Low-energy-consumption intelligent bird repelling method and system based on visual identification

Abstract

The invention discloses a low-energy-consumption intelligent bird repelling method and system based on visual recognition, and the method comprises the following steps: a radar outputs a trigger signal after monitoring a moving target, a camera in a dormant state collects an original image according to the trigger signal, and the original image is preprocessed to obtain a preprocessed image; inputting the preprocessed image into a pre-trained improved YOLOv5s-EB model to obtain an identification result; and judging the risk level of the bird species according to the identification result and the trigger signal, and matching a repelling means. The method has the advantages of being high in recognition precision and low in energy consumption, meanwhile, efficient expelling is achieved, and the error expelling rate is reduced.

Description

Technical Field

[0001] This invention relates to the field of visual recognition, specifically a low-energy intelligent bird-repelling method and system. Background Technology

[0002] Safety incidents and economic losses caused by bird activity are becoming increasingly prominent. In particular, bird nesting in power systems causes short circuits in more than 30% of power grid failures, resulting in incalculable losses. Bird strikes at airports seriously threaten flight safety.

[0003] Existing technologies mostly employ single-mode bird deterrence devices and rudimentary visual bird deterrence devices. Single-mode bird deterrence devices rely solely on fixed sound waves, flashes, or other single methods, which birds easily adapt to, resulting in rapid attenuation of deterrence effectiveness. Furthermore, they lack recognition capabilities and have a high false trigger rate. Rudimentary visual bird deterrence devices integrate only basic visual modules, employing traditional recognition algorithms with low accuracy. They also lack low-power design, relying on mains power or requiring frequent battery replacements, leading to poor battery life. In summary, existing technologies have the following drawbacks: 1. Insufficient recognition accuracy and specificity: Traditional visual algorithms have large errors in identifying bird species and distances, and cannot distinguish between protected birds and harmful birds, resulting in a high false alarm rate.

[0004] 2. It consumes too much energy, lacks intelligent sleep mode and adaptive power consumption adjustment mechanism, has poor battery life, and is difficult to operate for a long time in outdoor scenarios without mains power.

[0005] 3. The methods of dispersal are too simplistic and lack a tiered strategy, which makes birds adapt easily and leads to a rapid decline in the success rate of dispersal.

[0006] 4. Lacks remote collaborative management capabilities; data is stored in a scattered manner, making it impossible to achieve centralized cloud monitoring, fault early warning, and strategy optimization.

[0007] 5. Poor environmental adaptability; single visual perception is prone to failure under backlight, obstruction, or inclement weather. Summary of the Invention

[0008] The purpose of this invention is to provide a low-energy intelligent bird deterrence method and system based on visual recognition, which has the advantages of high recognition accuracy and low energy consumption, while achieving efficient bird deterrence and reducing the false bird deterrence rate.

[0009] To achieve the above objectives, the specific solution adopted by the present invention is as follows: a low-energy intelligent bird-repelling method based on visual recognition, comprising the following steps: After the radar detects a moving target, it outputs a trigger signal. The camera, which is in a dormant state, acquires the original image based on the trigger signal. After preprocessing the original image, a preprocessed image is obtained. The preprocessed image is input into the pre-trained improved YOLOv5s-EB model to obtain the recognition result; The risk level of the bird species is determined based on the identification results and trigger signals, and the appropriate repulsion measures are matched accordingly.

[0010] As an optimization of the aforementioned low-energy intelligent bird deterrence method based on visual recognition, the YOLOv5s-EB model includes a backbone network, a neck network, and a head network. The C3 module in the backbone network is replaced with an inverted residual module, and the convolutional layers in the inverted residual module are replaced with depthwise separable convolutions.

[0011] As another optimization of the aforementioned low-energy intelligent bird-repelling method based on visual recognition, a lightweight attention module is embedded between the backbone network and the neck network.

[0012] As another optimization scheme for the aforementioned low-energy intelligent bird-repelling method based on visual recognition, the method for preprocessing the original image is to perform adaptive histogram equalization, Gaussian filtering for noise reduction, and white balance correction on the original image.

