Fan forest large-scale video monitoring system based on image algorithm

Abstract

The present application relates to the technical field of video monitoring, in particular to a large-scale wind turbine forest area video monitoring system based on image algorithm. The system first acquires monitoring video frames and wind turbine operation state data containing relative wind direction, and performs time lag compensation and visual main motion direction calibration based on the acquisition time difference to construct an image wind direction reference system that fuses working conditions and visual features. Second, a global affine transformation model is fitted to eliminate global rigid displacement caused by cabin vibration, and local relative motion vectors are obtained. Then the vectors are decomposed into the reference system upwind and crosswind directions, and the one-way motion weight is calculated according to the time domain flip frequency of the upwind component to filter out vegetation interference by weighting and suppressing the two direction components to generate an abnormal motion intensity map. Finally, an adaptive alarm threshold is generated based on the background noise distribution of the intensity map to identify abnormal areas. The present application effectively solves the false alarm and missed alarm problems in strong wind and cabin vibration environment, and improves the reliability of security in wind power forest area.

Description

Technical Field

[0001] This invention relates to the field of video surveillance technology, specifically to a large-scale video surveillance system for wind turbine forest areas based on image algorithms. Background Technology

[0002] Wind farms are often built in remote, densely vegetated forest areas. To prevent fires and illegal intrusions, video surveillance is an essential security facility. Since wind turbine towers are typically over 100 meters high and widely distributed, installing security cameras on the nacelles has become an economical and efficient deployment method, utilizing the nacelle's open field of view and existing power and communication links for large-scale monitoring. However, during operation, the wind turbine nacelles are affected by wind loads and blade rotation, resulting in irregular vibrations and yaw movements. This causes global displacement in the monitoring image, including translation, rotation, and scaling, severely undermining the assumptions of motion detection algorithms based on background modeling (such as Gaussian mixture models). Furthermore, forest areas experience high and drastically changing wind speeds, causing ground vegetation to sway dramatically under strong winds. Frame difference methods or optical flow methods struggle to distinguish this high-frequency background dynamics from earlier smoke diffusion or human movement, easily misreporting swaying trees as abnormal targets.

[0003] Existing image stabilization algorithms can eliminate global displacement caused by nacelle vibration to some extent, but they struggle to handle dynamic background interference from large-scale vegetation swaying violently in the wind. Meanwhile, conventional target detection algorithms are mostly based on static texture features, making them susceptible to interference at long distances, under low pixel counts, and in complex lighting conditions. To reduce false alarm rates, systems are often forced to raise detection thresholds, leading to missed detections of minor fires or intrusions during high-risk, windy weather. Furthermore, while relying solely on physical data from the wind turbine system can reflect macroscopic wind conditions, it cannot accurately correct for local terrain turbulence deviations in the image. Therefore, existing video detection algorithms, lacking a deep fusion mechanism of multi-source operating data and visual features, cannot promptly and accurately identify abnormal areas within the wind turbine forest area, affecting the reliability and environmental adaptability of wind farm security systems and failing to meet the refined management needs of unmanned wind farms. Summary of the Invention

[0004] To address the technical problem that existing video detection algorithms lack a deep fusion mechanism between multi-source operating data and visual features, thus failing to accurately and promptly identify abnormal areas within wind turbine forest areas, this invention aims to provide a large-scale video monitoring system for wind turbine forest areas based on image algorithms. The specific technical solution adopted is as follows:

[0005] This invention provides a large-scale video monitoring system for wind turbine forest areas based on image algorithms. The system includes:

[0006] The data acquisition module is used to acquire video frames from the wind turbine forest area monitoring camera and the wind turbine operating status data in real time; among which, the status data includes the relative wind direction;

[0007] The image wind direction reference system acquisition module is used to perform time lag compensation based on the acquisition time difference between the state data and the video frame, and determine the reference state data aligned with the current video frame; according to the visual main motion direction of the current video frame, the relative wind direction contained in the reference state data is calibrated to construct an image wind direction reference system containing the downwind direction and the crosswind direction.

[0008] The abnormal motion intensity map acquisition module is used to perform full-image motion estimation on the current video frame and fit a global affine transformation model, thereby eliminating global rigid displacements and obtaining the local relative motion vector of each pixel in the current video frame; the local relative motion vector is projected and decomposed to the downwind and crosswind directions of the image wind direction reference frame; based on the direction flipping frequency of the downwind component of each pixel in the time domain, the unidirectional motion weight of each pixel is obtained; and then the downwind and crosswind components are weighted and suppressed to generate an abnormal motion intensity map;

[0009] The abnormal area identification module is used to obtain an adaptive alarm threshold that fluctuates with wind conditions based on the background noise distribution characteristics of the abnormal motion intensity map, and then identify abnormal areas.

[0010] Furthermore, the method for obtaining the image wind direction reference system is as follows:

[0011] Set up an image plane coordinate system that includes a positive direction definition and a cabin physical coordinate system with the cabin centerline as the reference. The difference between the relative wind direction contained in the reference state data and the corresponding angle of the current pan-tilt azimuth angle of the monitoring camera in the cabin physical coordinate system is circumferentially normalized and used as the current initial image wind direction angle.

[0012] Temporal stability detection is performed on the visual main motion direction of the current video frame. If and only if the rate of change of the angle of the visual main motion direction for a consecutive preset number of frames is lower than a preset stability threshold, the absolute value of the minimum circumferential angle difference between the corresponding angle of the visual main motion direction of the current video frame in the image plane coordinate system and the wind direction angle of the initial image is used as the deviation analysis value.

[0013] When the deviation analysis value is less than the preset deviation threshold, the direction angle of the vector obtained by weighted summation of the unit vector corresponding to the visual main motion direction of the current video frame and the unit vector corresponding to the wind direction angle of the initial image is used as the calibrated target image wind direction angle; wherein, the weight of the initial image wind direction angle is set to be greater than the weight of the visual main motion direction.

[0014] When the deviation analysis value is greater than or equal to the preset deviation threshold, or when the visual main motion direction fails the temporal stability detection, the initial image wind direction angle is taken as the target image wind direction angle.

[0015] Based on the wind direction angle of the target image, a unit vector is constructed to determine the downwind and crosswind directions of the image plane coordinate system, which serve as the image wind direction reference system.

[0016] Furthermore, the method for obtaining the main visual motion direction is as follows:

[0017] Perform full-map motion estimation on the current video frame and its previous adjacent video frame to obtain the basic motion vector field; construct an orientation histogram containing at least two angular intervals based on the basic motion vector field.

[0018] Statistical analysis of the direction histogram is performed, and the center angle of the angle interval with the highest frequency is taken as the main visual motion direction of the current video frame.

[0019] Furthermore, the method for obtaining the local relative motion vector is as follows:

[0020] Based on the random sampling consensus algorithm, a six-parameter affine transformation model is fitted using sample points in the basic motion vector field; the affine transformation model represents the global rigid displacement.

