Intelligent snapshot and video recording device and method for bird activities around overhead transmission line

Abstract

The invention provides an intelligent snapshot video recording device and method for bird activities around an overhead transmission line, and relates to the technical field of electric power protection. The device comprises a camera module, an edge calculation module, a communication module and a power supply module, the camera module adopts a first frame rate to carry out video recording on an overhead transmission line channel, the first frame rate is smaller than a second frame rate, and the first frame rate is small, that is, the overhead transmission line channel is subjected to video recording by adopting a low-frequency surface; the activity condition of the birds around the line is shot in a silent triggering mode; when the edge calculation module recognizes that birds exist in the image, a snapshot picture is sent to the camera module, and a snapshot video recording instruction of second frame rate video recording is adopted, that is, snapshot video recording is started only when the birds fly over the overhead line or approach the overhead line, so that the flow consumption can be greatly reduced, and the energy consumption of the device is saved; operation and maintenance personnel are assisted to quickly find the bird activity process, and fault analysis and evidence obtaining are facilitated.

Description

Technical Field

[0001] This application relates to the field of power protection technology, and in particular to an intelligent device and method for capturing and recording bird activity around overhead transmission lines. Background Technology

[0002] In recent years, several incidents have occurred in Tibet and other regions where rare wild birds, such as black-necked cranes, have accidentally collided with power transmission lines, resulting in injuries or deaths. To protect these birds from injury or death caused by collisions with power lines, relevant organizations have conducted research and applied bird warning systems. For example, warning balls used in aviation are installed on overhead power lines to alert birds to the presence of overhead lines.

[0003] Birds exhibit different characteristics when flying over overhead power lines. To analyze the specific flying characteristics of birds when facing overhead power lines, it is necessary to identify and capture the process of birds flying over overhead power lines, so as to provide technical support for the analysis of birds accidentally colliding with power lines and birds avoiding overhead power lines.

[0004] In addition, after installing warning devices (such as warning balls used in aviation), it is necessary to analyze the effectiveness of the warning devices and monitor the behavior of birds flying over overhead lines with and without warning devices. Therefore, it is also necessary to identify and capture the process of birds flying over overhead lines to provide technical support for the improvement of warning devices.

[0005] Therefore, how to accurately and efficiently identify and capture the process of birds flying over overhead power lines has become an urgent technical problem to be solved. Summary of the Invention

[0006] In view of the above problems, this application is made to provide an intelligent device and method for capturing and recording bird activity around overhead transmission lines to overcome or at least partially solve the above problems. The technical solution is as follows: In a first aspect, a smart video capture device for capturing bird activity around overhead transmission lines is provided, the device comprising: The camera module is installed on the power transmission tower and is used to record video of the overhead power transmission line channel at the first frame rate. The edge computing module is used to analyze each frame of the image recorded by the camera module in real time using a pre-trained intelligent recognition model. When birds are identified in the image, the module sends a capture and recording instruction to the camera module, which captures the image and records the video at a second frame rate. The first frame rate is less than the second frame rate. The camera module is also used to capture images of bird activity around overhead power lines according to the capture and recording instructions, adjust the first frame rate to the second frame rate, record video using the second frame rate, and submit the captured images of bird activity around overhead power lines and the recorded video to the communication module. The communication module is used to upload captured images and recorded videos of birds activity around overhead power lines to the backend server. The power supply module is used to power the camera module, edge computing module, and communication module.

[0007] In one possible implementation, the intelligent recognition model in the edge computing module is a lightweight version of YOLOv12, where YOLOv12 is an attention-centric real-time object detection model whose network structure includes a backbone network, a feature enhancement module, and a detection head.

[0008] In one possible implementation, the lightweight YOLOv12 model is achieved by reducing redundant layers. Due to the differences in the characteristics of different target types, some layers exhibit overcompensation in feature downsampling. For specific object recognition, an experimental method is used to try removing some layers, analyzing the accuracy of the recognition results. After training the model, several images are selected for recognition. If the change in the Intersection over Union (IOU) value is less than a preset threshold, the redundant layer reduction operation is considered effective. Through continuous iteration, redundant layers in the network structure are searched and reduced. After reducing redundant layers, the YOLOv12 network structure has fewer layers, achieving network lightweighting.

[0009] In one possible implementation, the preset threshold is 0.02.

