Panda behavior identification method and system based on large model

Abstract

The invention discloses a panda behavior recognition method and system based on a large model, and relates to the technical field of animal monitoring. The method comprises the following steps: collecting a multi-scene panda video stream, performing frame extraction to obtain a data set, and performing cleaning, image enhancement and SAM visual large model semi-automatic labeling to generate a training set; sAM is used as a backbone network, behavior classification branches are connected, a scene feature auxiliary judgment module is additionally arranged to construct a model, and optimization is carried out through transfer learning; inputting to-be-identified data, accurately segmenting features by the model, outputting a result in combination with scene features, and feeding back supplementary annotations for data with substandard confidence. The system comprises a data acquisition module, a preprocessing module, a model training and optimizing module, a recognition reasoning module and a result output and feedback module. The method is suitable for multiple complex scenes, the recognition accuracy and efficiency are improved, model iterative optimization is achieved, and reliable technical support is provided for panda protection monitoring.

Description

Technical Field

[0001] This invention relates to the field of animal monitoring technology, specifically to a method and system for recognizing giant panda behavior based on a large model. Background Technology

[0002] As a rare and endangered species endemic to my country, monitoring the behavior patterns of giant pandas is a core foundation for species conservation, population reproduction research, and habitat optimization, and is of great significance for formulating scientific conservation strategies. With the development of artificial intelligence and computer vision technologies, traditional monitoring methods relying on manual observation and post-event interpretation are gradually being replaced by intelligent recognition technologies. Deep learning-based giant panda behavior recognition methods have become a key direction for improving monitoring efficiency and reducing human interference. Currently, related technologies have begun to be applied to behavior recognition in single scenarios, whether in captivity or the wild. Through video frame segmentation, feature extraction, and model training, some behaviors are automatically recognized, providing technical support for giant panda conservation efforts. However, this also places higher demands on the technology's adaptability to different scenarios, recognition accuracy, and practicality.

[0003] Existing technologies still have significant limitations and cannot meet the needs of multi-scenario, efficient, and accurate monitoring. For example, the published patent CN116844228A, "A Method for Recognizing the Actions of Captive Giant Pandas Based on Spatiotemporal Channel Attention Mechanism," is only designed for captive scenarios and does not cover common scenarios in actual monitoring, such as complex environments in the wild, low light at night, and occlusion, resulting in insufficient scenario adaptability. The published patent CN111666881B, "A Method for Tracking and Analyzing Giant Panda Walking, Bamboo Eating, and Estrus Behaviors," uses traditional background extraction algorithms and single feature analysis without integrating environmental scene features. Its recognition accuracy is limited in complex scenarios, and it does not involve efficient annotation schemes and model iteration mechanisms, resulting in a large workload for data annotation and difficulty in continuously optimizing model performance. In addition, most existing technologies lack visualization output and data feedback loops, resulting in insufficient practicality and operability, and failing to provide comprehensive and reliable technical support for giant panda conservation. Summary of the Invention

[0004] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides a method and system for recognizing giant panda behavior based on a large model. This addresses the deficiencies of existing giant panda behavior recognition technologies, such as incomplete scene coverage, low data annotation efficiency, insufficient recognition accuracy in complex scenarios, and a lack of effective model iteration mechanisms and practical output methods.

[0005] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: a method and system for recognizing giant panda behavior based on a large model, comprising the following steps: S1. Acquire video stream data of giant pandas in different scenarios using an image acquisition device, and perform frame extraction processing on the video stream data to obtain a frame image dataset; the different scenarios include captive scenarios, wild scenarios, daytime scenarios, nighttime scenarios, and scenarios with occlusion. S2. Clean the frame image dataset and remove invalid images; expand the dataset using image enhancement technology to obtain an expanded dataset; implement semi-automatic annotation of giant panda target and behavior categories based on the SAM visual big model, and perform annotation processing on the expanded dataset to generate a training dataset with annotation information; the invalid images include blurred images, overexposed images, and underexposed images; S3. Using the SAM visual large model as the backbone network, a behavior classification branch is connected to the feature output end of the backbone network, and a scene feature auxiliary judgment module is added to construct an initial recognition model; the training dataset is input into the initial recognition model, and the model parameters are fine-tuned through transfer learning to obtain an optimized giant panda behavior recognition model; the scene feature auxiliary judgment module is used to extract environmental features from the image and fuse them with the behavior features. S4. Input the giant panda image or video stream to be identified into the optimized giant panda behavior recognition model. The model completes the accurate segmentation and feature extraction of the giant panda target through the SAM visual large model backbone network, and then outputs the preliminary behavior recognition result through the behavior classification branch. Combined with the scene feature auxiliary judgment module, the final giant panda behavior recognition result is output. S5. The final panda behavior recognition result is digitally output and a behavior report is generated; if the confidence level of the final panda behavior recognition result is lower than the preset threshold, the corresponding data is fed back to the annotation stage of S2, and after supplementing the annotation, it is re-entered into the optimization model of S3.