[0021] Based on the affine transformation model, the theoretical background displacement vector of each pixel in the current video frame is obtained;

[0022] The vector difference between the vector corresponding to each pixel in the basic motion vector field and its theoretical background displacement vector in the current video frame is used as the local relative motion vector of each pixel in the current video frame.

[0023] Furthermore, the method for obtaining the unidirectional motion weights is as follows:

[0024] In response to the detection that the pan-tilt unit of the surveillance camera is in a stationary and locked state, for any pixel in the current video frame, the downwind component of that pixel is stored in the corresponding first-in-first-out preset storage queue in real time.

[0025] When the frame number of the current video frame is greater than the length of the preset storage queue, the direction of the downwind component in the preset storage queue is flipped according to the specified conditions.

[0026] The ratio of the number of direction flips in the preset storage queue to the length of the preset storage queue is used as the direction flip frequency;

[0027] The result of mapping the direction reversal frequency using the negative exponential decay function is used as the weight for unidirectional motion.

[0028] When the frame number of the current video frame is less than or equal to the length of the preset storage queue, 1 is used as the unidirectional motion weight.

[0029] Furthermore, the specified conditions are:

[0030] For any downwind component in the storage queue, the absolute value of the downwind component and its next adjacent downwind component are both greater than the preset dead zone threshold.

[0031] The product of the downwind component and its next adjacent downwind component is less than 0.

[0032] Furthermore, the method for obtaining the abnormal motion intensity map is as follows:

[0033] For any pixel in the current video frame, the absolute value of the downwind component of that pixel is taken as the first value; the absolute value of the crosswind component of that pixel is taken as the second value.

[0034] The weighted sum of the first and second values ​​is used as the base intensity value of the pixel.

[0035] The product of the unidirectional motion weight of the pixel and the base intensity value is taken as the abnormal motion intensity of the pixel.

[0036] The image composed of the abnormal motion intensity of each pixel in the current video frame is used as the abnormal motion intensity map.

[0037] Furthermore, the method for obtaining the adaptive alarm threshold is as follows:

[0038] The exponentially weighted moving average algorithm is used to smooth the abnormal motion intensity map in the time domain to obtain the risk accumulation map.

[0039] Obtain the mean and standard deviation of the remaining pixels in the risk accumulation map, excluding the pixel with the highest value in the previous preset proportion, and use them as the background mean and background standard deviation, respectively.

[0040] The sum of the background mean and the weighted background standard deviation is used as the dynamic floating threshold;

[0041] The dynamic floating threshold is compared with the preset minimum noise floor threshold, and the larger value is used as the adaptive alarm threshold that fluctuates with the wind conditions.

[0042] Furthermore, the method for obtaining the abnormal region is as follows:

[0043] The risk accumulation map is binarized using an adaptive alarm threshold, and connected regions are extracted as suspected target regions.

[0044] The system obtains the installation latitude and longitude, current focal length, pan-tilt azimuth angle, and pan-tilt pitch angle of the surveillance camera. Based on a preset digital elevation model, it uses the principle of monocular perspective projection to map the pixel coordinates of the suspected target area to geographic coordinates, thereby obtaining the abnormal area.

[0045] Furthermore, the method for obtaining the reference state data is as follows:

[0046] The system acquires wind speed data in real time from the status data and uses the average of all wind speed data within a preset time period as the wind speed reference value.

[0047] When the wind speed reference value is greater than the preset wind speed threshold, the acquisition time difference is obtained based on the installation height of the monitoring camera and the wind speed reference value, combined with the preset relationship model.

[0048] When the wind speed reference value is less than or equal to the preset wind speed threshold, the preset fixed duration will be used as the data collection time difference.

[0049] Starting from the current moment, the state data collected at historical moments of the time difference are used as the reference state data for aligning the current video frame.

[0050] The present invention has the following beneficial effects:

[0051] This invention first compensates for the time lag between the acquisition of state data and video frames to determine the reference state data aligned with the current video frame. This helps solve the problem of time-dimensional asynchrony between high-altitude wind turbine data and ground visual data, ensuring the physical consistency of multi-source data. To correct local wind direction deviations caused by terrain undulations, the relative wind direction contained in the reference state data is calibrated according to the main visual motion direction of the current video frame, constructing an image wind direction reference system that includes both downwind and crosswind directions. This provides a geometric benchmark that accurately reflects the true swaying direction of vegetation, improving subsequent decoupling accuracy. To eliminate background motion interference caused by nacelle vibration and yaw, the current video frame undergoes full-image motion estimation and a global affine transformation model is fitted to eliminate global rigid displacements. This obtains the local relative motion vector of each pixel in the current video frame, accurately reflecting the independent motion state of objects relative to the ground, which helps focus on the dynamic changes of the scene itself. Furthermore, the local relative motion vector is projected and decomposed onto the image wind direction reference system. Wind direction and crosswind direction transform indiscriminate pixel motion into orthogonal components with clear physical meaning, namely, windward and crosswind directions. Furthermore, based on the direction flipping frequency of the windward direction component of each pixel in the time domain, the reciprocating motion characteristics of the object along the wind axis are accurately reflected. This allows for the accurate acquisition of the unidirectional motion weight of each pixel, accurately reflecting the credibility of whether the current motion belongs to a unidirectional transport target (such as smoke) or a reciprocating oscillating interference (such as vegetation), which helps to distinguish the dynamic background from the real target under strong winds. Weighted suppression of the windward and crosswind components generates an abnormal motion intensity map, which helps to significantly attenuate vegetation interference energy while preserving the real target signal, significantly improving the signal-to-noise ratio. Based on the background noise distribution characteristics of the abnormal motion intensity map, an adaptive alarm threshold that fluctuates with wind conditions is accurately obtained, accurately reflecting the background noise level under the current wind conditions, thus accurately identifying abnormal areas. This facilitates the generation of high-precision alarm intelligence and effectively improves the environmental adaptability and detection reliability of the wind farm forest area security system. Attached Figure Description

[0052] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0053] Figure 1 This is a structural block diagram of a large-scale video monitoring system for wind turbine forest areas based on image algorithms, provided in one embodiment of the present invention.

[0054] Figure 2This is a schematic diagram of a computer device provided according to an embodiment of the present invention. Detailed Implementation

[0055] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of the large-scale video monitoring system for wind turbine forest areas based on image algorithms proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0056] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0057] The following description, in conjunction with the accompanying drawings, details the specific solution of the large-scale video monitoring system for wind turbine forest areas based on image algorithms provided by this invention.