[0010] In one possible implementation, the snapshot recording instruction includes the current recording duration; The camera module is also used to capture bird activity around the overhead transmission line according to the capture recording instruction containing the current recording duration, adjust the first frame rate to the second frame rate, use the second frame rate, and record according to the current recording duration, and submit the captured bird activity pictures around the overhead transmission line and the recorded video of the current recording duration to the communication module. The communication module is used to upload captured images of birds' activities around overhead power lines and recorded videos of the current recording duration to the backend server.

[0011] In one possible implementation, the device is used in scenarios where no warning device is installed on overhead power transmission lines.

[0012] In one possible implementation, the device is used in scenarios where warning devices are installed on overhead power transmission lines.

[0013] Secondly, a method for intelligently capturing and recording bird activity around overhead transmission lines is provided, the method comprising: The camera module installed on the transmission tower is controlled to record video at the first frame rate facing the overhead transmission line channel; The edge computing module uses a pre-trained intelligent recognition model to analyze each frame of the image recorded by the camera module in real time. When birds are identified in the image, a capture and recording instruction is sent to the camera module to capture the image and record the video at a second frame rate; where the first frame rate is less than the second frame rate. The camera control module captures images of birds around the overhead transmission line according to the capture and recording command, adjusts the first frame rate to the second frame rate, records the video using the second frame rate, and submits the captured images of birds around the overhead transmission line and the recorded video to the communication module. The communication module uploads captured images and recorded videos of birds activity around overhead power lines to the backend server.

[0014] In one possible implementation, the method further includes: Based on YOLOv12, a lightweight version of YOLOv12 is created to obtain an intelligent recognition model. YOLOv12 is an attention-centric real-time object detection model, and its network structure includes a backbone network, a feature enhancement module, and a detection head.

[0015] In one possible implementation, YOLOv12 is used as a foundation, and YOLOv12 is lightweighted to obtain an intelligent recognition model, including: Collect images of bird activity around overhead power lines as a training set; YOLOv12 is trained using the training set to obtain the first weight file; Sequentially select one layer from the multiple layers set in YOLOv12 and remove it, and use the training set to train YOLOv12 after performing the single-layer removal operation to obtain the second weight file; Select several images, apply the first weight file to identify birds, and record the Intersection over Union (IOU) values; apply the second weight file to identify birds, and record the IOU values. For the IOU values ​​under the first weight file and the second weight file; Determine if the change in IOU value of the same image is less than a preset threshold. If yes, the redundancy reduction layer is considered successful and the next round of operation is executed. If no, the redundancy reduction layer is considered to have failed and the next round of operation is executed. Through continuous iteration, redundant layers in the network structure are searched and reduced to obtain an intelligent recognition model. After reducing redundant layers, the YOLOv12 network structure has fewer layers, achieving network lightweighting.

[0016] Based on the above technical solution, this application provides an intelligent image capture and recording device and method for bird activity around overhead transmission lines. The device includes a camera module installed on a transmission tower for recording video of the overhead transmission line channel at a first frame rate; an edge computing module for analyzing each frame of the image recorded by the camera module in real time using a pre-trained intelligent recognition model, and sending an image capture and recording command to the camera module when birds are detected in the image, wherein the first frame rate is less than the second frame rate; the camera module is also used to capture bird activity around the overhead transmission line according to the image capture and recording command, adjust the first frame rate to the second frame rate, record video at the second frame rate, and submit the captured images of bird activity around the overhead transmission line and the recorded video to a communication module; the communication module is used to upload the captured images of bird activity around the overhead transmission line and the recorded video to a background server; and a power supply module is used to supply power to the camera module, the edge computing module, and the communication module. As can be seen, in the device provided in this application embodiment, the camera module records video facing the overhead transmission line channel at a first frame rate, which is lower than the second frame rate. The lower frame rate indicates a low-frequency recording method, allowing for silent triggering of bird activity around the line. When the edge computing module detects birds in the image, it sends a capture and recording command to the camera module, instructing it to capture an image and record video at the second frame rate. This means that capture and recording are only initiated when birds fly over or approach the overhead line, significantly reducing bandwidth consumption, saving device energy, and assisting maintenance personnel in quickly identifying bird activity, thus aiding in fault analysis and evidence collection. Furthermore, the device provides uploads the captured images and recorded videos of bird activity around the overhead transmission line to a backend server. The backend server can then analyze changes in bird behavior before and after the installation of the warning device, supporting bird protection and bird-related fault analysis and handling. This facilitates rapid fault analysis and contributes to ecological environmental protection. It can also analyze the effectiveness of the warning device, providing technical support for its improvement. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the description of the embodiments of this application will be briefly introduced below.

