Power plant security target detection method and system based on scene physical constraint

Abstract

The application provides a power plant security target detection method and system based on scene physical constraints to overcome the high false detection rate and the easy missed detection of real small targets caused by strong reflection and backlight interference in the existing power plant security target detection, and relates to the technical fields of computer vision and intelligent security of power plants. The method comprises the following steps: establishing a scene constraint baseline reflecting the distribution law of reflection interference and the change law of illumination in the scene; determining the reflection interference area and the candidate target protection area in the current frame image; controlling the exposure parameters of the camera to obtain an input image for target detection; extracting target detection features, and limiting the intensity of the suppression by using the candidate target protection area to obtain corrected target detection features. The application solves the technical problems of high false detection rate and easy missed detection of real small targets caused by strong reflection and backlight interference.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and intelligent security technology for power plants, specifically a method and system for detecting security targets in power plants based on scene physical constraints. Background Technology

[0002] Numerous fixed video surveillance points are installed in thermal power plants, wind farms, and substation areas to automatically detect security targets such as those not wearing safety helmets, not wearing work clothes, not using safety belts properly, personnel falling to the ground, area intrusion, and open flames and smoke. Compared with general park monitoring scenarios, power plant scenarios have the following unique characteristics: the images consistently contain highly reflective objects such as stainless steel pipes, galvanized components, aluminum casings, smooth equipment surfaces, and water surfaces; outdoor cameras are typically installed at fixed heights and tilt angles, and the direction of sunlight changes regularly over time; indoor fixed light sources, welding arc light, and maintenance lighting can create localized bright areas. These factors easily create bright spots, bright bands, or false outlines near small targets such as safety helmets, reflective work clothes, and safety belts, seriously interfering with target detection.

[0003] Most existing methods process images based solely on brightness or color characteristics, failing to utilize the crucial factor that power plant cameras are permanently fixed and that their materials and surface morphology are relatively stable. Consequently, it is difficult to distinguish between false targets formed by specular reflection and genuine safety helmets, reflective strips, or safety belts based on their causes. Uniformly suppressing high-brightness areas across the entire image may weaken the edge and color features of genuine targets, leading to a higher false negative rate. Conversely, continuously performing multi-exposure fusion or complex model inference on all images significantly increases the processing burden on edge computing devices, making it difficult to strike a balance between real-time performance and accuracy.

[0004] To address the aforementioned issues, a technical solution based on the physical constraints of a fixed power plant scenario is needed. This solution would allow reflection interference modeling, exposure control, and model feature suppression to work together to reduce false alarms caused by metal surfaces, water surfaces, and backlighting while preserving the true characteristics of small targets, and also take into account the real-time processing capabilities of edge devices. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a power plant security target detection method and system based on scene physical constraints, so as to overcome the problems of high false detection rate and easy omission of real small targets due to strong reflection and backlight interference from metal surfaces, water surfaces and other strong reflections in existing power plant security target detection.

[0006] The technical solution adopted by this invention to solve its technical problem is: a power plant security target detection method based on scene physical constraints, comprising the following steps: S1. Obtain scene attribute information of fixed monitoring points in the power plant, and establish a scene constraint baseline based on the scene attribute information that reflects the distribution pattern of reflection interference and the pattern of illumination change in the scene; wherein, the scene attribute information includes at least camera pose parameters, surface material reflectivity of objects in the monitored scene, and surface morphology information. S2. Match the current frame image with the scene constraint baseline, and determine the reflection interference area and candidate target protection area in the current frame image based on the matching result; the candidate target protection area is the image area in the current frame image that belongs to the actual security target and is determined from the current frame image based on the evidence of the existence of the target in the current frame image. S3. Based on the reflection interference area and the candidate target protection area, control the camera's exposure parameters to obtain the input image for target detection; S4. Input the target detection input image into the target detection network to extract target detection features; use the reflection interference region to suppress the feature response of the target detection features, and use the candidate target protection region to limit the intensity of the suppression to obtain the corrected target detection features; S5. Based on the corrected target detection features, classify and locate the security target, and output the target detection results.

[0007] Furthermore, the scene constraint baseline includes: the consistency of reflection direction calculated based on the camera's line of sight, the device's surface normal or curvature, the fixed light source direction, or the sunlight direction that changes over time; and the brightness distribution, highlight frequency, chromaticity stability, gradient direction stability, and common highlight periods of each image grid obtained statistically from the reference frame within at least one observation period. The camera pose parameters include camera intrinsic parameters, installation position, installation height, pitch angle, and orientation.

[0008] Furthermore, the method for determining the reflection interference area includes: Extract feature information from the current frame image relative to the scene constraint baseline, the feature information including at least one of highlight level, low chromaticity level, brightness residual and bright spot gradient direction; Based on the feature information, the reflection confidence of each image location is obtained by weighting the calculation according to the material weight coefficient set corresponding to the surface material reflectivity of the corresponding location. The reflection confidence score is compared with a preset threshold for reflection confidence score, and image locations that reach or exceed the preset threshold for reflection confidence score are segmented by thresholding to form initial reflection regions; Based on the consistency of reflection direction in the scene constraint baseline, the specular reflection direction is calculated as the predicted reflection direction by combining the surface normal and the light source direction, and the initial reflection area is directionally expanded along the predicted reflection direction to obtain the reflection interference area.