[0058] Example 1:

[0059] This invention proposes a large-scale video monitoring system for wind turbine forest areas based on image algorithms. Please refer to [link / reference]. Figure 1 The diagram shows a structural block diagram of a large-scale video monitoring system for wind turbine forest areas based on image algorithms provided in an embodiment of the present invention. The system includes: a data acquisition module 10, an image wind direction reference system acquisition module 20, an abnormal motion intensity map acquisition module 30, and an abnormal area identification module 40.

[0060] The data acquisition module 10 is used to acquire video frames from the wind turbine forest area monitoring camera and the status data of wind turbine operation in real time; among which, the status data includes the relative wind direction.

[0061] Specifically, to ensure simultaneous analysis of the monitoring scene from both visual and physical perspectives, this embodiment acquires video frames from the wind turbine forest area monitoring camera in real time via a video stream interface, and reads the wind turbine's operational status data in real time via the wind turbine monitoring system's data interface. This status data includes at least relative wind direction (i.e., the angle between the wind direction and the nacelle's central axis) and wind speed data, accurately reflecting the environmental wind conditions around the wind turbine and the wind load on the nacelle. Simultaneously, the camera's current focal length (or zoom ratio), pan / tilt azimuth angle, and pan / tilt pitch angle are acquired in real time, providing necessary parameters for calculating camera intrinsic parameters using focal length and extrinsic parameters using azimuth and pitch angles to achieve monocular positioning. This embodiment sets the video frame acquisition frequency to 25Hz to ensure the continuity of visual information. Considering the characteristics of industrial control data, the update frequency of status data is typically low; therefore, the status data acquisition frequency is set to 1Hz. Implementers can set the video frame acquisition frequency and status data acquisition frequency according to actual conditions, without limitation here.

[0062] Since the video frame acquisition frequency (e.g., 25Hz) is usually higher than the state data acquisition frequency (e.g., 1Hz), in order to achieve data alignment, the system uses zero-order hold or linear interpolation to upsample the low-frequency state data to the same temporal resolution as the video frame, or specifies that each video frame is associated with the most recently updated state data before its acquisition time. Zero-order hold and linear interpolation are existing concepts and will not be elaborated upon further.

[0063] The image wind direction reference system acquisition module 20 is used to perform time lag compensation based on the acquisition time difference between the state data and the video frame, and determine the reference state data aligned with the current video frame; according to the visual main motion direction of the current video frame, the relative wind direction contained in the reference state data is calibrated to construct an image wind direction reference system containing the downwind direction and the crosswind direction.

[0064] Specifically, given that wind turbine hub height is typically over 100 meters, there is a significant physical time lag (i.e., acquisition time difference) in the transmission of airflow from high altitudes to the ground vegetation layer. Directly using current operational data to analyze vegetation dynamics in ground images would lead to the failure of physical constraints, as the swaying state of the ground vegetation would actually be a response to past high-altitude wind conditions, not the current wind conditions, thus affecting the accuracy of subsequent analysis. Therefore, this embodiment compensates for the time lag based on the acquisition time difference between state data and video frames, retrospectively acquiring historical state data matching the moment the airflow reaches the ground. This accurately determines the reference state data aligned with the current video frame, effectively solving the problem of temporal asynchrony between high-altitude physical data and ground visual data, ensuring the physical consistency of multi-source data.

[0065] Considering that the undulating terrain of the forest area may generate local turbulence effects, resulting in a deviation between the actual wind direction on the ground and the physical wind direction at high altitude; and in order to compensate for the limitations of relying solely on physical sensor data, this embodiment further calculates the visual main motion direction of the current video frame and uses its visual features to calibrate the relative wind direction contained in the reference state data, ensuring that the constructed image wind direction reference system accurately reflects the real physical axis of vegetation swaying in the current scene. Finally, an image wind direction reference system including the downwind direction and the crosswind direction is constructed, providing a reliable geometric benchmark for subsequent accurate separation of vegetation interference and target movement.

[0066] Preferably, in one feasible method of this embodiment, the method for obtaining the reference state data is as follows: real-time acquisition of wind speed data included in the state data; in order to eliminate the fluctuation noise of single instantaneous wind speed data and improve the robustness of the data, the average value of all wind speed data within a preset time period is used as the wind speed reference value; in this embodiment, the duration of the preset time period is set to 10 seconds to ensure that short-term fluctuations caused by gusts or sensor noise can be smoothed out. The implementer can set the duration of the preset time period according to the actual situation, which is not limited here, but the end time of the preset time period must be the current time.

[0067] Considering that in extremely low wind speeds or calm conditions, a simple physical transmission model (height / wind speed) can lead to an infinitely large calculated time delay, causing calculation anomalies, this embodiment sets a preset wind speed threshold of 0.5 m / s to ensure that dynamic calculation logic is only activated under effective wind speeds. Implementers can set the preset wind speed threshold based on the wind turbine's cut-in wind speed or the characteristics of the site environment; no limitation is imposed here. When the wind speed reference value is greater than the preset wind speed threshold, it indicates that the current wind transmission state is effective. Considering that wind conduction in actual forest environments is not an ideal vertical uniform motion, and is affected by turbulence, vegetation resistance, and terrain, the actual physical lag time often deviates from the theoretical calculation. To accurately simulate this complex physical process, this embodiment uses engineering experience based on a large amount of field experimental data to fit a preset relationship model to calculate the acquisition time difference. The specific formula is: In the formula, T is the time difference for data acquisition; H is the installation height of the surveillance camera; and C is the wind speed reference value. 'b' is a preset scaling factor used to correct for transmission rate changes caused by vertical wind speed attenuation; 'b' is a preset base offset time used to compensate for inherent data transmission and computation delays in the system; this embodiment sets... The time difference is set to 1.2 seconds for b and 0.5 seconds for c, ensuring that the data acquisition time difference more accurately reflects the actual delay in the transmission of airflow from the hub height to the ground vegetation layer. Implementers can set the time difference according to the calibration data of the specific wind field. b is not limited here. When the wind speed reference value is less than or equal to the preset wind speed threshold, it indicates that the current state is calm or light, and the physical transmission model is no longer applicable. In order to prevent calculation overflow and maintain the continuous operation of the system, this embodiment uses a preset fixed duration as the acquisition time difference; wherein, in this embodiment, the preset fixed duration is set to 2 seconds, and the implementer can set the size of the preset fixed duration according to actual engineering experience, which is not limited here;

[0068] Finally, starting from the current moment, the state data collected at historical moments of the retrospective acquisition time difference is used as the reference state data for aligning the current video frame. It should be noted that the airflow is conducted from the hub or wind turbine to the ground vegetation. Although the monitoring camera is installed on the nacelle, its height is usually very close to the hub height. Therefore, this embodiment assumes that the installation height of the monitoring camera is equal to the height of the wind turbine hub, and that the airflow is conducted vertically to the ground.