[0018] Figure 1 This paper shows a structural diagram of an intelligent video capture device for capturing bird activity around overhead transmission lines provided in an embodiment of this application. Figure 2 A flowchart is shown below illustrating the intelligent method for capturing and recording bird activity around overhead transmission lines provided in an embodiment of this application. Figure 3A flowchart of a method for reducing redundant layers provided in an embodiment of this application is shown. Detailed Implementation

[0019] Exemplary embodiments of the present application will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present application are shown in the drawings, it should be understood that the present application may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this application will be thorough and complete, and will fully convey the scope of the present application to those skilled in the art.

[0020] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such use can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the term "comprising" and its variations should be interpreted as open-ended terms meaning "including but not limited to."

[0021] Currently, monitoring birds flying over overhead power lines involves capturing video footage with cameras and manually tracing the long video streams to locate the birds' flight paths. The problem is that birds typically fly very fast, often completing the crossing in just a few seconds. This brief flight time makes it impossible to capture with traditional image monitoring. Continuous recording followed by sifting through long videos to find the bird's flight path is labor-intensive, resource-intensive, and lacks automation. Even using artificial intelligence algorithms to analyze long videos and extract flight segments would consume significant computing resources and lack real-time performance, failing to meet analytical requirements.

[0022] To address the aforementioned technical problems, this application provides an intelligent video capture device for capturing bird activity around overhead transmission lines, such as... Figure 1 As shown, the intelligent video recording device for capturing bird activity around the overhead transmission line may include: Camera module 10 is installed on the power transmission tower and is used to record video of the overhead power transmission line channel at the first frame rate; The edge computing module 20 is used to analyze each frame of the image recorded by the camera module 10 in real time using a pre-trained intelligent recognition model. When a bird is identified in the image, the module sends a capture and recording instruction to the camera module 10 to capture the image and record the video at a second frame rate. The first frame rate is less than the second frame rate. The camera module 10 is also used to capture bird activity around the overhead transmission line according to the capture and recording instruction, adjust the first frame rate to the second frame rate, record the video using the second frame rate, and submit the captured bird activity pictures and recorded video around the overhead transmission line to the communication module 30. Communication module 30 is used to upload captured images and recorded videos of bird activity around overhead power lines to the backend server; Power module 40 is used to supply power to the camera module, edge computing module, and communication module.

[0023] Furthermore, the intelligent capture and recording device for bird activity around overhead transmission lines provided in this application embodiment may also include installation accessories, etc. Figure 1 (not shown in the figure), this embodiment does not impose any limitations on this.

[0024] The camera module 10 can determine the frame rate by controlling the exposure time. The first frame rate and the second frame rate can be set according to actual needs, such as the first frame rate being 1 frame / second and the second frame rate being 24 or 30 frames / second. This embodiment does not impose any restrictions on this.

[0025] In the device provided in this application embodiment, the camera module 10 records video facing the overhead transmission line channel at a first frame rate lower than the second frame rate. The lower frame rate indicates a low-frequency recording method, allowing for silent triggering of bird activity around the line. When the edge computing module 20 detects birds in the image, it sends a capture image and recording command to the camera module 10 at the second frame rate. This means that capture and recording are only initiated when birds fly over or approach the overhead line, significantly reducing bandwidth consumption, saving device energy, and assisting maintenance personnel in quickly identifying bird activity, thus aiding in fault analysis and evidence collection. Furthermore, the device uploads the captured images and recorded videos of bird activity around the overhead transmission line to a backend server. The backend server can then analyze changes in bird behavior before and after the installation of the warning device, supporting bird protection and bird-related fault analysis and handling. This facilitates rapid fault analysis and contributes to ecological environmental protection. It can also analyze the effectiveness of the warning device, providing technical support for its improvement.

[0026] This application provides a possible implementation method in which the intelligent recognition model in the edge computing module 20 is a model after lightweight processing of YOLOv12. YOLOv12 is an attention-centered real-time target detection model, and its network structure includes a backbone network, a feature enhancement module, and a detection head.

[0027] Specifically, YOLOv12 is the latest generation model in the YOLO series for object detection, and its core structure has achieved a breakthrough innovation in balancing real-time performance and accuracy. The overall architecture continues the single-stage detection paradigm, but it has reconstructed the three major components—the backbone network, the feature enhancement module, and the detection head—through multi-level cross-dimensional fusion and dynamic computation optimization.