[0009] Furthermore, the method for determining the candidate target protection area includes: Obtain evidence of target presence in the current frame image. The evidence of target presence includes at least one of the following: target motion consistency information obtained through inter-frame motion detection, human contour information quickly obtained on a low-resolution image by a lightweight human detector or head and shoulder detector independent of the target detection network, and infrared contour information obtained through the infrared imaging channel. Compare the evidence of the existence of the target with a set protection threshold; When the evidence of the target reaches or exceeds the set protection threshold, the corresponding image region is determined as the candidate target protection region.

[0010] Further, step S3 includes: S31. Obtain the determined candidate target protection area; S32. Calculate the overexposure ratio of the high brightness of the reflection interference area and the underexposure of the candidate target protection area within the working area. S33. When neither the overexposure ratio nor the underexposure level reaches the trigger condition, single-frame processing is adopted, and the static reflection area of ​​the corresponding time period is read from the scene constraint baseline as the calculation prior for subsequent feature correction; the static reflection area is the area recorded in the scene constraint baseline that has fixed reflection characteristics in the corresponding time period; the trigger condition includes the overexposure ratio threshold and the underexposure level threshold. When at least one of the overexposure ratio and the underexposure degree reaches the trigger condition, multi-exposure acquisition and fusion processing is performed: The exposure order and exposure amount of short exposure frames, medium exposure frames, and long exposure frames are determined based on the scene constraint baseline. The exposure amount of the short exposure frame is determined based on the upper quantile of the brightness of the reflection interference region, so that the upper quantile of the brightness of the reflection interference region is lower than the overexposure threshold. The exposure amount of the long exposure frame is determined based on the lower quantile of the brightness of the candidate target protection region, so that the lower quantile of the brightness of the candidate target protection region reaches the range of extractable contours, and the long exposure time is limited according to the target movement speed. Static background registration is performed using the aforementioned exposure frame as a geometric reference, and local registration of the moving region is performed using target optical flow; For each exposure frame, a basic fusion weight is generated, consisting of exposure suitability, local contrast, and registration reliability. Using the candidate target protection region as a spatial prior, the fusion weight is adjusted as follows: (1) For the image region within the candidate target protection area, increase the fusion weight of long exposure frames or medium exposure frames; wherein, when there is a moving target in the region, the fusion weight of medium exposure frames is increased first, and when the region is a static background, the fusion weight of long exposure frames is increased first. (2) For image regions within the reflection interference area that do not belong to the candidate target protection area, increase the fusion weight of short exposure frames and decrease the fusion weight of long exposure frames; (3) For the underexposed areas of the candidate target protection area, increase the fusion weight of medium exposure frames or long exposure frames; (4) For the motion region, reduce the fusion weight of the exposure frame with insufficient registration reliability; when the candidate target protection region overlaps with the reflection interference region, the weight adjustment rule of the candidate target protection region shall prevail in the overlapping region, that is, mode (1) shall be executed first, and the operation of increasing the weight of short exposure frame and decreasing the weight of long exposure frame in mode (2) shall not be executed or the operation of increasing the weight of short exposure frame and decreasing the weight of long exposure frame shall be restricted. After normalizing the weights of each exposure frame, they are fused to generate a fused image, which is used as the input image for target detection.

[0011] Further, step S4 includes: S41. Input the target detection input image into the target detection network, and extract multi-scale target features through the target feature branch of the target detection network. l = 1, 2, 3, ..., L, where L is the number of scales; S42. Generate a material category code based on the surface material reflectivity in the scene attribute information, and generate a brightness residual based on the brightness deviation of the target detection input image relative to the scene constraint baseline; S43. Input the reflection interference area, the material category code, and the brightness residual into the interference feature branch of the target detection network to extract interference features; S44. On the multi-scale feature layer, the reflection interference region is scaled to the corresponding scale, and the reflection suppression coefficient at each scale is determined according to the material category code. ; S45, Regarding the first Target features at each scale The following formula is used to obtain the first... Correction features at each scale : ; in, This represents the reflection interference area at the corresponding scale. This represents the protected area for candidate targets at the corresponding scale. The compensation features are extracted from non-highlight textures and brightness residuals by the interference feature branch. The compensation coefficient; S46. Based on the correction characteristics of each scale The corrected target detection features F are obtained, and the classification and localization of security targets are performed.

[0012] Furthermore, it also includes deployment and adaptation steps: Spatially converge the scene constraint baselines to obtain a point scene vector that characterizes the overall scene characteristics of the current monitoring point. Point adaptation parameters are generated or selected based on the point scene vector; the point adaptation parameters are used to adjust at least one of the reflection suppression coefficient, feature channel scaling coefficient, feature channel bias, and detection threshold of the target detection network; Using the target detection network as the basic model, the basic model, the point adaptation parameters, the scene constraint baseline, and the preset task rules are bound together to form a deployable unit associated with the current camera channel; the preset task rules include task types and event rules; the task type corresponds to a spatial position constraint relationship, and the spatial position constraint relationship is contained in the event rules; When deploying at new locations, at least some parameters of the base model remain unchanged, and only the location adaptation parameters and the scene constraint baseline are updated.