[0013] As an alternative optimization of the aforementioned low-energy intelligent bird-repelling method based on visual recognition, the method for determining the risk level of bird species based on recognition results and trigger signals is as follows: Based on the bird species identified in the identification results, a pre-set bird species risk database is queried to obtain the basic risk coefficient for that bird species; The comprehensive risk value is obtained by weighting the basic risk coefficient based on the number of birds in the identification results and the distance of the birds in the trigger signal. Based on the threshold range of the comprehensive risk value, it is determined to be a normal state, a low-risk level, or a high-risk level.

[0014] As another optimization scheme for the aforementioned low-energy intelligent bird deterrence method based on visual recognition: when the comprehensive risk value is less than 2, it is judged as a normal state; When the comprehensive risk value is greater than or equal to 2 and less than 5, it is judged as a low-risk level; When the comprehensive risk value is greater than or equal to 5, it is judged as a high-risk level.

[0015] As another optimization scheme for the aforementioned low-energy intelligent bird-repelling method and system based on visual recognition, the matching method of repelling means is as follows: When the condition is determined to be normal, the original image is captured at a low frame rate and uploaded every 30 minutes. When the risk level is determined to be low, directional acoustic waves are activated to drive them away. When a high-risk level is determined, lasers, enemy sounds, and strobe lights are activated simultaneously to drive away the foes.

[0016] A low-energy intelligent bird-repelling system based on visual recognition includes a cloud platform, a multimodal perception terminal module, an edge AI low-energy computing module, and a bird-repelling module. The multimodal sensing terminal module includes a camera, radar, and infrared sensor. After the radar detects a moving target, it outputs a trigger signal. The camera, which is in a dormant state, acquires the original image based on the trigger signal. After preprocessing the original image, a preprocessed image is obtained. The edge AI low-energy computing module includes a pre-trained improved YOLOv5s-EB model. The pre-processed image is input into the pre-trained improved YOLOv5s-EB model to obtain the recognition result. The bird deterrence module includes a sound wave generator, a laser emitter, and a strobe light. It determines the risk level of the bird species based on the identification results and trigger signals, and matches the deterrence method accordingly.

[0017] As an optimized solution for the aforementioned low-energy intelligent bird-repelling system based on visual recognition: the multimodal sensing terminal module includes a shell, and the camera, radar and infrared sensor are all installed inside the shell.

[0018] As another optimization of the aforementioned low-energy intelligent bird-repelling system based on visual recognition: the outer shell is coated with an anti-ultraviolet coating.

[0019] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention provides a low-energy intelligent bird deterrence method and system based on visual recognition. It uses radar for monitoring, while the camera and edge AI low-energy computing module are in a dormant state. When the radar detects a moving target, it generates a trigger signal, which triggers the ring camera and edge AI low-energy computing module, greatly reducing energy consumption.

[0020] 2. This invention adopts an improved YOLOv5s-EB model, combining the inverse residual module with deep learning convolution, which significantly reduces the amount of computation while ensuring high accuracy, effectively filters out non-target interference, and significantly improves the accuracy and reliability of bird identification.

[0021] 3. This invention employs multimodal perception, which determines the risk level based on the bird's type, number, and distance, and matches it with corresponding graded repulsion measures to avoid problems such as insufficient bird adaptation or response, thereby minimizing interference and optimizing the repulsion effect. Detailed Implementation

[0022] The technical solution of the present invention will be further described in detail below with reference to specific embodiments. Parts not described or disclosed in detail in the following embodiments of the present invention should be understood as prior art known or should be known by those skilled in the art.

[0023] Example 1 A low-energy intelligent bird-repelling method based on visual recognition includes the following steps: Real-time monitoring is achieved using radar, while the camera, edge AI low-power computing module, and de-escalation module remain in sleep mode to reduce energy consumption. In this embodiment, millimeter-wave radar is used. Upon detecting a moving target, the radar outputs a trigger signal. The camera, in sleep mode, acquires raw images based on the trigger signal. After preprocessing, a preprocessed image is obtained. Specifically, the radar scans its monitoring area at a fixed frequency to monitor for moving targets in real time. When the radar detects a moving target, it filters it to prevent irrelevant targets such as flying insects from frequently waking the camera. When the radar confirms that one or more moving targets match bird characteristics, it generates a trigger signal and transmits it to the control unit. Upon receiving the trigger signal, the control unit wakes up the camera and quickly turns it towards the direction of the moving target indicated by the radar. The camera acquires one or more frames of raw images and transmits them to the edge AI computing module.