[0028] In the YOLOv12 network architecture, the backbone network may include: Layer 0 (Conv) is the input preprocessing layer, typically a 6×6 convolution, which quickly downsamples and extracts low-level features (such as edges, textures, etc.).

[0029] The first layer, Conv, further downsamples (e.g., 3×3 convolution) to compress the spatial dimension and increase the number of channels.

[0030] Layer 2, C3k2, is a key module. C3k2 is an improved cross-stage partial connection module in YOLOv12 that enhances feature reuse and reduces computation through grouped convolutions and residual connections.

[0031] Layer 3 (Conv) is a transition layer that adjusts the number of channels and prepares for the next stage.

[0032] The fourth layer, C3k2, is applied again to extract mid-layer features (such as local structure of objects).

[0033] The 5th layer Conv is downsampled to a higher semantic level.

[0034] Layer 6 A2C2f, A2C2f = A2 (region attention) + C2f (cross-stage feature fusion); The A2 module divides the feature map into four sub-regions, calculates attention weights for each, and focuses on key regions (such as bird activity areas); the C2f structure preserves gradient flow and improves the ability to detect small targets.

[0035] Layer 7 (Conv) transitions to the deep network.

[0036] Layer 8, A2C2f, is a deep feature enhancement layer that captures global contextual information (such as the relationship between power transmission lines and the background).

[0037] In the YOLOv12 network architecture, the feature enhancement module may include: The 9th layer, Upsample, such as bilinear interpolation, recovers high-resolution feature maps for subsequent fusion with smaller-scale features.

[0038] The 10th layer, Concat, combines features from shallow layers (such as the 6th layer) and upsampled features, fusing high- and low-level semantic information.

[0039] The 11th layer, A2C2f, applies attention weighting to the fused features to suppress irrelevant background (such as sky noise).

[0040] The 12th layer, Upsample, performs another upsampling to prepare for feature fusion at a larger scale.

[0041] The 13th layer, Concat, stitches together shallower features (such as the 4th layer) to enhance the localization accuracy of small targets (such as birds).

[0042] Layer 14, A2C2f, performs final optimization on the mesoscale features, balancing details and semantics.

[0043] On layer 15 (Conv), adjust the number of channels to adapt to the input of the detection head.

[0044] The 16th layer, Concat, merges features from different paths (such as the 14th and 11th layers) to construct a multi-scale pyramid.

[0045] Layer 17, A2C2f, applies attention to large-scale features (such as the entire transmission tower) to highlight key areas.

[0046] Layer 18 Conv, channel alignment before output.

[0047] The 19th layer, Concat, is the final feature aggregation, covering targets at all scales.

[0048] Layer 20, C3k2, performs feature compression and enhancement before detection to reduce redundant computation.

[0049] In the YOLOv12 network architecture, the detection head can include: The 21st layer, Detect (detection output layer), has the following dynamic detection head: Output feature maps at three scales (e.g., 20×20, 40×40, 80×80) to detect large, medium, and small targets respectively; use loss optimization for bounding box regression to adapt to the irregular shape of birds; and use classification branch to solve the class imbalance problem between birds and the background in power transmission scenarios.

[0050] The backbone network employs heterogeneous depthwise separable convolutional blocks and introduces a multi-branch parallel processing mechanism. The basic unit consists of 3×3 depthwise convolutions, 1×1 pointwise convolutions, and dilated convolutional branches, with the outputs of each branch dynamically fused through channel attention weights. This design significantly improves the receptive field while maintaining lightweight architecture, resulting in a 37% improvement in low-level feature retention for small object detection. The network body has a 5-level pyramid structure, with each level employing an improved straddle spatial pyramid pooling (S3PP) module for downsampling. This module captures multi-scale contextual information by connecting dilated convolutional layers with different dilation rates in parallel, reducing computation by 21% compared to the traditional SPPF (fast spatial pyramid pooling) module.

[0051] The feature enhancement module innovatively introduces a bidirectional cross-scale attention mechanism. Based on the FPN (Feature Pyramid Network) path, a bottom-up semantic compensation path is added, with each path integrating channel-space dual-dimensional attention units. High-level features are first upsampled through sub-pixel convolution, multiplied pixel-level with low-level features, and then weighted for local saliency; low-level features are calibrated through deformable convolution before being concatenated with high-level features. This bidirectional interactive structure improves the mAP (mean Average Precision) by 4.2%, particularly enhancing the feature integrity of occluded targets. To address the differences between the various levels of the feature pyramid, a dynamic convolution kernel strategy is adopted: shallow layers use a large 5×5 kernel to enhance geometric feature extraction, while deeper layers switch to 3×3 kernel groups to reduce computational overhead.