[0013] Furthermore, it also includes an event determination step: The personnel tracking results are obtained by associating targets in consecutive frames using a multi-target tracking algorithm. Spatial relationship verification is performed according to the task type and its corresponding spatial constraints in the preset task rules. The target detection results are associated with the personnel tracking results, and the target detection results are verified according to the spatial constraints corresponding to the task type. The spatial relationship verification includes: the safety helmet target should be located in the personnel's head and shoulder area; the work clothes or reflective strip target should be located in the personnel's torso area; and the safety belt and its attachment points should have a preset spatial relationship with the high-altitude work area and the personnel's torso area. Calculate the overlap ratio between the detection bounding box of the target detection result that passed the spatial relationship verification and the reflection interference area; When the detection frame highly overlaps with the reflection interference area, maintains a fixed spatial position at multiple times, and lacks personnel movement or infrared contour support, the event confidence of the detection result is reduced. When the detection frame overlaps with the reflection interference area, but has continuous human body trajectory or infrared contour support, the detection result is retained using the candidate target protection area; The retained detection results are checked for temporal continuity, and security events are generated from the detection results that pass the check.

[0014] Furthermore, it also includes a feedback update step: Acquire false alarm and missed alarm information that has been manually or multi-sensor verified, map the false alarm and missed alarm back to the image coordinates of the corresponding camera, and generate an error spatial distribution map; When an error repeatedly occurs at a fixed location, the overlap ratio with the reflection interference area reaches a set condition, and the error changes regularly with the sunshine period or the fixed light source status, the error is classified as a scene reflection error, and the frequency of bright appearance, material weight coefficient set, reflection confidence preset threshold, or common bright period in the corresponding position in the scene constraint baseline is updated. When an error occurs as a person or equipment moves and does not meet the fixed reflection characteristics, the error is classified as a target recognition error, and the corresponding sample is added to the hard sample set to update the location adaptation parameters or to update the basic model after centralized verification. The basic model is managed by version number. After an update, new locations will use the new version, while existing locations will retain the original version until they are actively upgraded. The updated version will be tested at some locations first. If the accuracy, recall, false alarm rate, or processing delay does not meet the set conditions, the model will be restored to the scene constraint baseline or model version before the update.

[0015] A power plant security target detection system based on scene physical constraints, the system being used to execute the aforementioned power plant security target detection method based on scene physical constraints, comprising: The scene constraint modeling module is used to perform step S1; A reflection interference calculation module is used to perform step S2; The acquisition and control module is used to execute step S3; The feature correction and target detection module is used to perform steps S4 and S5.

[0016] The beneficial effects of this invention are as follows: Through the coordinated operation of steps S1 to S5, this invention systematically improves the problem of "lack of scene physical prior guidance" in the existing technology. S1 quantifies the camera pose, the reflective properties and morphological information of the equipment surface material, and the law of illumination change into a scene constraint baseline, enabling the system to grasp the spatiotemporal distribution law of reflection interference from the causal level. S2 uses this baseline to separate fixed reflection pseudo-targets and real moving targets into reflection interference regions and candidate target protection regions. S3 uses these two regions as spatial priors to control exposure parameters as needed, avoiding the computational burden caused by continuously running complex models. S4 uses the reflection interference regions to perform structured suppression of target detection features, and uses the candidate target protection regions to limit the suppression intensity, so that the corrected features retain real target information and suppress reflection interference information. S5 performs classification and localization based on the corrected features, thereby systematically solving the technical problems of high false detection rate and easy missed detection of real small targets caused by strong reflection and backlight interference in the entire link from acquisition to detection. Attached Figure Description

[0017] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0018] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0019] like Figure 1 As shown, the power plant security target detection method based on scene physical constraints of the present invention includes the following steps: S1. Scene constraint baseline establishment: Obtain scene attribute information of fixed monitoring points in the power plant, and establish a scene constraint baseline based on the scene attribute information that reflects the distribution law of reflection interference and the law of light change in the scene; wherein, the scene attribute information includes at least camera pose parameters, surface material reflectivity properties of objects in the monitored scene, and surface morphology information. S2. Determination of Reflection Interference Region and Candidate Target Protection Region: The current frame image is matched with the scene constraint baseline, and the reflection interference region and candidate target protection region in the current frame image are determined based on the matching result; the candidate target protection region is an image region in the current frame image that belongs to the actual security target, determined from the evidence of the existence of the target (various security objects that need to be detected and protected) in the current frame image; the evidence of the existence of the target is used to indicate the possibility that a real security target exists in the image region; S3. Exposure parameter control: Based on the reflection interference area and the candidate target protection area, control the camera's exposure parameters to obtain the input image for target detection; S4. Feature Correction: Input the target detection input image into the target detection network to extract target detection features; use the reflection interference region to suppress the feature response of the target detection features, and use the candidate target protection region to limit the intensity of the suppression, to obtain the corrected target detection features; S5. Target Detection: Based on the corrected target detection features, perform classification and localization of security targets, and output the target detection results.