[0069] Preferably, in one feasible embodiment of this method, the method for obtaining the image wind direction reference system is as follows: To unify the geometric reference of physical data and visual data, this embodiment first sets up an image plane coordinate system (i.e., a two-dimensional coordinate system with the vertical upward direction of the image as 0° and clockwise rotation as the positive direction), and simultaneously sets up a nacelle physical coordinate system with the nacelle centerline as the 0° reference and clockwise rotation as the positive direction. To map the physical wind direction in the wind turbine coordinate system onto the image plane of the camera's viewpoint, this embodiment ensures that the relative wind direction (relative to the nacelle centerline) included in the reference state data and the current pan-tilt azimuth angle of the monitoring camera (relative to the nacelle centerline) are both defined in the nacelle physical coordinate system, making them comparable. This embodiment performs circular normalization on the difference between the relative wind direction included in the reference state data and the current pan-tilt azimuth angle of the monitoring camera in the corresponding angles of the nacelle physical coordinate system (ensuring that the resulting angle falls within the nacelle physical coordinate system). or Within a preset period (e.g., through modulo operation), the initial image wind direction angle is used to accurately reflect the theoretical projection direction of the high-altitude wind flow in the current monitoring image under the assumption of ideal flat terrain.

[0070] To correct for ground wind direction deviation caused by terrain turbulence and to prevent the limitations of purely physical data, this embodiment further obtains the absolute value of the minimum circumferential angle difference between the angle corresponding to the main visual motion direction of the current video frame in the image plane coordinate system and the wind direction angle of the initial image. This value is used as the deviation analysis value to ensure that the deviation calculation is not affected by the periodic expression of angles (such as 0° and 360°), and accurately reflects the degree of consistency between the physically predicted wind direction and the actual visually observed wind direction in the current scene. The method for obtaining the main visual motion direction is as follows: using optical flow or block matching algorithms, perform full-image motion estimation on the current video frame and its previous adjacent video frame to obtain the basic motion vector field; construct a direction histogram containing at least two angle intervals based on the basic motion vector field. In this embodiment, 360° is divided into 36 angle intervals, that is, one interval every 10°. The implementer can set the number and size of the angle intervals according to the actual situation, which is not limited here; perform statistics on the direction histogram, considering that the peak interval represents the dominant group motion trend in the picture (such as a large area of ​​vegetation swaying in the wind), and then take the center angle of the angle interval with the highest frequency as the main visual motion direction of the current video frame. It should be noted that, considering the swaying of vegetation may cause high-frequency jumps in the visual main motion direction calculated in a single frame, thus causing oscillations in the reference frame, this embodiment performs temporal stability detection on the visual main motion direction to ensure the robustness of the reference frame. The system maintains a historical visual direction queue. The visual feature is considered valid and used for subsequent calibration only if the rate of change of the angle of the visual main motion direction for a consecutive preset number of frames is lower than a preset stability threshold; otherwise, the visual direction is considered unstable, and the wind direction angle of the initial physical image is directly used. This embodiment sets the preset number of frames to 10 and the preset stability threshold to 5° / second to ensure the robustness of the reference frame. Implementers can set the preset number of frames and the preset stability threshold according to the actual situation, which is not limited here. The method of obtaining the basic motion vector field through optical flow or block matching algorithm is a well-known technique and will not be described in detail here.

[0071] The smaller the deviation analysis value, the closer the visually observed ground wind direction matches the upper-level wind direction predicted by the physical model. In this case, the visual features can effectively reflect the details of the actual wind conditions. The larger the deviation analysis value, the more significant the conflict between the two, which may be due to visual calculation errors (such as lighting interference) or sudden changes in wind direction caused by terrain. Therefore, this embodiment sets a preset deviation threshold of 30° to ensure that visual correction is only introduced when there is a basic consistency between physical and visual data, thus avoiding data contamination by errors. Implementers can set the size of the preset deviation threshold according to the complexity of the forest terrain and the anti-interference capability of the camera algorithm; no limitation is imposed here.

[0072] When the deviation analysis value is less than the preset deviation threshold, it indicates that the physical wind direction and the visual wind direction are roughly consistent, and the reliability of the visual features is high. In order to integrate the stability of physical data and the local adaptability of visual data, this embodiment uses the direction angle of the vector obtained by weighted summation of the unit vector corresponding to the main visual motion direction of the current video frame and the unit vector corresponding to the wind direction angle of the initial image as the calibrated target image wind direction angle. Specifically, firstly, the unit vector corresponding to the wind direction angle of the initial image is constructed. Unit vector corresponding to the main visual motion direction Then calculate the composite vector. ; final calculation The direction angle is used as the wind direction angle of the target image; where... The initial image wind direction angle; θ is the angle corresponding to the main visual motion direction of the current video frame in the image plane coordinate system; sin is the sine function; cos is the cosine function; To preset physical confidence weights; this embodiment uses vector weighting instead of direct angle weighting, effectively avoiding calculation errors caused by periodic angle jumps (such as the error of averaging 359° and 1° as 180°); this embodiment sets preset physical confidence weights. The physical confidence weight is 0.8, meaning it is greater than 0.5 and tends to be relatively large. This is because while wind turbine data may have local deviations, its overall trend is stable, whereas visual data is easily affected by local disturbances. Therefore, this embodiment adopts a strategy that prioritizes physical data and uses visual fine-tuning as a supplement. It only uses visual features to correct local constant deviations caused by terrain, preventing the reference frame from drifting due to tree swaying caused by over-reliance on visual data. This ensures that while introducing visual correction, the stability of the operating condition data is the primary factor. Implementers can set a preset physical confidence weight based on the reliability of the wind turbine data. The size is not limited here;

[0073] When the deviation analysis value is greater than or equal to the preset deviation threshold, it indicates that the visual features may be disturbed (such as changes in lighting or occlusion by moving targets) or the physical model may fail. In this case, a conservative strategy is better. In order to prevent incorrect visual guidance from causing the system to become inaccurate, the wind direction angle of the initial image is used as the wind direction angle of the target image.

[0074] To provide orthogonal geometric axes for subsequent motion decomposition, this embodiment constructs a unit vector based on the wind direction angle of the target image, that is, calculates the sine and cosine values ​​of this angle, specifically: downwind unit vector. Crosswind unit vector Where sin is the sine function and cos is the cosine function; The wind direction angle of the target image is used to determine the downwind direction (i.e., the unit vector along the direction of wind flow) and the crosswind direction (i.e., the unit vector perpendicular to the direction of wind flow) of the image plane coordinate system, which serve as the image wind direction reference system.

[0075] The abnormal motion intensity map acquisition module 30 is used to perform full-image motion estimation on the current video frame and fit a global affine transformation model, thereby eliminating global rigid displacement and obtaining the local relative motion vector of each pixel in the current video frame; the local relative motion vector is projected and decomposed to the downwind and crosswind directions of the image wind direction reference system; based on the direction flipping frequency of the downwind component of each pixel in the time domain, the unidirectional motion weight of each pixel is obtained; and then the downwind and crosswind components are weighted and suppressed to generate an abnormal motion intensity map.