[0052] The detection head achieves a unified approach of decoupling and dynamism. The classification and regression branches share basic features but employ independent enhancement paths: the classification branch aggregates scene semantics through a global context module, while the regression branch integrates a coordinate attention mechanism for precise localization. The innovation lies in proposing a DAG (Dynamic Anchor Generator), which predicts the anchor size distribution in real-time based on the input image content, replacing the preset anchor box scheme. The prediction layer introduces a task-aware feature selection mechanism, automatically assigning 3-5 optimal feature channel combinations to each output unit, improving computational efficiency by 30% while maintaining 99% of the original accuracy.

[0053] During training, a spectral normalization constraint is introduced to improve model robustness by restricting the Lipschitz constant of the weight matrix. The loss function employs a combination of an improved Focal Loss and CIoU (Complete Intersection over Union) loss, with a gradient reweighting mechanism added for difficult samples. Post-processing optimization utilizes a hybrid strategy of parallelized NMS (Non-Maximum Suppression) and decision tree filtering, reducing the false positive rate by 18% while maintaining recall.

[0054] This application provides a possible implementation method for lightweighting the YOLOv12 model by reducing redundant layers. Due to the differences in the characteristics of different target types, some layers exhibit overcompensation in feature downsampling. For specific object recognition, an experimental method is used to try removing some layers, analyzing the accuracy of the recognition results. After training the model, several images are selected for recognition. If the change in the IOU (Intersection Over Union) value is less than a preset threshold, the redundant layer reduction operation is considered effective. Through continuous iteration, redundant layers in the network structure are searched and reduced. After reducing redundant layers, the YOLOv12 network structure has fewer layers, achieving network lightweighting. Here, the IOU value is an indicator used in object detection tasks to evaluate the degree of overlap between the predicted bounding box and the ground truth bounding box.

[0055] This embodiment gradually simplifies the network by iteratively deleting layers and testing IOU stability, thereby improving inference efficiency. The lightweight model is more suitable for deployment on edge devices.

[0056] This application provides a possible implementation method in which the preset threshold can be 0.02.

[0057] This application provides a possible implementation method in which the snapshot recording instruction includes the current recording duration; The camera module 10 is also used to capture bird activity around the overhead transmission line according to the capture and recording instruction containing the current recording duration, adjust the first frame rate to the second frame rate, use the second frame rate, and record according to the current recording duration, and submit the captured bird activity pictures around the overhead transmission line and the recorded video of the current recording duration to the communication module 30. The communication module 30 is used to upload captured images of bird activity around overhead power lines and recorded videos of the current recording duration to the backend server.

[0058] In this embodiment, the current recording duration can be set according to actual needs to accommodate the recording of the complete process of birds flying over overhead lines, while also avoiding wasted recording time. Generally, birds fly over lines quickly, so the current recording duration can be set to 10 seconds. The duration can be appropriately increased or decreased based on factors such as the local bird species and environment.

[0059] The intelligent bird activity capture and recording device around overhead transmission lines provided in this application embodiment can be applied to scenarios where no warning devices are installed on overhead transmission lines, as well as scenarios where warning devices are installed on overhead transmission lines. The intelligent bird activity capture and recording device around overhead transmission lines uploads captured images and recorded videos of bird activity around overhead transmission lines to a backend server. In this way, the backend server can analyze changes in bird behavior before and after the installation of warning devices based on the captured images and recorded videos of bird activity around overhead transmission lines. This can support bird protection and bird-related fault analysis and handling, help to quickly analyze problems and faults, and contribute to ecological environmental protection. It can also analyze the effectiveness of warning devices and provide technical support for the improvement of warning devices.

[0060] Based on the intelligent capture and recording device for bird activities around overhead transmission lines provided in the above embodiments, and based on the same inventive concept, this application also provides an intelligent capture and recording method for bird activities around overhead transmission lines.