[0020] In this application, surface material reflectivity refers to the category of optical reflection characteristics of a device or object surface, used to characterize the intensity and directional distribution characteristics of specular reflection of incident light by different material surfaces. Surface material reflectivity can be specifically categorized into smooth metallic surfaces, metallic coated surfaces, water surfaces, and diffuse reflective surfaces, etc. Different material categories correspond to different reflection characteristic weights and reflection suppression coefficients.

[0021] In this application, surface morphology information refers to the geometric shape category of a device or object's surface, used to characterize the normal field distribution characteristics of different surface shapes and their variation with spatial position. Surface morphology information can be specifically categorized into planar, cylindrical, and irregular curved surfaces. Planar surfaces have a consistent normal and a fixed reflection area; cylindrical surfaces have a continuously changing normal with the cross-sectional angle, and the reflection area moves with the direction of illumination; irregular curved surfaces have complex normal variations and dispersed reflection areas, requiring point-by-point calculation. Surface morphology information, together with the reflective properties of the surface material, constitutes a complete physical description of reflection interference in the scene.

[0022] In this application, specifically, the scene constraint baseline includes: the consistency of reflection direction calculated based on the camera's line of sight, the device's surface normal or curvature, the fixed light source direction, or the sunlight direction that changes over time; and the brightness distribution, highlight frequency, chromaticity stability, gradient direction stability, and common highlight periods of each image grid obtained statistically from reference frames within at least one observation period; the camera pose parameters include camera intrinsic parameters, installation position, installation height, pitch angle, and orientation. The highlight frequency is: within at least one observation period, for each image grid, the proportion of reference frames whose brightness reaches or exceeds the highlight determination threshold to the total number of reference frames; the highlight determination threshold is determined based on the historical brightness upper quantile or overexposure threshold of that location; the total number of reference frames is the total number of valid reference frames excluding those containing personnel, vehicles, or mobile devices.

[0023] It should be noted that the target detection input image is fed into the target detection network, which is a pre-trained target detection network. In this invention, the target detection network adopts a dual-branch structure, including a target feature branch and an interference feature branch. The training data includes regular annotations and reflection region annotations. The regular annotations include the bounding boxes and category labels of security targets. The reflection region annotations are reflection interference region markers generated based on scene physical constraints, which are sampled and corrected and used as supervision signals to enable the interference feature branch to learn and recognize the spatial distribution of physical reflection interference. During training, candidate target protection region masks are generated using the annotated real target bounding boxes. (Set the region within the true target bounding box to 1, and the rest to 0), so that the feature correction formula in the training and inference phases... Maintaining a consistent distribution ensures that the network can learn during training. Regional desuppression, in The total loss function for training differential response patterns with normal suppression in the region is:

[0024] In the formula, The target detection loss, including classification loss and localization loss, is used to ensure the network's basic detection capability for security targets. The reflection region supervision loss is used to enable the interfering feature branch to learn the reflection region annotations. Consistency loss is used to constrain the output of the same real target in the original image and the reflection-controlled fused image to remain consistent, ensuring that the network does not lose its ability to recognize real targets when suppressing reflection interference; and These are the weighting coefficients for each loss term.

[0025] Through the aforementioned multi-task joint training, the network learns to distinguish between "physical reflection interference" and "real targets," rather than simply learning statistical correlation by increasing the number of strong light samples. The backbone network can be a convolutional neural network or a visual transformation network, and the optimizer, quantization, or pruning method can be selected according to the computing power of the edge device.

[0026] Example 1: Establishing the scenario constraint baseline.

[0027] This embodiment uses a fixed camera in the outdoor equipment area of ​​a power plant as an example. The camera is a network camera.

[0028] The scene attribute information includes camera pose parameters, surface material reflectivity of objects in the monitored scene, and surface morphology information.

[0029] The camera pose parameters include camera intrinsic parameters, installation position, installation height, pitch angle, and orientation. Camera intrinsic parameters can be obtained through a calibration plate, while the installation position, orientation, and pitch angle can be read from construction data or measured on-site.

[0030] The reflective properties of the surface material include smooth metallic surfaces, metallic coated surfaces, water surfaces, and diffuse reflective surfaces. The reflective properties of the equipment surface material can be read from the equipment register and 3D model, or the corresponding material category can be set by marking the equipment area in the screen with polygons in the initialization interface.

[0031] The surface morphology information includes planes, cylindrical surfaces, and irregular curved surfaces. For pipes, their surface morphology is marked as cylindrical and the projection axis and visible radius are recorded; for boxes and platforms, their surface morphology is marked as planes; for equipment with complex shapes, their surface morphology is marked as irregular curved surfaces; for pools, their surface morphology is marked as water surfaces (treated as planes with the normal facing upwards).

[0032] For fixed light sources, the direction of the light source is determined by its location and illumination direction. For outdoor natural light, the direction of sunlight is determined based on the location, date, and time. For surface locations... The direction of specular reflection from the surface outward can be represented as:

[0033] In the formula, For position The unit normal vector pointing outwards from the surface. For a moment By location The unit vector pointing to the light source. For a moment By location The outward-facing mirror reflection unit direction.

