset.

[0069] Preferably, the construction of the lightweight spatiotemporal decoupled behavior recognition model includes: The model input is set to a standardized video data tensor with batch size of B, number of channels of C, temporal length of T, and spatial resolution of H×W; In the spatial feature extraction stage, the MobileNetV3 network is used to process the input video tensor frame by frame, mapping each frame image to an n-dimensional feature vector, and the output temporal feature is represented as (B,T,n). In the temporal modeling stage, a multi-layer 1D convolutional network is used to process the temporal feature representation and output the fused temporal feature (B,512). In the classification decision-making stage, a fully connected layer and a Softmax activation function are used to process the fused temporal features, and the output probability distributions corresponding to five behaviors: foraging, standing, walking, flying, and preening are generated.

[0070] Preferably, the spatial feature extraction stage of the model uses an optimized MobileNetV3 network to extract high-level semantic features frame by frame, including: Depthwise separable convolution is used; the h-swish activation function is defined as:

[0071] In the formula, s is the input tensor, and ReLU6 is the ReLU activation function; using the SE attention mechanism, its calculation formula is:

[0072] In the formula, For the Sigmoid function, h-swish function, GAP is global average pooling. , This is the weight matrix of the fully connected layer. This is the input feature map.

[0073] Preferably, the temporal modeling stage uses a multi-layer 1D convolutional network to perform multi-scale temporal dynamic modeling of inter-frame feature sequences, including: Parallel 1D convolutional layers with different kernel sizes are used, and the output features of each layer are concatenated and fused through channels. Using a temporal attention mechanism, for features Calculate temporal weights using 1D convolution:

[0074] In the formula, F represents the input temporal feature with dimensions T×C. Let be the feature vector of frame t, Conv1D be a one-dimensional convolution operation, and softmax be a normalized exponential function. Let be the attention weight for frame t; Using a global temporal pooling layer, the calculation formula is as follows:

[0075] In the formula, λ is the weighting coefficient. For one-dimensional max pooling, For one-dimensional average pooling, The fused global temporal features; Residual connections and batch normalization are used between the 1D convolutional layers.

[0076] Preferably, the step of inputting the video of the red-crowned crane activity to be identified into the trained lightweight spatiotemporally decoupled behavior recognition model and outputting the recognition result of its behavior pattern includes: The video of red-crowned crane activity to be identified is preprocessed and adjusted to a standardized input tensor with a batch size of 1, a temporal length of T, and a spatial resolution of H×W. The standardized input tensor is input into the trained lightweight spatiotemporal decoupled behavior recognition model, and forward propagation is performed sequentially through the spatial feature extraction stage, the temporal modeling stage, and the classification decision stage. Obtain the probability distribution of five types of behaviors output in the classification decision stage, wherein the five types of behaviors include foraging, standing, walking, flying, and preening; The behavior category corresponding to the highest probability value in the probability distribution is selected as the behavior pattern recognition result of the current input video segment.

[0077] Preferably, the step of providing real-time early warning for abnormal behavior based on the identification results includes: The behavioral pattern recognition results are continuously acquired and recorded to form a real-time behavioral sequence of red-crowned crane activities; Define rules for identifying abnormal behavior, including: the continuous occurrence of specific high-risk behaviors, significant disruption of normal behavior rhythms, or significant deviation of preset normal behavior patterns from their temporal distribution. When the real-time behavior sequence triggers any of the abnormal behavior discrimination rules, an abnormal behavior warning signal is automatically generated and sent. The warning signal triggers an audio-visual alarm on the protected area management platform and includes information on the type of abnormal behavior, the time of occurrence, and the geographical location.

[0078] The foregoing has shown and described the basic principles, main features, and advantages of this application. Those skilled in the art should understand that this application is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of this application. Various changes and modifications can be made to this application without departing from the spirit and scope thereof, and all such changes and modifications fall within the scope of this application as claimed. The scope of protection claimed by this application is defined by the appended claims and their equivalents.