[0006] Preferably, the image enhancement techniques in S2 include horizontal flipping, brightness adjustment, and noise addition; the brightness adjustment range is ±20%; and the noise is Gaussian noise.

[0007] Preferably, the specific process of semi-automatic annotation in S2 is as follows: first, the outline of the giant panda target and the corresponding behavior category in the frame image are initially annotated by the SAM visual large model, and then the initial annotation results are manually verified and corrected to complete the annotation.

[0008] Preferably, the environmental features extracted by the scene feature auxiliary judgment module in S3 include water source area features, occlusion features, and light intensity features; the fusion adopts an attention mechanism to achieve weighted fusion of behavioral features and environmental features.

[0009] A giant panda behavior recognition system based on a visual big data model includes: Data acquisition module: Composed of an image acquisition device and a frame extraction unit. The image acquisition device is used to acquire video stream data of giant pandas in different scenes, and the frame extraction unit is used to extract frames from the video stream data to obtain a frame image dataset. The data preprocessing module includes a cleaning unit, an enhancement unit, and a labeling unit. The cleaning unit is used to remove invalid images from the frame image dataset. The enhancement unit is used to expand the dataset using image enhancement techniques. The labeling unit integrates the SAM visual model to achieve semi-automatic labeling of panda targets and behavior categories, generating a training dataset with labeled information. Model training and optimization module: includes a model building unit and a transfer learning unit; the model building unit is used to build an initial recognition model by using the SAM visual large model as the backbone network, connecting the behavior classification branch and adding a scene feature auxiliary judgment module; the transfer learning unit is used to input the training dataset into the initial recognition model, and fine-tune the parameters through transfer learning to obtain an optimized giant panda behavior recognition model. Recognition and reasoning module: It is used to receive images or video streams of giant pandas to be recognized, call the optimized giant panda behavior recognition model, and output the final giant panda behavior recognition result; The result output and feedback module includes an output unit and a feedback unit. The output unit is used to digitize and output the final panda behavior recognition result and generate a behavior report. The feedback unit is used to feed back the corresponding data to the annotation unit of the data preprocessing module when the confidence level of the recognition result is lower than a preset threshold.

[0010] Preferably, the image acquisition device includes an infrared camera and a high-definition camera; the annotation unit of the data preprocessing module is also equipped with a manual verification interface, which is used to manually correct the preliminary annotation results of the SAM visual large model.

[0011] Preferably, the model training and optimization module is also equipped with a cloud training platform, which provides the computing resources required for transfer learning; the result output and feedback module is also equipped with a visualization unit to intuitively display the data information in the behavior report.

[0012] (III) Beneficial Effects This invention provides a method and system for recognizing giant panda behavior based on a large model. It has the following beneficial effects: This invention significantly improves the accuracy of giant panda behavior recognition in complex scenarios by collecting data across multiple scenarios, enhancing images, cleaning and optimizing data quality, and combining the precise segmentation and scene feature fusion mechanism of the SAM visual big data model. Simultaneously, the semi-automatic annotation mode reduces manual workload, transfer learning lowers model training costs, and the collaborative operation of various modules improves overall processing and recognition efficiency, achieving high efficiency and low cost. Furthermore, the confidence feedback mechanism promotes data feedback for supplementary annotation, helping the model to continuously iterate and optimize to adapt to more real-world scenarios. It also supports multi-format input and visual output, clearly presenting recognition results, providing reliable technical support for giant panda monitoring and protection, and demonstrating outstanding practicality. Detailed Implementation