[0034] calculate The angle between the surface location and the line of sight pointing from the surface to the camera's optical center is considered; the smaller the angle, the higher the consistency of the reflection direction. For cylindrical pipes, the varying normal field is calculated along the pipe's cross-section to obtain strip-shaped high-reflectivity candidate areas that move with the direction of sunlight, rather than permanently shielding the entire pipe area.

[0035] Reference frames are acquired over one or more complete lighting cycles. Areas containing people, vehicles, or mobile devices are excluded using people detection and foreground motion detection. Then, the brightness distribution, highlight frequency, chromaticity stability, and gradient direction stability are statistically analyzed at each location, either pixel-wise or grid-wise. Specifically, brightness distribution is characterized by the mean brightness and upper quantile; highlight frequency is calculated as the ratio of highlight occurrences to the total number of statistical frames; chromaticity stability is measured by the variance of the chromaticity values ​​of the grid at different times—a smaller variance indicates more stable chromaticity and better conformity to the color characteristics of specular reflection; gradient direction stability is measured by the dispersion of the principal gradient direction of the grid at different times—a smaller dispersion indicates more stable gradient direction and better conformity to the geometric characteristics of fixed surface reflection. Statistical parameters include mean brightness, upper quantile brightness, highlight frequency, chromaticity stability, gradient direction stability, and common highlight periods. Finally, a scene description item is generated for each grid.

[0036] In the formula, Indicates the material category, Indicates surface morphology. Indicates the surface normal or normal range. Indicates the frequency of highlight occurrence. Indicates brightness distribution. Indicates color stability, Indicates gradient direction stability. This indicates the most common time periods for highlighting. This indicates the area to which the operation belongs. The scene description items together constitute the scene constraint baseline.

[0037] The scene constraint baseline includes: the consistency of reflection direction calculated based on the camera's line of sight, the device's surface normal or curvature, the fixed light source direction, or the sunlight direction that changes over time; and the brightness distribution, highlight frequency, chromaticity stability, gradient direction stability, and common highlight periods of each image grid obtained statistically from the reference frame within at least one observation period.

[0038] Example 2: Determination of reflection interference area and candidate target protection area.

[0039] This embodiment illustrates the specific process of determining the reflection interference region and the candidate target protection region in the current frame image.

[0040] After the current frame enters the edge computing node, it is first registered with the baseline image using feature points of the fixed background to correct pixel shifts caused by slight camera vibrations or maintenance. The following features are calculated for each grid cell: Brightness This indicates the degree to which the grid brightness exceeds the upper quantile of the historical brightness at that location; Low chromaticity This indicates the degree to which the bright area approaches the colorless specular highlight; Brightness residual This indicates the positive deviation of the current brightness from the estimated diffuse reflectance value at that location; Bright spot gradient direction This indicates the degree of consistency between the current principal direction of the bright spot and the reflection direction predicted based on the surface normal and the incident light direction; Historical characteristics This indicates whether the current time falls within the common highlighted period of that position; Sports protection features This indicates whether the grid moves in a consistent manner with the trajectory of a person, the outline of a human body, or a personal object.

[0041] Determining the area of ​​reflected interference: Extract feature information from the current frame image relative to the scene constraint baseline, the feature information including highlight level. Low chromaticity Brightness residual Bright spot gradient direction and historical characteristics ; Based on the aforementioned feature information, a weighted calculation is performed according to the material weight coefficient set corresponding to the surface material reflectivity at the corresponding location to obtain the reflectivity confidence score for each image location. The reflectivity confidence score can be calculated using the following formula: ; In the formula, Features for sports protection; , , , and According to position Material category The corresponding feature weights are read, and the weights are categorized by material type. The components together constitute the set of material weight coefficients for the corresponding positions. The target protection weights are defined by the sigmoid function, which limits the results to the range of 0 to 1. The directional feature weights for smooth metallic surfaces can be higher than those for metallic coated surfaces, while the weights for historical and spatial fluctuation features of water surfaces can be increased accordingly. For water surface regions, the motion protection features can be reduced. The weights are adjusted to avoid water surface undulations being misinterpreted as human movement, thus reducing the confidence level of the reflection. The reflection confidence The image is compared with a preset confidence threshold for reflection, which ranges from 0.3 to 0.8, with an initial value preferably of 0.6. The threshold is dynamically adjusted based on the frequency of highlight occurrence in the scene constraint baseline, or is set differently according to different material categories. Image positions that reach or exceed the preset confidence threshold for reflection are segmented to form initial reflection regions. Based on the consistency of reflection direction in the scene constraint baseline, the specular reflection direction is calculated as the predicted reflection direction by combining the surface normal and the light source direction. The initial reflection area is then oriented and expanded along the predicted reflection direction to cover the continuous bright band on the metal pipe. For water surface reflection, the reflection area can be expanded by combining the high-frequency brightness fluctuations of adjacent grids. The reflection interference area M is obtained after directional expansion.