[0076] Specifically, it is known that wind turbine nacelles experience irregular swaying under wind loads during operation, resulting in global rigid displacements in the monitoring image, including translation, rotation, and scaling. If this background motion is not removed, it will severely interfere with the detection of small targets (such as smoke and people). To separate the independent motion information of objects within the scene from the superimposed motion, ensuring that subsequent analysis is not affected by camera shake, this embodiment performs full-image motion estimation on the current video frame and fits a global affine transformation model. The affine transformation model can accurately characterize the background rigid transformation law caused by nacelle vibration. Then, the global rigid displacement is removed using the affine transformation model, obtaining the local relative motion vector of each pixel in the current video frame. This accurately reflects the independent motion state of objects relative to the ground, facilitating a focus on the dynamic changes of the scene itself. After removing the global displacement, to utilize physical priors to distinguish different types of motion, this embodiment projects and decomposes the local relative motion vector onto the downwind and crosswind directions of the image wind direction reference frame. This transforms the indistinguishable pixel motion into two orthogonal components with clear physical meaning, which is beneficial for targeted analysis of the unique vegetation pattern of downwind swaying.

[0077] Considering that tree branches and leaves typically oscillate back and forth around their equilibrium position at high frequency under wind power, while smoke drift or human movement exhibits a low-frequency unidirectional transport process, in order to completely decouple these two motion modes in the time domain, this embodiment obtains the unidirectional motion weight of each pixel based on the direction flip frequency of the downwind component in the time domain. This accurately reflects the credibility of the current pixel's motion mode belonging to unidirectional transport; the larger the unidirectional motion weight, the more likely the corresponding pixel is smoke drift or human movement; the smaller the unidirectional motion weight, the more likely the corresponding pixel is vegetation.

[0078] Finally, in order to generate a feature map with a high signal-to-noise ratio, enabling the system to filter out background noise to the greatest extent while preserving the real target signal, this embodiment uses unidirectional motion weights to suppress the downwind and crosswind components. Essentially, this significantly attenuates the motion energy of areas identified as vegetation interference (i.e., areas with high flip-over frequency characteristics), while fully preserving or enhancing the motion energy of areas identified as unidirectional targets (i.e., areas with low flip-over frequency characteristics). The resulting abnormal motion intensity map accurately reflects the real abnormal dynamics in the scene after removing vegetation interference, significantly improving the contrast between the target signal and background interference, which is beneficial for building a robust alarm threshold decision logic in the future.

[0079] Preferably, in one feasible embodiment of this method, the local relative motion vector is obtained as follows: In order to accurately extract the rigid motion model of the background from the complex full-image motion, this embodiment is based on the random sampling consensus algorithm, using sample points in the basic motion vector field to fit a six-parameter affine transformation model. The specific process is as follows: First, several groups (e.g., 50 groups) of macroblock sample points are randomly selected from the basic motion vector field; for each group of sample points, a set of affine parameters (including translation, rotation, and scaling parameters) is calculated using the least squares method. This embodiment uses the following formula to describe the affine transformation relationship: ;in, It represents the x-coordinate of the k-th pixel in the previous adjacent video frame of the current video frame; It is the ordinate of the k-th pixel in the previous adjacent video frame of the current video frame; It is a combination of the horizontal scaling factor and the rotational cosine component; It is a combination of horizontal shear and rotational sinusoidal components; It is a combination of vertical shear and rotational sinusoidal components; It is a combination of the scaling factor in the vertical direction and the rotation cosine component; This represents the x-coordinate of the k-th pixel in the current video frame; The ordinate of the k-th pixel in the current video frame; These are the horizontal translation parameters; The vertical translation parameters are used as follows: Further, the Euclidean distances between all macroblock vectors in the entire image and the predicted vectors of the model are calculated, and the number of inliers whose distances are less than a preset threshold (e.g., 0.5 pixels, which can be set by the implementer according to the actual situation and is not limited here) is counted. Finally, the set of parameters with the largest number of inliers is selected as the optimal global motion model for the current frame, thus obtaining the optimal parameter combination.

[0080] The affine transformation model represents global rigid displacement, accurately describing the overall geometric transformation of the image caused by cabin vibration. This facilitates the mathematical separation of the camera's uncontrolled motion from the real motion of objects within the scene. The random sample consensus algorithm is a well-known technique and will not be elaborated upon further.

[0081] In order to obtain the theoretical motion of each pixel position in the whole image due to background shaking, this embodiment obtains the theoretical background displacement vector of each pixel in the current video frame based on the affine transformation model. That is, by using the fitted affine parameters, the theoretical displacement value of each pixel coordinate after rigid transformation is calculated, which accurately reflects the motion vector that each pixel should have under the assumption that the scene is stationary and only the camera shakes.

[0082] In order to eliminate background interference and retain only the independent dynamics within the scene, this embodiment uses the vector difference between the vector corresponding to each pixel in the basic motion vector field and its theoretical background displacement vector in the current video frame as the local relative motion vector of each pixel in the current video frame. This is equivalent to mathematically stabilizing the background, so that the remaining vector field only contains local non-rigid motion information such as tree swaying and target movement.

[0083] Preferably, in one feasible method of this embodiment, the method for obtaining unidirectional motion weight is as follows: Considering that the rotation of the pan-tilt unit will cause the image coordinate system to shift, so that the same pixel coordinate no longer corresponds to the same physical object, thereby causing the temporal historical data to become invalid, in response to the detection that the pan-tilt unit of the surveillance camera is in a stationary locked state, in order to establish a temporal motion model for each pixel, for any pixel in the current video frame, this embodiment stores the downwind component of the pixel in real time into the corresponding first-in-first-out preset storage queue; it should be noted that the downwind component of the pixel is the last to enter the preset storage queue to maintain the continuity of the time sequence; this embodiment sets the length of the preset storage queue to 50 frames to ensure that the statistical window has a sufficient temporal span to capture the complete vegetation swaying cycle. The implementer can set the length of the preset storage queue according to the typical swaying frequency of vegetation in the scene, which is not limited here;

[0084] When the frame number of the current video frame is greater than the length of the preset storage queue, i.e., the historical queue is full, it indicates that the pixel has accumulated enough historical data for stable frequency statistics. To accurately identify the reciprocating oscillation characteristics of the pixel in the downwind direction, this embodiment determines the direction flip of the downwind component in the preset storage queue according to specified conditions. The specified conditions are: for any downwind component in the storage queue, the absolute values ​​of the downwind component and its next adjacent downwind component are both greater than a preset dead zone threshold, indicating that the motion amplitude at that point is sufficiently significant, rather than a small random jitter caused by image noise; simultaneously, the product of the downwind component and its next adjacent downwind component is less than 0, indicating that the pixel's motion direction has reversed between two adjacent moments (i.e., one reciprocating oscillation). This embodiment sets the preset dead zone threshold to 0.5 pixels / frame to ensure effective filtering of background noise. Implementers can set the size of the preset dead zone threshold according to the camera sensor signal-to-noise ratio, which is not limited here.