[0024] After system startup, the control unit defaults to putting the camera, edge AI low-power computing module, and de-escalation module into sleep mode, activating only the millimeter-wave radar for real-time monitoring to minimize system power consumption. The millimeter-wave radar operates according to preset functional scanning logic. Based on the environmental characteristics of the monitoring area, it performs a 360-degree scan of the monitoring range at a fixed frequency adapted to the scene, continuously capturing the presence of moving targets within the area. During the scan, it collects the target's motion parameters in real time and transmits them to the control unit.

[0025] After the radar transmits moving target data to the control unit, a built-in target feature filtering mechanism is activated to exclude irrelevant targets. This mechanism identifies key target features to distinguish between effective monitoring targets such as birds and wild animals and interfering targets such as flying insects, fallen leaves, and dust, avoiding energy waste and response redundancy caused by irrelevant targets frequently waking up subsequent modules. When the control unit confirms that one or more moving targets meet the preset effective features, it immediately generates a trigger signal containing key information such as the target's location and approximate distance, which is simultaneously transmitted to the camera and the edge AI low-power computing module.

[0026] Upon receiving a trigger signal, the camera quickly wakes up from sleep mode. Based on the target location information transmitted by the control unit, it rapidly adjusts the shooting angle via a built-in steering drive mechanism to precisely align with the target's active area. It then initiates image acquisition, capturing one or more frames of raw images that clearly reflect the target's features. After acquisition, the raw image is transmitted in real-time to the edge AI low-power computing module via a data bus. This module is simultaneously awakened and initiates an image preprocessing process: first, it denoises the raw image, filtering out noise caused by changes in ambient light and atmospheric scattering; second, it enhances the image, optimizing the contrast between the target and background to highlight the target's outline and details; finally, it crops the image based on the target's location information, retaining the effective area containing the target and removing redundant background to reduce the computational load for subsequent AI recognition, adapting to the low-power operation requirements of the edge computing module. The final output is a clear, focused preprocessed image, laying the foundation for accurate target recognition.

[0027] Before recognizing the original image, it needs to be preprocessed. Specifically, the preprocessing methods for the original image include adaptive histogram equalization, Gaussian filtering for noise reduction, and white balance correction. Adaptive histogram equalization involves: dividing the original image into several non-overlapping local regions; calculating the gray-level histogram for each local region; performing an equalization transformation on the gray-level histogram using a cumulative distribution function to make the gray-level distribution of the local region more uniform; and using a bilinear yield algorithm to smooth the transition of the transformation function between adjacent local regions, avoiding gray-level abrupt changes at the boundaries of local regions and ensuring a natural gray-level variation throughout the image. This significantly enhances the contrast between bird targets and complex environments, making details such as bird outlines and feather textures clearer, especially improving the visibility of small birds.

[0028] Adaptive histogram equalization works as follows: Based on the resolution of the original image captured by the camera, a dynamic block-segmentation strategy is used to divide the local region. The original image is divided into several non-overlapping rectangular blocks, with the block size adaptively adjusted according to the size of the bird target. For each segmented local region, a gray-level histogram is calculated block by block using the gray-level statistical algorithm built into the edge AI module. A cumulative distribution function (CDF) is constructed based on the gray-level histogram, and the gray-level values ​​of the region are mapped and transformed using the CDF. The original gray-level distribution range is uniformly expanded to the full gray-level range of 0-255, making the gray-level distribution more dispersed in areas where gray levels were originally concentrated, thereby enhancing the contrast of local details. To avoid gray-level abrupt changes at block boundaries caused by local region segmentation, a bilinear interpolation algorithm is used to fuse the gray-level transformation functions of two adjacent local regions. Specifically, for pixels at block boundaries, the transformed gray-level values ​​of both the pixel's own block and adjacent blocks are referenced, and weights are assigned according to the distance from the pixel to the center of the two blocks. A weighted average gray-level value is calculated as the final pixel value, ensuring that the gray-level changes of the entire image are natural and consistent, preserving the local optimization effect while eliminating block effect interference.