[0013] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0014] Example 1: This invention provides a method and system for recognizing giant panda behavior based on a large model, comprising the following steps: S1. Acquire video stream data of giant pandas in different scenarios using image acquisition devices, and perform frame extraction processing on the video stream data to obtain a frame image dataset. The different scenarios include captive scenarios, wild scenarios, daytime scenarios, nighttime scenarios, and scenarios with occlusion. The acquisition devices are configured according to the characteristics of the scenarios. Fixed high-definition devices are used for captive scenarios, mobile infrared and high-definition devices are used for wild scenarios, infrared devices are given priority for nighttime scenarios, and multi-angle layouts are added for occluded scenarios. The video stream acquisition frame rate is 24-30 frames / second, and the frame extraction interval is 1-5 frames / second to ensure the integrity of the behavior and avoid redundancy, ultimately forming a frame image dataset covering multiple scenarios, multiple time periods, and multiple behaviors. S2. Clean the frame image dataset and remove invalid images; expand the dataset using image enhancement techniques to obtain an expanded dataset; implement semi-automatic annotation of panda target and behavior categories based on the SAM visual model, and perform annotation processing on the expanded dataset to generate a training dataset with annotation information; the invalid images include blurry images, overexposed images, and underexposed images. During cleaning, the standard is a pixel sharpness ≥85% and a brightness value of 100-250, and the quantitative judgment formula is as follows: Where C represents image pixel sharpness and L represents image brightness, invalid images that do not meet the standards are removed, and the percentage of valid images retained is no less than 70%; the image enhancement technology adopts a combination of horizontal flipping, brightness adjustment, and noise addition, where the brightness adjustment can be adaptively adjusted according to the actual scene lighting, with the adjustment range controlled within ±15%-25%, and the core calculation method is as follows: In the formula, L0 is the original image brightness, Lnew is the adjusted brightness, and k is the adjustment coefficient, where k∈[−0.2,0.2]. The Gaussian noise intensity is selected in the range of 0.01-0.05 to avoid excessive noise affecting image features. Through this combined strategy, the dataset size is expanded to 2-3 times the original effective dataset, effectively improving the model's adaptability to different lighting and environmental interference, and enhancing generalization performance. The training dataset generated after annotation must contain three core annotation information: target contour coordinates, behavior category labels, and scene type labels. Semi-automatic annotation is divided into two stages: batch preliminary annotation and manual verification and correction. In the first stage, the SAM vision large model batch processes and expands the dataset, automatically identifying the giant panda target contour (recognition accuracy ≥90%) and matching behavior categories such as eating, climbing, resting, and moving. In the second stage, preliminary results are retrieved through the manual verification interface, and the annotators focus on reviewing the annotation content that is incomplete due to occlusion or blurred behavior boundaries, and make corrections and improvements. After verification and correction, the annotation accuracy must reach more than 95% to ensure the annotation quality of the training dataset and provide reliable data support for model training. S3. Using the SAM visual large model as the backbone network, a behavior classification branch is connected to the feature output end of the backbone network. A scene feature auxiliary judgment module is also added to construct an initial recognition model. The training dataset is input into the initial recognition model, and the model parameters are fine-tuned through transfer learning to obtain an optimized giant panda behavior recognition model. The scene feature auxiliary judgment module is used to extract environmental features from the image and fuse them with the behavior features. The environmental features include water source area features, occlusion features, and light intensity features. The fusion uses an attention mechanism to achieve weighted fusion, highlighting the core role of the behavior features. The calculation method is as follows: Where Fb is the behavioral feature, Fe is the environmental feature, α is the weighting coefficient of the behavioral feature (α∈[0.6,0.8]), and β is the weighting coefficient of the environmental feature (β∈[0.2,0.4]), and α+β=1. This fusion method allows the model to modify behavioral judgments based on environmental context, such as prioritizing the matching of "drinking water" behavior near water sources, improving recognition accuracy in complex scenarios. The behavioral classification branch has 3-5 hidden layers for refined classification. During transfer learning, the dataset is divided into training, validation, and test sets in a 7:2:1 ratio. The backbone network only fine-tunes the parameters of the last 3-5 layers, and the parameter update formula is: In the formula, θt is the model parameter in the t-th iteration, θt+1 is the updated parameter, η is the learning rate (η∈[10−5,10−3]), and ∇Loss(θt) is the gradient of the loss function at parameter θt; after 50-100 iterations, training stops when the accuracy of the validation set is stable for 5 consecutive iterations. S4. Input the giant panda image or video stream to be identified into the optimized giant panda behavior recognition model. The model completes the accurate segmentation and feature extraction of the giant panda target through the SAM visual large model backbone network, and then outputs the preliminary behavior recognition result through the behavior classification branch. Combined with the scene feature auxiliary judgment module, the final giant panda behavior recognition result is output. The input data is first scaled to a uniform size of 640×640, the backbone network segmentation IoU value is ≥0.85, the behavior classification branch outputs the probability values ​​of various behaviors, and the scene feature auxiliary module outputs a unique behavior category label after correction. S5. The final panda behavior recognition result is digitized and a behavior report is generated. If the confidence level of the final panda behavior recognition result is lower than a preset threshold, the corresponding data is fed back to the annotation stage in S2, and after supplementary annotation, it is re-input into the optimization model in S3. The digitized output includes behavior category, recognition time, scene type, and confidence level. The preset threshold is 80%-90%, and the confidence level judgment formula is: Where Conf is the recognition confidence level and T is the preset threshold (T∈[0.8,0.9]); data below the threshold are automatically pushed to the annotation stage, and after supplementary annotation, they are re-added to the training set. The model parameters are updated every 10-20 iterations.