[0061] Figure 2 A flowchart of the intelligent capture and video recording method for bird activity around overhead transmission lines provided in an embodiment of this application is shown, such as... Figure 2 As shown, the intelligent video capture method for bird activity around the overhead transmission line may include the following steps S201 to S204: Step S201: Control the camera module installed on the transmission tower to record video at the first frame rate facing the overhead transmission line channel; In step S202, the edge computing module uses a pre-trained intelligent recognition model to analyze each frame of the image recorded by the camera module in real time. When birds are identified in the image, a capture and recording instruction is sent to the camera module to capture the image and record the video at the second frame rate; wherein the first frame rate is less than the second frame rate. Step S203: The camera module is controlled to capture images of bird activity around the overhead transmission line according to the capture and recording instruction, and the first frame rate is adjusted to the second frame rate. The second frame rate is used for recording, and the captured images of bird activity around the overhead transmission line and the recorded video are submitted to the communication module. In step S204, the communication module uploads the captured images and recorded videos of bird activity around the overhead transmission line to the backend server.

[0062] In this embodiment, the first frame rate and the second frame rate can be set according to actual needs, such as the first frame rate being 1 frame / second and the second frame rate being 24 or 30 frames / second. This embodiment does not impose any restrictions on this.

[0063] In this embodiment, the camera module records video at a first frame rate facing the overhead transmission line channel. This first frame rate is lower than the second frame rate, meaning a low-frequency recording is used to capture bird activity around the line in a silent triggering manner. When the edge computing module detects birds in the image, it sends a capture and recording command to the camera module, instructing it to capture an image and record video at the second frame rate. This means that capture and recording are only initiated when birds fly over or approach the overhead line, significantly reducing bandwidth consumption, saving device energy, and assisting maintenance personnel in quickly identifying bird activity, thus aiding in fault analysis and evidence collection. Furthermore, the captured images and recorded videos of bird activity around the overhead transmission line are uploaded to a backend server. The backend server can then analyze changes in bird behavior before and after the installation of the warning device, supporting bird protection and bird-related fault analysis and handling. This facilitates rapid fault analysis and contributes to ecological environmental protection. It can also analyze the effectiveness of the warning device, providing technical support for its improvement.

[0064] This application provides a possible implementation method, which may further include the following step A1: Step A1 involves lightweighting YOLOv12 to obtain an intelligent recognition model. YOLOv12 is an attention-centric real-time object detection model whose network structure includes a backbone network, a feature enhancement module, and a detection head. The lightweight intelligent recognition model is more suitable for deployment on edge devices.

[0065] This application provides a possible implementation method, and step A1 above may specifically include the following steps A1-1 to A1-7: Step A1-1: Collect images of bird activity around overhead power lines as a training set; Step A1-2: Use the training set to train YOLOv12 to obtain the first weight file; Steps A1-3: Select one layer from the multiple layers set in YOLOv12 and remove it, and use the training set to train YOLOv12 after the single-layer removal operation to obtain the second weight file. Steps A1-4: Select several images, apply the first weight file to identify birds, and record the Intersection over Union (IOU) values; apply the second weight file to identify birds, and record the IOU values. Steps A1-5: For the IOU values ​​under the first weight file and the second weight file; Steps A1-6: Determine whether the change in IOU value of the same image is less than a preset threshold. If yes, the redundancy reduction layer is considered successful and the next round of operation is executed. If no, the redundancy reduction layer is considered to have failed and the next round of operation is executed. Steps A1-7 involve iterating continuously to search for redundant layers in the network structure and implementing reduction operations to obtain an intelligent recognition model. After reducing redundant layers, the YOLOv12 network structure has fewer layers, achieving network lightweighting.

[0066] This embodiment improves inference efficiency by iteratively deleting layers and testing IOU stability.

[0067] In one specific embodiment, since edge computing terminal resources are limited, running YOLOv12 directly on the edge computing terminal requires a lot of computing resources, which further occupies the already scarce computing resources. Here, the algorithm is lightweighted to reduce the occupation of valuable computing resources.

[0068] 1) Optimize the network structure The specific approach involves using a more efficient backbone design, specifically employing EfficientNetV2 (an efficient convolutional neural network model) to construct the backbone. This avoids the performance bottleneck issues of traditional methods by uniformly scaling the network's three dimensions (depth, width, and resolution) instead of a single dimension.

[0069] The specific method involves replacing the YOLOv12 backbone network with EfficientNet V2. Since EfficientNetV2 has efficient feature extraction capabilities and better parameter utilization, this allows the model to extract image features better with the same computational resources, while reducing the number of parameters and computational cost, thus accelerating inference speed.