[0085] To quantify the degree of reciprocating motion of a pixel, the ratio of the number of direction flips in a preset storage queue to the length of a preset storage queue is used as the direction flip frequency. This accurately reflects the density of the pixel's reciprocating swaying along the downwind axis per unit time. A higher direction flip frequency indicates that the pixel better matches the characteristics of wind-driven vegetation swaying; a lower direction flip frequency indicates that the pixel better matches the target characteristics of unidirectional motion. Furthermore, to convert the frequency feature into a confidence level for suppressing interference, this embodiment uses the result of mapping the direction flip frequency using a negative exponential decay function as the unidirectional motion weight. The specific formula is as follows: In the formula, G is the weight of unidirectional motion; R is the attenuation control coefficient; R is the direction reversal frequency; exp is an exponential function with the natural constant as the base; in this embodiment, it is set as follows: A value of 10 ensures a steep cutoff characteristic for mid-to-high frequency oscillating signals; implementers can set this value according to their tolerance for false alarm rates. The magnitude of the weight is not limited here; this embodiment uses negative exponential nonlinear mapping to ensure that when the frequency increases slightly, the weight rapidly decays to 0, thereby achieving strong suppression of vegetation disturbance; while when the frequency is close to 0, the weight remains high, preserving the true target signal.

[0086] When the frame number of the current video frame is less than or equal to the length of the preset storage queue, it indicates that the system is in a cold start phase, and historical data is insufficient to support reliable frequency statistics. To prevent potential targets from being missed during the initialization phase, this embodiment adopts a high-sensitivity strategy, using 1 as the weight for unidirectional motion (i.e., assuming all motion is a suspected target and not suppressing it). Furthermore, to ensure the effectiveness of frequency statistics, this embodiment only performs calculations when the image wind direction reference system is relatively stable. That is, when the rate of change of the target image wind direction angle for multiple consecutive frames (e.g., 10 frames) is lower than the preset stability threshold, the aforementioned direction flip detection and update are performed; otherwise, it is considered a reference system oscillation. To prevent misjudgments due to unstable reference, this embodiment also uses 1 as the weight for unidirectional motion, suspends the interference suppression function, and prioritizes the target detection rate. Additionally, when the pan-tilt unit of the surveillance camera is detected to be not in a stationary locked state, i.e., the pan-tilt unit is in patrol or manual control, to prevent false alarms caused by data pollution due to pixel misalignment, this embodiment uses 1 as the weight for unidirectional motion.

[0087] Preferably, in one feasible embodiment, the method for obtaining the abnormal motion intensity map is as follows: For any pixel in the current video frame, the absolute value of the downwind component of the pixel is taken as the first amplitude; the absolute value of the crosswind component of the pixel is taken as the second amplitude; in order to comprehensively consider the motion energy in the downwind direction (which may include vegetation interference) and the motion energy in the crosswind direction (which may include target drift characteristics), the weighted sum of the first and second amplitudes is taken as the base intensity value of the pixel, accurately reflecting the overall motion activity of the pixel in all directions; wherein, the formula for calculating the base intensity value is: In the formula, S is the base intensity value of the pixel. This represents the windward component of that pixel. This is the first value; This represents the crosswind component of the pixel. This is the second value; It is an absolute value function; As the first preset weight; This is the second preset weight. This embodiment sets... It is 1.2. The sensitivity is set to 0.8 to ensure high sensitivity to the main motion in the downwind direction, while retaining a certain crosswind response capability to capture irregular drift of the target. The implementer can set the sensitivity according to the required downwind motion detection. Set according to the requirements for crosswind motion detection sensitivity No specific restrictions are imposed here;

[0088] To preserve real unidirectional moving targets (such as smoke or people) while strongly suppressing oscillating vegetation interference using temporal features, the product of the unidirectional motion weight and the base intensity value of a pixel is used as the anomalous motion intensity of that pixel, accurately reflecting the true anomaly confidence after removing vegetation interference. To provide an intuitive global anomaly distribution view for subsequent connected component analysis and thresholding, the image composed of the anomalous motion intensities of each pixel in the current video frame is used as the anomalous motion intensity map.

[0089] The abnormal area identification module 40 is used to obtain an adaptive alarm threshold that fluctuates with the wind conditions based on the background noise distribution characteristics of the abnormal motion intensity map, and then identify abnormal areas.

[0090] Specifically, to address the issue that fixed thresholds cannot balance sensitivity and false alarm rate under different wind speeds, and to enable the system to automatically adjust detection standards based on real-time environmental background noise, this embodiment obtains an adaptive alarm threshold that fluctuates with wind conditions based on the background noise distribution characteristics of the abnormal motion intensity map. This is beneficial for increasing the threshold to suppress false alarms in strong wind and high noise environments, and for decreasing the threshold to prevent missed alarms in light wind and low noise environments. This ensures that the alarm logic is always in the optimal signal-to-noise ratio state, thereby accurately identifying abnormal areas that significantly deviate from the current background distribution. This is conducive to generating high-precision alarm intelligence and significantly improving the environmental adaptability of the security system.

[0091] Preferably, in one feasible embodiment, the adaptive alarm threshold is obtained as follows: First, the abnormal motion intensity map is smoothed in the temporal domain using an exponentially weighted moving average algorithm to obtain a risk accumulation map; specifically, for any pixel in the current video frame, the formula for calculating the risk accumulation analysis value of that pixel is: In the formula, F is the cumulative risk analysis value of that pixel. This is the cumulative risk analysis value of the pixel in the previous adjacent video frame of the current video frame. This represents the abnormal motion intensity of the pixel. This is the third preset weight. Specifically, if a pixel only exhibits a high value in a single frame (transient noise), its energy will be... Rapid decay; if the pixel maintains a high value (real target) for multiple consecutive frames, its cumulative value will steadily increase; this embodiment sets... A value of 0.1 ensures sufficient smoothing capability against transient noise while maintaining a moderate response speed to the appearance of real targets. Implementers can set this value according to their tolerance for alarm delay. The size of the image is not limited here. It should be noted that the risk accumulation analysis value of each pixel in the initial video frame is 0. Finally, the image composed of the risk accumulation analysis values ​​of each pixel in the current video frame is used as the risk accumulation map; the exponentially weighted moving average algorithm is well-known and will not be elaborated further.