[0015] Image enhancement techniques in S2 include horizontal flipping, brightness adjustment, and noise addition. The brightness adjustment range is ±20%. The noise is Gaussian noise. The dataset is expanded using a combination of horizontal flipping, brightness adjustment, and Gaussian noise addition. The brightness adjustment can be adaptively adjusted according to the actual scene lighting, with the range controlled within ±15%-25%. The Gaussian noise intensity is selected in the range of 0.01-0.05 to avoid excessive noise affecting image features. Through this combined strategy, the dataset size is expanded to 2-3 times the original effective dataset, effectively improving the model's adaptability to different lighting and environmental interference, and enhancing generalization performance.

[0016] The semi-automatic annotation process in S2 is as follows: First, the SAM visual model is used to initially annotate the panda target outline and corresponding behavior category in the frame image. Then, the initial annotation results are manually verified and corrected to complete the annotation. The semi-automatic annotation is divided into two stages: batch initial annotation and manual verification and correction. In the first stage, the SAM visual model batch processes and expands the dataset, automatically identifies the panda target outline (recognition accuracy ≥ 90%) and matches it with behavior categories such as eating, climbing, resting, and moving. In the second stage, the initial results are retrieved through the manual verification interface. Annotators focus on reviewing the annotation content that is incomplete due to occlusion or has blurred behavior boundaries, and make corrections and improvements. After verification and correction, the annotation accuracy needs to reach more than 95% to ensure the annotation quality of the training dataset and provide reliable data support for model training. The environmental features extracted by the scene feature-assisted judgment module in S3 include water source area features, occlusion features, and light intensity features; the fusion adopts an attention mechanism to achieve weighted fusion of behavioral features and environmental features; (1) Environmental feature extraction details: Water source area features focus on the texture and gray distribution features of the water body; occlusion features extract the shape, area ratio and relative position relationship with the giant panda target; light intensity features are extracted with the average pixel brightness and contrast as the core dimensions to ensure that the features can accurately reflect environmental differences.

[0017] (2) Feature fusion details: When using the attention mechanism for weighted fusion, the core role of behavioral features is highlighted. The weighting coefficient of behavioral features is set in the range of 0.6-0.8, and the weighting coefficient of environmental features is set in the range of 0.2-0.4. Through this fusion method, the model can combine the environmental context to correct the behavior judgment. For example, in the vicinity of water source area, the "drinking water" behavior is matched first, which improves the recognition accuracy in complex scenarios.