[0079] Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for recognizing the behavioral patterns of red-crowned cranes based on artificial intelligence vision technology, characterized in that, Includes the following steps: S1. Collect raw video data containing the activities of red-crowned cranes by deploying high-resolution camera equipment in nature reserves; S2. Preprocess and label the original video data: split the frames at fixed time intervals, use the perceptual hash algorithm to remove redundant frames, and label the red-crowned crane individuals in the key frames with bounding boxes and behavior labels to construct a red-crowned crane behavior dataset with spatiotemporal labels. S3. Based on the red-crowned crane behavior dataset, execute a multi-level data augmentation strategy, including applying spatial transformation to image data and applying temporal velocity change to video sequences, to generate an augmented training dataset. S4. Construct a lightweight spatiotemporal decoupled behavior recognition model and train the model end-to-end using the enhanced training dataset; wherein, in the spatial feature extraction stage of the model, an optimized MobileNetV3 network is used to extract high-level semantic features frame by frame, and in the temporal modeling stage, a multi-layer 1D convolutional network is used to perform multi-scale temporal dynamic modeling of inter-frame feature sequences. S5. Input the video of the red-crowned crane activity to be identified into the trained lightweight spatiotemporal decoupled behavior recognition model, output the recognition result of its behavior pattern, and provide real-time warning of abnormal behavior based on the recognition result.

2. The method for recognizing red-crowned crane behavior patterns based on artificial intelligence vision technology according to claim 1, characterized in that, S1 includes: By relying on fixed camera arrays and portable shooting equipment, multi-view video data of red-crowned crane activity scenes are acquired, raw video streams are collected in real time, and video sequences under multiple time periods and weather conditions are obtained. Preliminary timestamp synchronization and equipment information recording are performed on them. Set the main monitoring viewpoint and obtain the shooting parameters and geographical coordinates of the main monitoring viewpoint; Based on the wetland topography and the activity patterns of red-crowned cranes, the shooting direction and focal length of the auxiliary monitoring perspective were adjusted to cover the blind spots of the main perspective. Calculate the overlap between the secondary monitoring view and the primary monitoring view, and combine the geographic coordinates of the primary monitoring view to convert the overlap into a spatial coverage optimization parameter. Integrate multi-view video streams with spatial coverage parameters to generate synchronous and complementary raw video datasets.

3. The method for recognizing red-crowned crane behavior patterns based on artificial intelligence vision technology according to claim 2, characterized in that, S2 includes: Based on the original video dataset, the original high frame rate video is split into frames at fixed intervals to obtain an initial image sequence. The image fingerprints of adjacent frames in the initial image sequence are calculated using a perceptual hashing algorithm, and their Hamming distances are also calculated. A similarity threshold is set, and redundant image frames with a Hamming distance less than the similarity threshold are removed, leaving key frames that represent individual red-crowned cranes to form a set of key frames after redundancy removal.

4. The method for recognizing the behavior patterns of red-crowned cranes based on artificial intelligence vision technology according to claim 3, characterized in that, S2 further includes: Based on the redundancy-removed keyframe set, the bounding boxes of individual red-crowned cranes in each frame image are annotated using the VIA annotation script to generate a target detection annotation file. Based on the original video dataset, the video was segmented segment by segment according to behavioral semantics using the VIA annotation script, and a labeling system for five typical behaviors including foraging, standing, walking, flying, and preening was established. By integrating the object detection annotation files and behavior labeling system, each video segment and its corresponding keyframe are assigned behavior category and timestamp information; By combining all the labeled video segments and keyframes, a dataset of red-crowned crane behavior with spatiotemporal labels is constructed.

5. The method for recognizing red-crowned crane behavior patterns based on artificial intelligence vision technology according to claim 4, characterized in that, S3 includes: Spatial transformation is applied to the image data in the red-crowned crane behavior dataset. By performing random rotation transformation, motion blur processing, scale transformation and horizontal flipping in a random combination, the multi-angle posture, motion blur and visual morphology of red-crowned cranes at different distances and orientations in the natural environment are simulated to generate spatially enhanced image data. Temporal velocity variations are applied to video sequences in the red-crowned crane behavior dataset. By linearly adjusting the video playback speed within the range of 0.8 to 1.2 times while maintaining the integrity of the behavioral semantics, the differences in behavioral rhythms among different red-crowned crane individuals are simulated, and temporally enhanced video sequences are generated. All the spatially enhanced image data and the temporally enhanced video sequences are integrated to generate the enhanced training dataset.

6. The method for recognizing red-crowned crane behavior patterns based on artificial intelligence vision technology according to claim 1, characterized in that, S4 includes: The model input is set to a standardized video data tensor with batch size of B, number of channels of C, temporal length of T, and spatial resolution of H×W; In the spatial feature extraction stage, the MobileNetV3 network is used to process the input video tensor frame by frame, mapping each frame image to an n-dimensional feature vector and outputting a temporal feature representation. In the temporal modeling stage, a multi-layer 1D convolutional network is used to process the temporal feature representation and output fused temporal features. In the classification decision-making stage, a fully connected layer and a Softmax activation function are used to process the fused temporal features, and the output probability distributions corresponding to five behaviors: foraging, standing, walking, flying, and preening are generated.