[0070] In addition, the feature fusion part was optimized. Based on the feature maps output by EfficientNet V2 at different stages, and combined with the original feature fusion mechanism of YOLOv12, a cross-scale connection feature fusion method was adopted to enhance the transmission and fusion of feature information, enabling the model to better capture targets at different scales, thereby improving the accuracy of target detection.

[0071] 2) Reduce redundant layers By reducing redundant layers, YOLOv12 is lightweighted, and the deep structure of the network is improved. Visualization tools and feature map analysis are used to identify layers that contribute little to the final detection results. Redundant layers are excessively long layers. Due to the differences in the characteristics of different target types, some layers overcompensate for feature downsampling. For specific object recognition, an experimental method is used to try removing some layers, and the accuracy of the recognition results is analyzed. After training the model, several images are selected for recognition. If the change in the recognition IOU value is less than 0.02, the reduction of redundant layers is considered effective. Through continuous iteration, redundant layers in the network structure are searched and reduced, improving the network's operational efficiency. After reducing redundant layers, the YOLOv12 network structure has fewer layers, achieving network lightweighting.

[0072] See Figure 3 After collecting images of bird activity around overhead power lines as a training set, YOLOv12 was trained using this set to obtain the first weight file, the original pt (a binary file format used to store model weights or the complete model). Then, single layers were selected from multiple layers defined in YOLOv12 (such as layers 1-3, 5, 7, 9, 12, 15, and 18 as described earlier) for removal operations. The YOLOv12 model after these single-layer removal operations was then trained using the training set to obtain the second weight file, the modified pt weight file. Finally, 10 bird images were randomly selected, and the birds were identified using the first weight file, and the results were recorded. Intersection over Union (IoU) values ​​are compared; birds are identified using a second weight file, and their IoU values ​​are recorded; for the first and second weight files, the IoU values ​​under different weight files are compared; it is determined whether the change in IoU value for the same image is less than 0.02, i.e., whether the difference in IoU for the same image does not exceed 0.02. If so, the redundancy reduction layer is considered successful, retained, and the next round of operation is continued; if not, the redundancy reduction layer is considered a failure, discarded, and the next round of operation is continued; through continuous iteration, redundant layers in the network structure are searched and reduced; after reducing redundant layers, the YOLOv12 network structure has fewer layers, achieving network lightweighting.

[0073] The improvement effect of the intelligent recognition model before and after the weight reduction was analyzed, and the results are shown in Table 1.

[0074] Table 1. Optimization effect of lightweight model

[0075] This application embodiment provides a possible implementation method. The capture and recording instruction mentioned in step S203 includes the current recording duration. Therefore, step S203 can also control the camera module to capture bird activities around the overhead transmission line according to the capture and recording instruction containing the current recording duration, adjust the first frame rate to the second frame rate, use the second frame rate, and record according to the current recording duration. The captured images of bird activities around the overhead transmission line and the recorded video of the current recording duration are submitted to the communication module; and the captured images of bird activities around the overhead transmission line and the recorded video of the current recording duration are uploaded to the background server.

[0076] In this embodiment, the current recording duration can be set according to actual needs to accommodate the recording of the complete process of birds flying over overhead lines, while also avoiding wasted recording time. Generally, birds fly over lines quickly, so the current recording duration can be set to 10 seconds. The duration can be appropriately increased or decreased based on factors such as the local bird species and environment.

[0077] Those skilled in the art will clearly understand that the specific working process of the systems, devices, and modules described above can be referred to the corresponding process in the foregoing method embodiments. For the sake of brevity, it will not be repeated here.

[0078] Those skilled in the art will understand that the technical solution of this application, or all or part of it, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several program instructions to cause an electronic device (e.g., a personal computer, server, or network device) to execute all or part of the steps of the methods described in the embodiments of this application when running the program instructions. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, portable hard drive, read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk.

[0079] Alternatively, all or part of the steps of the foregoing method embodiments can be implemented by hardware (such as electronic devices like personal computers, servers, or network devices) associated with program instructions. The program instructions can be stored in a computer-readable storage medium. When the program instructions are executed by the processor of the electronic device, the electronic device executes all or part of the steps of the methods described in the embodiments of this application.

[0080] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit it. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that within the spirit and principles of this application, modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein; and these modifications or substitutions do not cause the corresponding technical solutions to leave the protection scope of this application.