[0042] Determination of candidate target protection areas: Obtain evidence of target presence in the current frame image, including target motion consistency information (corresponding to motion protection features). (This refers to whether the grid moves in a consistent manner with the person's trajectory), at least one of the following: human contour information quickly acquired on a low-resolution image by a lightweight human detector or head and shoulder detector independent of the target detection network, and infrared contour information acquired through the infrared imaging channel. Evidence of the existence of the target (such as...) The value is compared with a set protection threshold; the set protection threshold can be a uniform threshold or a differentiated threshold. Unified threshold: When only a single piece of evidence is obtained or when different pieces of evidence have the same dimensions and comparable reliability, all evidence shares the same threshold; in the case of only a single piece of evidence, if the evidence reaches or exceeds the set protection threshold, it is determined to be a candidate target protection area. Differentiation threshold: When different pieces of evidence have different dimensions or different levels of reliability, each piece of evidence corresponds to its own threshold. Each piece of evidence is independently compared with its own threshold. Any piece of evidence that reaches or exceeds its corresponding threshold is determined to be a candidate target protection area. When evidence of multi-dimensional targets is fused, the protection threshold can be dynamically adjusted: if the detection rate of the true target in the candidate target protection area remains low, the threshold is lowered; if the false protection ratio of non-target areas remains high, the threshold is raised. When any piece of acquired evidence reaches or exceeds its corresponding protection threshold, the corresponding image area is determined as the candidate target protection area P. The reflection interference region M and the candidate target protection region P are output separately, and they are not merged into a simple deletion region.

[0043] This embodiment achieves accurate localization of reflection interference areas through scene physical constraint-driven reflection confidence calculation and directional expansion. Simultaneously, by utilizing candidate target protection areas, it compares the target presence evidence (including at least one of the following: target motion consistency information, human contour information obtained through a lightweight human detector or head and shoulder detector independent of the target detection network, and infrared contour information obtained through the infrared imaging channel) acquired in the current frame image with a set protection threshold. If any acquired evidence reaches or exceeds the threshold, the area is determined as a candidate target protection area, effectively distinguishing between fixed reflection false targets and real moving targets. This ensures that real targets within the reflection interference area are reliably preserved and false target false alarms are significantly suppressed, providing accurate spatial priors for subsequent exposure control and feature correction.

[0044] Those skilled in the art should understand that the determination of the reflection interference area and the candidate target protection area in step S2 is not limited to the above implementation method: Different feature combinations can be used to calculate reflectance confidence. For example, in scenarios with stable lighting where only differences in material reflectance need to be distinguished, only two features, brightness and material weight, can be used for simplified calculation to reduce the computational overhead of edge nodes. In scenarios with ample computational resources, multi-scale feature fusion can also be introduced, that is, reflectance confidence is calculated at multiple image resolution scales, and then the final reflectance confidence map is obtained by weighted summation or taking the maximum value, so as to improve the adaptability to multi-scale targets.

[0045] Different threshold strategies can be used to determine reflection interference areas. For example, an adaptive threshold can be used, which dynamically adjusts the threshold according to the overall brightness distribution of the current frame to maintain consistent sensitivity in reflection determination under different lighting conditions; alternatively, a multi-threshold hierarchical determination can be used, setting different thresholds for strong and weak reflection areas to obtain a more refined division of reflection areas.

[0046] Different fusion methods can be used to determine the protection area of ​​candidate targets. For example, weighted average fusion, maximum value fusion, or confidence-based selection fusion can all be used to protect the real targets within the reflection area.

[0047] Example 3: Exposure parameter control.

[0048] The candidate target protection area is determined by step S2, and this step directly uses the candidate target protection area.

[0049] Within the working area, the overexposure ratio of the reflection interference region M and the underexposure degree of the candidate target protection region are calculated respectively. The overexposure ratio refers to the proportion of pixels whose brightness reaches or exceeds the overexposure threshold within the coverage area of ​​the reflection interference region, out of the total number of pixels in that region; the overexposure threshold is a brightness threshold used to determine whether a pixel has reached the overexposure standard. The underexposure degree is the proportion of pixels whose brightness is lower than a preset underexposure threshold within the candidate target protection region, or the degree to which the average brightness of the candidate target protection region is lower than a preset expected brightness value. The underexposure degree threshold can be determined based on the minimum brightness required for the target detection network to effectively extract the target contour and texture.

[0050] If neither the overexposure ratio of the highlighted area covered by the reflection interference region M nor the underexposure degree of the candidate target protection area meets the triggering conditions, then multi-exposure acquisition is not required. Single-frame processing is used, directly using the current frame as the input image for target detection, and the static reflection region of the corresponding time period is read from the scene constraint baseline as the calculation prior. The static reflection region is the region recorded in the scene constraint baseline that has fixed reflection characteristics within the corresponding time period, that is, the image region in the scene constraint baseline where the frequency of highlight occurrence is higher than a set frequency threshold. The triggering conditions include a high-brightness overexposure ratio threshold and an underexposure degree threshold. The high-brightness overexposure ratio threshold is used to determine whether the overexposure degree of the reflection interference area M is severe enough to trigger multiple exposure acquisition. Its value is determined according to the specific scene and imaging conditions to balance image quality and edge computing efficiency. The preferred value range is 10% to 25%, and the preferred initial value is 20%. The specific basis includes the target scale (low value of 10% to 15% for small targets and high value of 20% to 25% for medium and large targets), scene lighting conditions (low value for low light at night or welding arc interference scenes), and the computing resources of the edge device. The threshold is adaptively optimized through a feedback update step.