[0092] It is known that when a large area of ​​real targets (such as a large-scale fire) exists in the image, directly calculating the average of the entire image will increase the background noise level, leading to a higher threshold and missed targets (i.e., the self-masking effect). To obtain a robust and accurate background noise level, this embodiment employs a robust statistical background estimation strategy. The mean and standard deviation of the remaining pixels in the risk accumulation map, excluding the highest-valued pixels in the first preset proportion, are used as the background mean and standard deviation, respectively. This embodiment sets the preset proportion to 20% to ensure that potentially high-confidence target areas are excluded, and noise statistics are performed only using the background area. The implementer can set the preset proportion based on the expected minimum target proportion; no limit is imposed here. This strategy ensures that even if significant abnormal targets appear in the image, the statistically obtained background parameters can still represent the true noise level of the environment, thereby maintaining stable detection sensitivity. To set an anomaly threshold based on statistical principles, the sum of the background mean and the weighted background standard deviation is used as a dynamic floating threshold to accurately reflect the statistical upper limit of background fluctuations under the current wind conditions. The formula for calculating the dynamic floating threshold is: In the formula, D is the dynamic floating threshold. The mean value is the background value. The background standard deviation; This is the fourth preset weight. This embodiment sets... The value is set to 3.0 to ensure that only areas significantly exceeding the background fluctuation range (e.g., 3 standard deviations) are considered abnormal. Implementers can set this value according to their needs for controlling the false alarm rate (e.g., the 3σ principle). The size of σ is not limited here; the 3σ principle is well known and will not be elaborated further.

[0093] Considering that in windless and extremely static scenes, the calculated dynamic threshold may be extremely small, even lower than the inherent dark current noise level of the sensor, in order to prevent false alarms caused by electronic noise, this embodiment sets a preset minimum noise floor threshold of 5.0 to ensure that the alarm threshold is always higher than the camera's inherent noise floor. The implementer can set the size of the preset minimum noise floor threshold according to the camera's signal-to-noise ratio parameter, which is not limited here. The dynamic floating threshold is compared with the preset minimum noise floor threshold, and the larger value is used as the adaptive alarm threshold that fluctuates with the wind conditions to ensure that the system still has basic noise resistance in extremely static environments.

[0094] Preferably, in one feasible embodiment, the method for obtaining the abnormal region is as follows: In order to locate significantly abnormal target regions from the continuous risk map, this embodiment uses an adaptive alarm threshold to binarize the risk accumulation map, so that background noise below the adaptive alarm threshold is set to 0, and suspected targets above the adaptive alarm threshold are set to 1. Then, through a connected component analysis algorithm (such as 8-neighborhood search), connected regions are extracted as suspected target regions to determine the target position and outline on the current image plane, which is beneficial for subsequent geographic positioning. In order to convert the two-dimensional image coordinates into three-dimensional geographic locations with practical value to assist emergency response, the installation latitude and longitude, height, current focal length, pan-tilt azimuth angle, and pan-tilt pitch angle of the surveillance camera are obtained. Based on a preset ground digital elevation model, i.e., the three-dimensional elevation data of the forest terrain, the pixel coordinates of the suspected target region are mapped to geographic coordinates using the principle of monocular perspective projection to obtain the abnormal region. Specifically, a line of sight ray is constructed using the camera's internal and external parameters and the pan-tilt angle, and the spatial intersection point of the ray with the surface of the preset ground digital elevation model is calculated. The latitude and longitude of the intersection point is the geographic coordinate of the target. Among them, binarization processing, connected component analysis algorithm, monocular perspective projection principle and ray intersection algorithm are well known and will not be described in detail.

[0095] In summary, this embodiment first acquires monitoring video frames and wind turbine operating status data including relative wind direction, and performs time delay compensation and visual main motion direction calibration based on the acquisition time difference to construct an image wind direction reference system that integrates operating conditions and visual features. Secondly, it fits a global affine transformation model to eliminate global rigid displacement caused by nacelle vibration, obtaining local relative motion vectors. Subsequently, the vectors are decomposed into downwind and crosswind directions of the reference system, and unidirectional motion weights are calculated based on the time-domain flip frequency of the downwind component. Weighted suppression is applied to the two directional components to filter out vegetation interference, generating an abnormal motion intensity map. Finally, an adaptive alarm threshold is generated based on the background noise distribution of the intensity map to identify abnormal areas. This invention effectively solves the problems of false alarms and missed alarms under strong wind and nacelle vibration environments, improves the reliability of security in wind farm forest areas, and thus achieves all-weather, high-precision unattended monitoring.

[0096] Example 2:

[0097] This invention also proposes a large-scale video monitoring device for wind turbine forest areas based on image algorithms. The device includes a memory and a processor. The memory stores executable program code, and the processor is used to call and execute the executable program code to execute the large-scale video monitoring system for wind turbine forest areas based on image algorithms provided in this application. Specifically, the device may be a chip, component, or module. The chip may include a connected processor and memory; the memory stores instructions, and when the processor calls and executes the instructions, the chip can execute the large-scale video monitoring system for wind turbine forest areas based on image algorithms provided in the above embodiments.

[0098] Furthermore, this application also protects a computer device; please refer to [link to relevant documentation]. Figure 2 The computer device includes a memory 401, a processor 402, and a computer program 403 stored in the memory 401 and running on the processor 402. When the processor 402 executes the computer program 403, the computer device can execute any of the aforementioned image algorithm-based large-scale video monitoring systems for wind turbine forest areas.

[0099] Example 3:

[0100] The present invention also provides a computer-readable storage medium storing computer program code, which, when executed on a computer, causes the computer to perform the aforementioned method steps to implement the image algorithm-based large-scale video monitoring system for wind turbine forest areas provided in the above embodiments.

[0101] Example 4:

[0102] The present invention also provides a computer program product that, when run on a computer, causes the computer to perform the aforementioned related steps to realize the large-scale video monitoring system for wind turbine forest areas based on image algorithms provided in the above embodiments.

[0103] In this embodiment, the device, computer-readable storage medium, computer program product, or chip are all used to execute the corresponding methods provided above. Therefore, the beneficial effects they can achieve can be referred to the beneficial effects in the corresponding methods provided above, and will not be repeated here.

[0104] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0105] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

Claims

1. A large-scale video monitoring system for wind turbine forest areas based on image algorithms, characterized in that, The system includes: The data acquisition module is used to acquire video frames from the wind turbine forest area monitoring camera and the wind turbine operating status data in real time; among which, the status data includes the relative wind direction; The image wind direction reference system acquisition module is used to perform time lag compensation based on the acquisition time difference between the state data and the video frame, and determine the reference state data aligned with the current video frame; according to the visual main motion direction of the current video frame, the relative wind direction contained in the reference state data is calibrated to construct an image wind direction reference system containing the downwind direction and the crosswind direction. The abnormal motion intensity map acquisition module is used to perform full-image motion estimation on the current video frame and fit a global affine transformation model, thereby eliminating global rigid displacements and obtaining the local relative motion vector of each pixel in the current video frame; the local relative motion vector is projected and decomposed to the downwind and crosswind directions of the image wind direction reference frame; based on the direction flipping frequency of the downwind component of each pixel in the time domain, the unidirectional motion weight of each pixel is obtained; and then the downwind and crosswind components are weighted and suppressed to generate an abnormal motion intensity map; The abnormal area identification module is used to obtain an adaptive alarm threshold that fluctuates with wind conditions based on the background noise distribution characteristics of the abnormal motion intensity map, and then identify abnormal areas.