[0018] Example 2: A giant panda behavior recognition system based on a visual big data model includes: The data acquisition module consists of an image acquisition device and a frame extraction unit. The image acquisition device is used to acquire video stream data of giant pandas in different scenes. The frame extraction unit is used to extract frames from the video stream data to obtain a frame image dataset. It supports dynamic device scheduling, automatically switches to infrared camera mode when the nighttime light intensity is <100 lux, and triggers multi-angle device synchronous acquisition when the proportion of obstructions is >30%. The frame extraction unit is integrated into the edge computing device, with a processing latency of 100-500 milliseconds. The data is transmitted to the preprocessing module via 4G / 5G and LoRa protocols. The data preprocessing module includes a cleaning unit, an enhancement unit, and a labeling unit. The cleaning unit removes invalid images from the frame image dataset. The enhancement unit expands the dataset using image enhancement techniques. The labeling unit integrates the SAM visual model to achieve semi-automatic labeling of panda targets and behavioral categories, generating a training dataset with labeled information. The cleaning unit has a processing rate of ≥100 images / second and generates a cleaning log containing the number of invalid images and their causes. The enhancement unit supports dynamic configuration of brightness adjustment and noise intensity. The labeling unit supports offline batch labeling and online real-time labeling, with a labeling efficiency of ≥50 images / minute, and regularly generates labeling quality reports. The model training and optimization module includes a model building unit and a transfer learning unit. The model building unit uses the SAM visual large model as the backbone network, connects the behavior classification branch, and adds a scene feature auxiliary judgment module to build an initial recognition model. The transfer learning unit inputs the training dataset into the initial recognition model and fine-tunes the parameters through transfer learning to obtain an optimized giant panda behavior recognition model. The model building unit provides a visual building interface and supports customizing the backbone network version and the number of hidden layers in the classification branch. The transfer learning unit supports parallel computing on 10-20 GPUs (single GPU memory ≥ 16GB), supports breakpoint resume training, and outputs real-time monitoring indicators such as training loss value and validation accuracy. Recognition and Inference Module: This module receives images or video streams of giant pandas to be recognized, calls the optimized giant panda behavior recognition model, and outputs the final giant panda behavior recognition result. It supports JPG and PNG format images and MP4 and AVI format video streams as input. Input data is automatically standardized. The inference response time is 500-1000 milliseconds. When inferring video streams in real time, duplicate behavior results are automatically filtered. The results output and feedback module includes an output unit and a feedback unit. The output unit digitizes the final panda behavior recognition results and generates a behavior report. The feedback unit feeds back the corresponding data to the annotation unit of the data preprocessing module when the confidence level of the recognition result is lower than a preset threshold. The output unit supports the output of structured text reports (CSV format) and visualization charts (line charts, bar charts, heatmaps). The feedback unit monitors the confidence level in real time. Data below the threshold is marked and a manual verification reminder is triggered. Supplemented annotation data is automatically fed back into the training dataset.

[0019] The image acquisition equipment includes an infrared camera and a high-definition camera; the annotation unit of the data preprocessing module is also equipped with a manual verification interface, which is used to manually correct the preliminary annotation results of the SAM visual large model; (1) Image acquisition equipment parameters: HD camera with pixel specifications of 5 million to 10 million pixels, infrared camera with detection distance of 10 to 30 meters, both of which are waterproof, dustproof, and have a wide temperature range of -20℃ to 40℃, suitable for complex outdoor environments.

[0020] (2) Manual verification interface function: It supports multiple annotators to log in and operate at the same time, and has the functions of annotation result backtracking and correction record saving; the interface can filter images with model annotation confidence of <90% and automatically push them to the manual review queue. Annotators can directly correct them through tools such as drawing and label editing, and the correction results are synchronized to the dataset of the annotation unit in real time. The model training and optimization module is also equipped with a cloud training platform, which provides the computing resources required for transfer learning; the result output and feedback module is also equipped with a visualization unit to intuitively display the data information in the behavior report. (1) Cloud training platform: It provides elastic computing resource scheduling function, which can dynamically allocate GPU / CPU resources according to the scale of training tasks and support the setting of training task priorities; it has data security storage function, and the training dataset and model parameter files are stored in encrypted form, and support access control to prevent data leakage.

[0021] (2) Visualization unit: Supports three core data display formats - behavior time series distribution chart (showing the trend of behavior changes in different time periods), behavior proportion bar chart (statistics the proportion of the frequency of various behaviors), and activity area heat map (presenting the key activity areas of giant pandas); supports data filtering (filtering by scene and time interval) and export functions, which makes it easy for users to quickly analyze core information.