[0029] Gaussian filtering denoising can suppress noise in the original image and improve the accuracy of subsequent recognition. Specifically, the Gaussian filter kernel is selected based on the type of outdoor noise. For common Gaussian distribution noise in the environment, a 3×3 Gaussian filter kernel is used (balancing denoising effect and computational efficiency, adapting to the low power consumption requirements of the edge module; if the monitored environment noise is strong, it can be expanded to a 5×5 kernel, but the computational logic needs to be optimized simultaneously to control energy consumption). The standard deviation σ of the filter kernel is set to an adaptive range of 0.8-1.2, which is automatically adjusted by the edge module according to the noise intensity of the original image: when the noise intensity is high, the value of σ is larger (1.0-1.2) to enhance the denoising capability; when the noise intensity is low, the value of σ is smaller (0.8-1.0) to reduce excessive blurring of target details.

[0030] The Gaussian filter kernel is traversed throughout the preprocessed image using a sliding window approach. For each pixel within the window, a weighted average gray value is calculated based on the weight distribution of the Gaussian kernel, and this average value replaces the original value of the pixel at the center of the window. Since the weight distribution of the Gaussian filter conforms to a normal distribution, it can effectively smooth out abrupt gray-level changes caused by noise. At the same time, because the weights are concentrated at the center, it can preserve the gray-level differences of the target edges to the greatest extent, avoiding excessive blurring of the bird's outline.

[0031] White balance correction includes: calculating the average value of the three RGB channels of the image; using the green channel as a reference, calculating the gain coefficients of the red and blue channels; and correcting the red and blue components of each pixel. This ensures color consistency in images of the same bird species collected at different times and under different weather conditions, significantly improving the accuracy and stability of the recognition model and avoiding misidentification of different bird species due to color differences caused by lighting.

[0032] Specifically, for the Gaussian-filtered image, the average grayscale value of all pixels across the RGB channels is calculated (i.e., the sum of the pixel grayscale values ​​for the red channel R, green channel G, and blue channel B is calculated separately and then divided by the total number of pixels). The green channel is chosen as the reference because it has the lowest sensitivity to changes in light in natural scenes, and the green and yellow characteristics of bird feathers are more stable in the green channel, effectively reducing the interference of extreme light on the reference.

[0033] Gain coefficient calculation logic: based on the average value G of the green channel. avg Using this as a standard, calculate the gain coefficient K of the red channel. r =G avg / R avg The gain coefficient K of the blue channel b =G avg / B avg If R avg >G avg This indicates that the overall image is reddish, K r <1, reduce the grayscale value of the red channel by adjusting the gain coefficient; if B avg >G avg This indicates that the image is generally bluish, K b A value less than 1 indicates a decrease in the grayscale value of the blue channel; conversely, a value greater than 1 indicates an increase in the grayscale value of the corresponding channel, thus achieving color balance across the three channels.

[0034] Pixel-level color correction: This involves correcting the R and B components of each pixel in the image point-by-point. The correction formula is: R... corrected =R original ×K r B corrected =B original ×K b The green channel retains its original grayscale value. During the correction process, the grayscale values ​​of the corrected R and B components need to be cropped to ensure that their range is limited to 0-255, avoiding pixel saturation (pure white or pure black) due to excessive gain coefficients, and ensuring natural color transitions. For example, bird images with warmer morning light (high proportion of red channel) will have their colors restored to the true body color of the bird species after correction, maintaining consistency with the colors of images under midday light conditions.

[0035] The preprocessed image is input into a pre-trained improved YOLOv5s-EB model to obtain the recognition result. The YOLOv5s-EB model includes a backbone network, a neck network, and a head network. The backbone network extracts multi-level, abstract feature maps from the preprocessed image. The neck network fuses the feature maps extracted by the backbone network at different scales, i.e., it performs feature fusion, enhancing the detection capability for small and occluded targets. The head network recognizes the feature maps after feature fusion and outputs the recognition result. In this embodiment, the C3 module in the backbone network is replaced with an inverse residual module, and the convolutional layers in the inverse residual module are replaced with depthwise separable convolutions, which include depthwise convolutions and pointwise convolutions. The inverse residual module first performs pointwise convolutions—firstly, it uses a 1×1 standard convolution to expand the number of channels in the low-dimensional input feature map, which increases the network's expressive capacity. Then, it performs depthwise convolutions: a 3×3 depthwise convolution is performed on the expanded feature map to efficiently extract spatial features. Finally, pointwise convolution is performed: 1×1 standard convolution is used again to compress the number of channels back to the target dimension, and residual connections are introduced to directly add the input to the output to alleviate gradient vanishing.