Claims

1. An intelligent device for capturing and recording bird activity around overhead transmission lines, characterized in that, The device includes: The camera module is installed on the power transmission tower and is used to record video of the overhead power transmission line channel at the first frame rate. The edge computing module is used to analyze each frame of the image recorded by the camera module in real time using a pre-trained intelligent recognition model. When birds are identified in the image, the module sends a capture and recording instruction to the camera module, which captures the image and records the video at a second frame rate. The first frame rate is less than the second frame rate. The camera module is also used to capture images of bird activity around overhead power lines according to the capture and recording instructions, adjust the first frame rate to the second frame rate, record video using the second frame rate, and submit the captured images of bird activity around overhead power lines and the recorded video to the communication module. The communication module is used to upload captured images and recorded videos of birds activity around overhead power lines to the backend server. The power module is used to supply power to the camera module, edge computing module, and communication module.

2. The apparatus according to claim 1, characterized in that, The intelligent recognition model in the edge computing module is a lightweight version of YOLOv12. YOLOv12 is an attention-centric real-time object detection model whose network structure includes a backbone network, a feature enhancement module, and a detection head.

3. The apparatus according to claim 2, characterized in that, The lightweight YOLOv12 model is achieved by reducing redundant layers. Due to the differences in the characteristics of different target types, some layers exhibit overcompensation in feature downsampling. For specific object recognition, an experimental method is used to try removing some layers and analyze the accuracy of the recognition results. After training the model, several images are selected for recognition. If the change in the Intersection over Union (IOU) value is less than a preset threshold, the redundant layer reduction operation is considered effective. Through continuous iteration, redundant layers in the network structure are searched and reduced. After reducing redundant layers, the YOLOv12 network structure has fewer layers, achieving network lightweighting.

4. The apparatus according to claim 3, characterized in that, The preset threshold is 0.

02.

5. The apparatus according to claim 1, characterized in that, The capture and recording instruction includes the current recording duration; The camera module is also used to capture bird activity around the overhead transmission line according to the capture recording instruction containing the current recording duration, adjust the first frame rate to the second frame rate, use the second frame rate, and record according to the current recording duration, and submit the captured bird activity pictures around the overhead transmission line and the recorded video of the current recording duration to the communication module. The communication module is used to upload captured images of birds' activities around overhead power lines and recorded videos of the current recording duration to the backend server.

6. The apparatus according to claim 1, characterized in that, The device is applicable to scenarios where no warning devices are installed on overhead power transmission lines.

7. The apparatus according to claim 1, characterized in that, The device is used in scenarios where warning devices are installed on overhead power transmission lines.

8. A method for intelligently capturing and recording bird activity around overhead transmission lines, characterized in that, The method includes: The camera module installed on the transmission tower is controlled to record video at the first frame rate facing the overhead transmission line channel; The edge computing module uses a pre-trained intelligent recognition model to analyze each frame of the image recorded by the camera module in real time. When birds are identified in the image, a capture and recording instruction is sent to the camera module to capture the image and record the video at a second frame rate; where the first frame rate is less than the second frame rate. The camera control module captures images of birds around the overhead transmission line according to the capture and recording command, adjusts the first frame rate to the second frame rate, records the video using the second frame rate, and submits the captured images of birds around the overhead transmission line and the recorded video to the communication module. The communication module uploads captured images and recorded videos of birds activity around overhead power lines to the backend server.

9. The method according to claim 8, characterized in that, The method further includes: Based on YOLOv12, a lightweight version of YOLOv12 is created to obtain an intelligent recognition model. YOLOv12 is an attention-centric real-time object detection model, and its network structure includes a backbone network, a feature enhancement module, and a detection head.

10. The method according to claim 9, characterized in that, Based on YOLOv12, a lightweight version of YOLOv12 is implemented to obtain an intelligent recognition model, including: Collect images of bird activity around overhead power lines as a training set; YOLOv12 is trained using the training set to obtain the first weight file; Sequentially select one layer from the multiple layers set in YOLOv12 and remove it, and use the training set to train YOLOv12 after performing the single-layer removal operation to obtain the second weight file; Select several images, apply the first weight file to identify birds, and record the Intersection over Union (IOU) values; apply the second weight file to identify birds, and record the IOU values. For the IOU values ​​under the first weight file and the second weight file; Determine if the change in IOU value of the same image is less than a preset threshold. If yes, the redundancy reduction layer is considered successful and the next round of operation is executed. If no, the redundancy reduction layer is considered to have failed and the next round of operation is executed. Through continuous iteration, redundant layers in the network structure are searched and reduced to obtain an intelligent recognition model. After reducing redundant layers, the YOLOv12 network structure has fewer layers, achieving network lightweighting.

Images