[0051] When the overexposure ratio or underexposure level of highlights reaches the trigger condition, multi-exposure acquisition and fusion processing is performed: Determine exposure parameters: Determine the exposure order and exposure amount of short exposure frames, medium exposure frames, and long exposure frames based on the scene constraint baseline.

[0052] The exposure amount of the short exposure frame is determined based on the upper quantile of the brightness of the reflection interference region M, so that the upper quantile of the brightness of the reflection interference region is lower than the overexposure threshold. The upper quantile refers to the brightness value at the upper quantile (e.g., the 95th percentile) after sorting the brightness values ​​of all pixels in the reflection interference region M from low to high, and is used to characterize the high brightness level of the region. The overexposure threshold is a brightness threshold used to determine whether a pixel has reached the overexposure standard, with a value range of 230 to 245 (8-bit image), and the initial value is preferably 240. This threshold, together with the aforementioned high brightness overexposure ratio threshold (10% to 25%), constitutes a linkage mechanism of "high brightness overexposure ratio statistics - short exposure amount determination" - the high brightness overexposure ratio threshold determines whether to trigger multi-exposure acquisition, and the overexposure threshold determines the level to which the brightness of region M needs to be suppressed in the short exposure frame. A lower value is used in strong reflection scenes (more sensitive), and a higher value is used in low-light scenes (preserving more dark information). This threshold can be adaptively optimized through feedback update steps.

[0053] The exposure amount of the long exposure frame is determined based on the lower quantile of the luminance of the candidate target protection area, so that the lower quantile of the luminance of the candidate target protection area reaches the range of extractable contours. The lower quantile refers to the luminance value at the lower quantile (e.g., the 5th percentile) after sorting the luminance values ​​of all pixels in the candidate target protection area from low to high, and is used to characterize the dark area level of that area. The range of extractable contours refers to the interval where the luminance value is not lower than 50 (8-bit image). Above this luminance level, the target detection network can effectively extract key features such as the edge of the safety helmet, the torso contour, and the texture of the seat belt.

[0054] The long exposure time is limited by the target's motion speed: at higher speeds, the long exposure time is shortened and the gain is increased to limit motion blur; at lower speeds, the long exposure time can be appropriately extended to obtain richer dark details. The balance between the two can be determined based on the detection accuracy requirements and motion blur tolerance of the specific scene. For points without motion speed sensors, the motion speed can be estimated by the change in target position between frames.

[0055] Multi-exposure frame registration: Static background registration is performed using the intermediate exposure frame as the geometric reference, and local registration of the moving area is performed using the target optical flow.

[0056] Image fusion and weight allocation: For each exposure frame, a basic fusion weight is generated, consisting of exposure suitability, local contrast, and registration reliability. Using the candidate target protection region as a spatial prior, the fusion weights are adjusted as follows: (1) For the image region within the candidate target protection area, increase the fusion weight of long exposure frames or medium exposure frames to ensure that the real target located in the reflection interference area maintains complete contour and texture features in the fused image; wherein, when there is a moving target in the region, the fusion weight of medium exposure frames is increased first, and when the region is a static background, the fusion weight of long exposure frames is increased first. (2) For image regions within the reflection interference area that do not belong to the candidate target protection area, increase the fusion weight of short exposure frames and decrease the fusion weight of long exposure frames; (3) For the underexposed areas of the candidate target protection area, increase the fusion weight of medium exposure frames or long exposure frames; (4) For moving regions, reduce the fusion weight of exposure frames with insufficient registration reliability; When the candidate target protection area overlaps with the reflection interference area, the weight adjustment rule of the candidate target protection area shall prevail in the overlapping area, that is, mode (1) shall be executed first, and the operation of increasing the weight of short exposure frames and decreasing the weight of long exposure frames in mode (2) shall not be executed or shall be restricted, so as to avoid the real target features being suppressed by short exposure. After weight normalization of each exposure frame, the images are fused to generate a fused image, which serves as the input image for target detection. The fused image uses the same image coordinate system as the current frame, ensuring that the reflection interference region M and the candidate target protection region P maintain spatial consistency in the fused image.

[0057] Therefore, multi-exposure processing is not a general HDR that uses a fixed exposure ratio for the entire image. Instead, it is determined by the coverage of the reflection interference area M at a fixed point, the brightness of the candidate target, and the motion state, which together determine whether to trigger, what exposure to use, and the fusion weight of each area.