2. The large-scale video surveillance system for wind turbine forest areas based on image algorithms as described in claim 1, characterized in that, The method for obtaining the image wind direction reference system is as follows: Set up an image plane coordinate system that includes a positive direction definition and a cabin physical coordinate system with the cabin centerline as the reference. The difference between the relative wind direction contained in the reference state data and the corresponding angle of the current pan-tilt azimuth angle of the monitoring camera in the cabin physical coordinate system is circumferentially normalized and used as the current initial image wind direction angle. Temporal stability detection is performed on the visual main motion direction of the current video frame. If and only if the rate of change of the angle of the visual main motion direction for a consecutive preset number of frames is lower than a preset stability threshold, the absolute value of the minimum circumferential angle difference between the corresponding angle of the visual main motion direction of the current video frame in the image plane coordinate system and the wind direction angle of the initial image is used as the deviation analysis value. When the deviation analysis value is less than the preset deviation threshold, the direction angle of the vector obtained by weighted summation of the unit vector corresponding to the visual main motion direction of the current video frame and the unit vector corresponding to the wind direction angle of the initial image is used as the calibrated target image wind direction angle; wherein, the weight of the initial image wind direction angle is set to be greater than the weight of the visual main motion direction. When the deviation analysis value is greater than or equal to the preset deviation threshold, or when the visual main motion direction fails the temporal stability detection, the initial image wind direction angle is taken as the target image wind direction angle. Based on the wind direction angle of the target image, a unit vector is constructed to determine the downwind and crosswind directions of the image plane coordinate system, which serve as the image wind direction reference system.

3. The large-scale video monitoring system for wind turbine forest areas based on image algorithms as described in claim 2, characterized in that, The method for obtaining the main visual motion direction is as follows: Perform full-map motion estimation on the current video frame and its previous adjacent video frame to obtain the basic motion vector field; construct an orientation histogram containing at least two angular intervals based on the basic motion vector field. Statistical analysis of the direction histogram is performed, and the center angle of the angle interval with the highest frequency is taken as the main visual motion direction of the current video frame.

4. The large-scale video monitoring system for wind turbine forest areas based on image algorithms as described in claim 3, characterized in that, The method for obtaining the local relative motion vector is as follows: Based on the random sampling consensus algorithm, a six-parameter affine transformation model is fitted using sample points in the basic motion vector field; the affine transformation model represents the global rigid displacement. Based on the affine transformation model, the theoretical background displacement vector of each pixel in the current video frame is obtained; The vector difference between the vector corresponding to each pixel in the basic motion vector field and its theoretical background displacement vector in the current video frame is used as the local relative motion vector of each pixel in the current video frame.

5. The large-scale video surveillance system for wind turbine forest areas based on image algorithms as described in claim 1, characterized in that, The method for obtaining the unidirectional motion weight is as follows: In response to the detection that the pan-tilt unit of the surveillance camera is in a stationary and locked state, for any pixel in the current video frame, the downwind component of that pixel is stored in the corresponding first-in-first-out preset storage queue in real time. When the frame number of the current video frame is greater than the length of the preset storage queue, the direction of the downwind component in the preset storage queue is flipped according to the specified conditions. The ratio of the number of direction flips in the preset storage queue to the length of the preset storage queue is used as the direction flip frequency; The result of mapping the direction reversal frequency using the negative exponential decay function is used as the weight for unidirectional motion. When the frame number of the current video frame is less than or equal to the length of the preset storage queue, 1 is used as the unidirectional motion weight.

6. The large-scale video surveillance system for wind turbine forest areas based on image algorithms as described in claim 5, characterized in that, The specified conditions are: For any downwind component in the storage queue, the absolute value of the downwind component and its next adjacent downwind component are both greater than the preset dead zone threshold. The product of the downwind component and its next adjacent downwind component is less than 0.

7. The large-scale video surveillance system for wind turbine forest areas based on image algorithms as described in claim 1, characterized in that, The method for obtaining the abnormal motion intensity map is as follows: For any pixel in the current video frame, the absolute value of the downwind component of that pixel is taken as the first value; the absolute value of the crosswind component of that pixel is taken as the second value. The weighted sum of the first and second values ​​is used as the base intensity value of the pixel. The product of the unidirectional motion weight of the pixel and the base intensity value is taken as the abnormal motion intensity of the pixel. The image formed by the abnormal motion intensity of each pixel in the current video frame is used as the abnormal motion intensity map.

8. The large-scale video surveillance system for wind turbine forest areas based on image algorithms as described in claim 1, characterized in that, The method for obtaining the adaptive alarm threshold is as follows: The exponentially weighted moving average algorithm is used to smooth the abnormal motion intensity map in the time domain to obtain the risk accumulation map. Obtain the mean and standard deviation of the remaining pixels in the risk accumulation map, excluding the pixel with the highest value in the previous preset proportion, and use them as the background mean and background standard deviation, respectively. The sum of the background mean and the weighted background standard deviation is used as the dynamic floating threshold. The dynamic floating threshold is compared with the preset minimum noise floor threshold, and the larger value is used as the adaptive alarm threshold that fluctuates with the wind conditions.

9. The large-scale video monitoring system for wind turbine forest areas based on image algorithms as described in claim 8, characterized in that, The method for obtaining the abnormal region is as follows: The risk accumulation map is binarized using an adaptive alarm threshold, and connected regions are extracted as suspected target regions. The system obtains the installation latitude and longitude, altitude, current focal length, azimuth angle, and tilt angle of the surveillance camera. Based on a preset digital elevation model, it uses the principle of monocular perspective projection to map the pixel coordinates of the suspected target area to geographic coordinates, thereby obtaining the abnormal area.

10. The large-scale video monitoring system for wind turbine forest areas based on image algorithms as described in claim 1, characterized in that, The method for obtaining the reference state data is as follows: The system acquires wind speed data in real time from the status data and uses the average of all wind speed data within a preset time period as the wind speed reference value. When the wind speed reference value is greater than the preset wind speed threshold, the acquisition time difference is obtained based on the installation height of the monitoring camera and the wind speed reference value, combined with the preset relationship model. When the wind speed reference value is less than or equal to the preset wind speed threshold, the preset fixed duration will be used as the data collection time difference. Starting from the current moment, the state data collected at historical moments of the time difference are used as the reference state data for aligning the current video frame.

Images