7. The method for recognizing red-crowned crane behavior patterns based on artificial intelligence vision technology according to claim 6, characterized in that, S4 further includes: Depthwise separable convolution is used; the h-swish activation function is defined as: In the formula, s is the input tensor, and ReLU6 is the ReLU activation function; using the SE attention mechanism, its calculation formula is: In the formula, For the Sigmoid function, h-swish function, GAP is global average pooling. , This is the weight matrix of the fully connected layer. This is the input feature map.

8. The method for recognizing the behavior patterns of red-crowned cranes based on artificial intelligence vision technology according to claim 7, characterized in that, S4 further includes: Parallel 1D convolutional layers with different kernel sizes are used, and the output features of each layer are concatenated and fused through channels. Using a temporal attention mechanism, for features Calculate temporal weights using 1D convolution: In the formula, F represents the input temporal feature with dimensions T×C. Let be the feature vector of frame t, Conv1D be a one-dimensional convolution operation, and softmax be a normalized exponential function. Let be the attention weight for frame t; Using a global temporal pooling layer, the calculation formula is as follows: In the formula, λ is the weighting coefficient. For one-dimensional max pooling, For one-dimensional average pooling, The fused global temporal features; Residual connections and batch normalization are used between the 1D convolutional layers.

9. The method for recognizing red-crowned crane behavior patterns based on artificial intelligence vision technology according to claim 1, characterized in that, S5 includes: The video of red-crowned crane activity to be identified is preprocessed and adjusted to a standardized input tensor with a batch size of 1, a temporal length of T, and a spatial resolution of H×W. The standardized input tensor is input into the trained lightweight spatiotemporal decoupled behavior recognition model, and forward propagation is performed sequentially through the spatial feature extraction stage, the temporal modeling stage, and the classification decision stage. Obtain the probability distribution of five types of behaviors output in the classification decision stage, wherein the five types of behaviors include foraging, standing, walking, flying, and preening; The behavior category corresponding to the highest probability value in the probability distribution is selected as the behavior pattern recognition result of the current input video segment.

10. The method for recognizing the behavior patterns of red-crowned cranes based on artificial intelligence vision technology according to claim 9, characterized in that, The S5 also includes: The behavioral pattern recognition results are continuously acquired and recorded to form a real-time behavioral sequence of red-crowned crane activities; Define rules for identifying abnormal behavior, including: the continuous occurrence of specific high-risk behaviors, significant disruption of normal behavior rhythms, or significant deviation of preset normal behavior patterns from their temporal distribution. When the real-time behavior sequence triggers any of the abnormal behavior discrimination rules, an abnormal behavior warning signal is automatically generated and sent. The warning signal triggers an audio-visual alarm on the protected area management platform and includes information on the type of abnormal behavior, the time of occurrence, and the geographical location.

11. A red-crowned crane behavior pattern recognition system based on artificial intelligence vision technology, characterized in that, include: The acquisition unit is used to acquire raw video data containing the activities of red-crowned cranes using high-resolution camera equipment deployed in nature reserves; The preprocessing unit is used to preprocess and label the original video data: split the frames at fixed time intervals, use the perceptual hash algorithm to remove redundant frames, and label the red-crowned crane individuals in the key frames with bounding boxes and behavior labels to construct a red-crowned crane behavior dataset with spatiotemporal labels. The data augmentation unit is used to execute a multi-level data augmentation strategy based on the red-crowned crane behavior dataset, including applying spatial transformation to image data and applying temporal velocity change to video sequences to generate an augmented training dataset. The recognition model training unit is used to construct a lightweight spatiotemporally decoupled behavior recognition model and to train the model end-to-end using the enhanced training dataset. In the spatial feature extraction stage of the model, an optimized MobileNetV3 network is used to extract high-level semantic features frame by frame, and in the temporal modeling stage, a multi-layer 1D convolutional network is used to perform multi-scale temporal dynamic modeling of the inter-frame feature sequence. The identification and early warning unit is used to input the video of the red-crowned crane activity to be identified into the trained lightweight spatiotemporal decoupled behavior recognition model, output the recognition result of its behavior pattern, and provide real-time early warning of abnormal behavior based on the recognition result.

Images