[0036] To compensate for the potential decrease in feature discrimination power due to lightweighting, a lightweight attention module is embedded before the backbone network output is fed into the neck network for multi-scale fusion. Specifically, global average pooling compresses the two-dimensional features of each channel into a scalar, capturing the global information of that channel. One-dimensional convolution is performed on the pooled channel vectors to achieve cross-channel information interaction, avoiding the huge number of parameters brought by fully connected layers and achieving lightweighting. Sigmoid activation generates weight coefficients for each channel. Recalibration multiplies the weight coefficients with the original feature map channel by channel, strengthening important feature channels and suppressing secondary channels.

[0037] The risk level of a bird species is determined based on the identification results and trigger signals, and appropriate deterrence measures are then applied. The method for determining the risk level of a bird species based on the identification results and trigger signals involves assessing the species and number of birds identified and the distance of the birds indicated by the trigger signals. Specifically, this includes the following steps: Based on the bird species identified in the results, a pre-set bird species risk database is consulted to obtain the basic risk coefficient for that species. It should be noted that this bird species risk database contains basic risk coefficients for common bird species, and these coefficients are based on the bird's ecological hazard, size, impact energy, and historical accident statistics. A comprehensive risk value is obtained by weighting the basic risk coefficients based on the number of birds in the identification results and the distance of the birds in the trigger signal; the more birds there are and the closer they are, the higher the risk.

[0038] Based on the threshold range of the comprehensive risk value, it is determined to be a normal state, a low-risk level, or a high-risk level.

[0039] When the comprehensive risk value is less than 2, it is judged as a normal state; when the comprehensive risk value is greater than or equal to 2 and less than 5, it is judged as a low-risk level; when the comprehensive risk value is greater than or equal to 5, it is judged as a high-risk level.

[0040] The method for matching removal methods is as follows: When the situation is determined to be normal, raw images are captured at a low frame rate and uploaded every 30 minutes for monitoring purposes only, without initiating any decoys. When the situation is determined to be low-risk, directional acoustic waves are activated for decoys. When the situation is determined to be high-risk, laser, predator sound, and strobe lights are activated simultaneously for decoys.

[0041] Example 2 A low-energy intelligent bird-repelling system based on visual recognition includes a cloud platform, a multimodal perception terminal module, an edge AI low-energy computing module, and a bird-repelling module.

[0042] The cloud platform is built on a B / S architecture, supporting access from both web and mobile devices. It features four core functions: first, data visualization of bird species, numbers, and distances; second, a four-level early warning mechanism for equipment failures, high-risk bird intrusions, and low battery status; third, remote operation and maintenance; and fourth, big data analysis to uncover bird activity patterns and optimize the timing and intensity of bird removal. The platform employs AES encrypted transmission technology to ensure data security, enabling centralized management of multiple devices and in-depth data utilization, constructing a complete operation and maintenance loop, and improving management efficiency.

[0043] In this embodiment, the bird-repelling system is equipped with a power supply module consisting of a monocrystalline silicon solar panel, a lithium iron phosphate battery, and a controller, which can meet the needs of the whole day and ensure long-term operation of the equipment in the absence of mains power. The multimodal sensing terminal module and the edge AI low-power computing module are connected via a CAN bus. The edge module uses 5G / Wi-Fi dual-mode communication to interface with the cloud platform, prioritizing 5G high-speed transmission. In the event of a network outage, it automatically switches to Wi-Fi and caches data, uploading synchronously once the network is restored. Image acquisition cycle: When no birds are present, recognition accuracy is improved upon bird detection. Abnormal data is uploaded in real time, while normal data is aggregated and uploaded every 30 minutes, ensuring stable data transmission and low latency, achieving full-process linkage of sensing, analysis, and control.

[0044] The multimodal sensing terminal module includes a housing, a camera, a radar, and an infrared sensor. The camera, radar, and infrared sensor are all housed within the housing, which is coated with an anti-UV coating. The camera uses a wide-angle lens, the radar is a 24GHz millimeter-wave radar, and the infrared sensor has an extremely short response time, making it suitable for day and night monitoring in all scenarios. The multimodal sensing terminal module can rotate 360°. When the radar detects a moving target, it outputs a trigger signal. The camera, which is in sleep mode, acquires the raw image based on the trigger signal. After preprocessing the raw image, a preprocessed image is obtained.