[0022] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for recognizing giant panda behavior based on a large model, characterized in that, Includes the following steps: S1. Acquire video stream data of giant pandas in different scenarios using an image acquisition device, and perform frame extraction processing on the video stream data to obtain a frame image dataset; the different scenarios include captive scenarios, wild scenarios, daytime scenarios, nighttime scenarios, and scenarios with occlusion. S2. Clean the frame image dataset and remove invalid images; expand the dataset using image enhancement technology to obtain an expanded dataset; implement semi-automatic annotation of giant panda target and behavior categories based on the SAM visual big model, and perform annotation processing on the expanded dataset to generate a training dataset with annotation information; the invalid images include blurred images, overexposed images, and underexposed images; S3. Using the SAM visual large model as the backbone network, a behavior classification branch is connected to the feature output end of the backbone network, and a scene feature auxiliary judgment module is added to construct an initial recognition model; the training dataset is input into the initial recognition model, and the model parameters are fine-tuned through transfer learning to obtain an optimized giant panda behavior recognition model; the scene feature auxiliary judgment module is used to extract environmental features from the image and fuse them with the behavior features. S4. Input the giant panda image or video stream to be identified into the optimized giant panda behavior recognition model. The model completes the accurate segmentation and feature extraction of the giant panda target through the SAM visual large model backbone network, and then outputs the preliminary behavior recognition result through the behavior classification branch. Combined with the scene feature auxiliary judgment module, the final giant panda behavior recognition result is output. S5. The final panda behavior recognition result is digitally output and a behavior report is generated; if the confidence level of the final panda behavior recognition result is lower than the preset threshold, the corresponding data is fed back to the annotation stage of S2, and after supplementing the annotation, it is re-entered into the optimization model of S3.

2. The giant panda behavior recognition method based on a large model according to claim 1, characterized in that: The image enhancement techniques in S2 include horizontal flipping, brightness adjustment, and noise addition; the brightness adjustment range is ±20%; the noise is Gaussian noise.

3. The giant panda behavior recognition method based on a large model according to claim 1, characterized in that: The specific process of semi-automatic annotation in S2 is as follows: First, the outline of the giant panda target and the corresponding behavior category in the frame image are initially annotated using the SAM visual large model. Then, the initial annotation results are manually verified and corrected to complete the annotation.

4. The giant panda behavior recognition method based on a large model according to claim 1, characterized in that: The environmental features extracted by the scene feature auxiliary judgment module in S3 include water source area features, occlusion features, and light intensity features; the fusion adopts an attention mechanism to achieve weighted fusion of behavioral features and environmental features.

5. A giant panda behavior recognition system based on a large visual model, characterized in that, The method for recognizing giant panda behavior based on a large visual model as described in any one of claims 1-4 includes: Data acquisition module: Composed of an image acquisition device and a frame extraction unit. The image acquisition device is used to acquire video stream data of giant pandas in different scenes, and the frame extraction unit is used to extract frames from the video stream data to obtain a frame image dataset. The data preprocessing module includes a cleaning unit, an enhancement unit, and a labeling unit. The cleaning unit is used to remove invalid images from the frame image dataset. The enhancement unit is used to expand the dataset using image enhancement techniques. The labeling unit integrates the SAM visual model to achieve semi-automatic labeling of panda targets and behavior categories, generating a training dataset with labeled information. Model training and optimization module: includes a model building unit and a transfer learning unit; the model building unit is used to build an initial recognition model by using the SAM visual large model as the backbone network, connecting the behavior classification branch and adding a scene feature auxiliary judgment module; the transfer learning unit is used to input the training dataset into the initial recognition model, and fine-tune the parameters through transfer learning to obtain an optimized giant panda behavior recognition model. Recognition and reasoning module: It is used to receive images or video streams of giant pandas to be recognized, call the optimized giant panda behavior recognition model, and output the final giant panda behavior recognition result; The result output and feedback module includes an output unit and a feedback unit. The output unit is used to digitize and output the final panda behavior recognition result and generate a behavior report. The feedback unit is used to feed back the corresponding data to the annotation unit of the data preprocessing module when the confidence level of the recognition result is lower than a preset threshold.

6. The giant panda behavior recognition system based on a large model according to claim 5, characterized in that: The image acquisition device includes an infrared camera and a high-definition camera; the annotation unit of the data preprocessing module is also equipped with a manual verification interface, which is used to manually correct the preliminary annotation results of the SAM visual large model.

7. The giant panda behavior recognition system based on a large model according to claim 5, characterized in that: The model training and optimization module is also equipped with a cloud training platform, which provides the computing resources required for transfer learning; the result output and feedback module is also equipped with a visualization unit to intuitively display the data information in the behavior report.