[0045] The edge AI low-power computing module uses the RK3588 lightweight AI chip as its core component and includes a pre-trained improved YOLOv5s-EB model and a control unit. The control unit receives trigger signals and outputs wake-up signals to activate the camera. Pre-processed images are input into the pre-trained improved YOLOv5s-EB model to obtain recognition results. The edge AI low-power computing module has 8GB of built-in local storage and supports offline data caching to ensure data integrity. This step enables local real-time intelligent analysis, balancing recognition accuracy and energy consumption, providing a rapid response for decoy decisions, and reducing reliance on the cloud.

[0046] The bird deterrence module includes a sound wave generator, a laser emitter, and a strobe light. It determines the risk level of the bird species based on the identification results and trigger signals, and matches the deterrence method accordingly.

[0047] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A low-energy intelligent bird-repelling method based on visual recognition, characterized in that, Includes the following steps: After the radar detects a moving target, it outputs a trigger signal. The camera, which is in a dormant state, acquires the original image based on the trigger signal. After preprocessing the original image, a preprocessed image is obtained. The preprocessed image is input into the pre-trained improved YOLOv5s-EB model to obtain the recognition result; The risk level of the bird species is determined based on the identification results and trigger signals, and the appropriate repulsion measures are matched accordingly.

2. The low-energy intelligent bird-repelling method based on visual recognition as described in claim 1, characterized in that: The YOLOv5s-EB model consists of a backbone network, a neck network, and a head network. The C3 module in the backbone network is replaced with an inverted residual module, and the convolutional layers in the inverted residual module are replaced with depthwise separable convolutions.

3. The low-energy intelligent bird-repelling method based on visual recognition as described in claim 2, characterized in that: A lightweight attention module is embedded between the backbone network and the neck network.

4. The low-energy intelligent bird-repelling method based on visual recognition as described in claim 1, characterized in that: The method for preprocessing the original image is to perform adaptive histogram equalization, Gaussian filtering for noise reduction, and white balance correction.

5. The low-energy intelligent bird-repelling method based on visual recognition as described in claim 1, characterized in that: The method for determining the risk level of bird species based on identification results and trigger signals is as follows: Based on the bird species identified in the identification results, a pre-set bird species risk database is queried to obtain the basic risk coefficient for that bird species; The comprehensive risk value is obtained by weighting the basic risk coefficient based on the number of birds in the identification results and the distance of the birds in the trigger signal. Based on the threshold range of the comprehensive risk value, it is determined to be a normal state, a low-risk level, or a high-risk level.

6. The low-energy intelligent bird-repelling method based on visual recognition as described in claim 5, characterized in that: When the overall risk value is less than 2, it is considered to be in a normal state; When the comprehensive risk value is greater than or equal to 2 and less than 5, it is judged as a low-risk level; When the comprehensive risk value is greater than or equal to 5, it is judged as a high-risk level.

7. The low-energy intelligent bird-repelling method based on visual recognition as described in claim 5, characterized in that: The method for matching removal methods is as follows: When the condition is determined to be normal, the original image is captured at a low frame rate and uploaded every 30 minutes. When the risk level is determined to be low, directional acoustic waves are activated to drive them away. When a high-risk level is determined, lasers, enemy sounds, and strobe lights are activated simultaneously to drive away the foes.

8. A low-energy intelligent bird-repelling system based on visual recognition, characterized in that: This includes a cloud platform, a multimodal sensing terminal module, an edge AI low-power computing module, and a decoy module; The multimodal sensing terminal module includes a camera, radar, and infrared sensor. After the radar detects a moving target, it outputs a trigger signal. The camera, which is in a dormant state, acquires the original image based on the trigger signal. After preprocessing the original image, a preprocessed image is obtained. The edge AI low-energy computing module includes a pre-trained improved YOLOv5s-EB model. The pre-processed image is input into the pre-trained improved YOLOv5s-EB model to obtain the recognition result. The bird deterrence module includes a sound wave generator, a laser emitter, and a strobe light. It determines the risk level of the bird species based on the identification results and trigger signals, and matches the deterrence method accordingly.

9. A low-energy intelligent bird-repelling system based on visual recognition as described in claim 8, characterized in that: The multimodal sensing terminal module includes a housing, and the camera, radar, and infrared sensor are all housed inside the housing.

10. A low-energy intelligent bird-repelling system based on visual recognition as described in claim 8, characterized in that: The outer shell is coated with an anti-ultraviolet coating.