[0058] The above method achieves on-demand, accurate, and coordinated processing of strong light reflection and backlight interference through regional trigger judgment, differential exposure parameter determination, and weight allocation guided by physical priors. Multi-exposure fusion is only initiated when overexposure or underexposure reaches the trigger condition, avoiding the processing burden caused by continuous calculations in general HDR schemes. The short exposure amount is determined based on the upper quantile of brightness in the reflection interference area to suppress highlight clipping, while the long exposure amount is determined based on the lower quantile of brightness in the candidate target protection area to supplement shadow details, achieving differentiated exposure control of "suppressing highlights and supplementing shadows". The fusion weight allocation uses the static reflection area in the scene constraint baseline as the calculation prior, increases the weight of short exposure frames in the reflection interference area, increases the weight of medium and long exposure frames in the underexposed candidate target area, and reduces the weight of unreliable registration frames in the moving area. This allows the fused image to retain the complete features of the real target while suppressing reflection interference, providing a higher quality input image for subsequent feature correction and target detection.

[0059] Example 4: Feature correction.

[0060] This embodiment illustrates the specific process of feature correction using the reflection interference region and the candidate target protection region.

[0061] The target detection network includes a target feature branch, an interference feature branch, a multi-scale feature correction module, and a detection head.

[0062] The target detection input image (the current frame image when multi-exposure acquisition and fusion processing is not triggered in step S3; the fused image generated in step S3 when multi-exposure acquisition and fusion processing is triggered) or the fused image is input into the target detection network, and multi-scale target features are extracted through the target feature branch of the target detection network. , l =1,2,3,…,L, where L is the number of scales.

[0063] Material category codes are generated based on the surface material reflectivity attributes in the scene attribute information. Specifically, smooth metallic surfaces are coded as 0, metallic coated surfaces as 1, water surfaces as 2, and diffuse reflective surfaces as 3. A brightness residual is generated based on the brightness deviation of the input image for target detection relative to the scene constraint baseline.

[0064] The reflection interference region M, the material type code, and the brightness residual are input into the interference feature branch of the target detection network, and the interference feature branch outputs compensation features. The basic suppression coefficient is determined based on the material category code and then adjusted according to the feature layer scale to obtain the reflection suppression coefficient at each scale. Among them, the high-resolution feature layer sets a lower reflection suppression coefficient for at least one small target among helmets and seat belts than for the other feature layers.

[0065] On the multi-scale feature layer, the reflection interference region M is scaled to the corresponding scale. The reflection suppression coefficient at each scale is determined based on the material category code. Specifically, the smooth metal surface area. A value of 0.8 is used for the metal coating area. The value is 0.5, for the water surface area. A value of 0.6 represents the diffuse reflective surface area. The value is 0.1.

[0066] For the l Target features at each scale The following formula is used to obtain the first... Correction features at each scale : ; in, This represents the reflection interference area at the corresponding scale. This represents the protected area for candidate targets at the corresponding scale. The compensation features are extracted from non-highlight textures and brightness residuals by the interference feature branch. For compensation coefficient, This is for element-wise multiplication. and A two-dimensional spatial mask is used, broadcast along the feature channel dimension during multiplication operations, so that each channel shares the same spatial suppression and protection mode; the compensation feature Only Feature correction is performed within the specified area to avoid introducing unnecessary texture compensation in areas with pure reflection interference where there is no evidence of real target protection. No compensation will be provided for the areas in question.

[0067] The meaning of the formula is: in the area of ​​reflection interference If no evidence of protection for the candidate target exists ( If =0), then the target feature Suppressed, the intensity of suppression is reduced by Control; if evidence of protection for candidate targets exists ( If the value is 1), then the suppression is lifted, and the true target features are preserved. Meanwhile, As a compensating feature, only The region (i.e., the region with real evidence of target protection) supplements the correction features with the non-highlight texture information learned by the interference branch, while... Pure reflection interference areas are not compensated to avoid introducing unnecessary texture information into the reflection areas.

[0068] Based on the correction characteristics of each scale After obtaining the corrected target detection features F, the detection head of the target detection network performs detection at each scale based on the corresponding corrected features. Perform security target classification and location. For small targets such as helmets and seat belts, the feature layer with the highest resolution among the multi-scale target features is set with a lower reflection suppression coefficient than the other feature layers to prevent the small target features from being completely cleared.

[0069] This embodiment achieves structured suppression of reflection interference and differentiated protection of the real target through dual-branch collaborative correction of the target feature branch and the interference feature branch. Within the reflection interference area, location features without target protection evidence are effectively suppressed, while location features with target protection evidence are suppressed, thus preserving the real target features. Simultaneously, the compensation features extracted by the interference feature branch only supplement the correction features within the candidate target protection area, effectively avoiding interference from reflection features on target detection. A differentiated suppression coefficient driven by material category encoding allows surfaces with different reflective properties to obtain appropriate suppression strengths, avoiding excessive weakening of low-reflectivity area features by uniform suppression. For small targets such as safety helmets and seat belts, a lower reflection suppression coefficient is set in the high-resolution feature layer to prevent the complete removal of small target features, ensuring detection accuracy. Unlike general target detection networks that rely solely on statistical correlation, this embodiment guides feature correction based on scene physical constraints, enabling the model to understand the distribution patterns of reflection interference from a causal perspective, providing reliable features with physical prior correction for final target detection.

[0070] Example 5: Target Detection.

[0071] This embodiment illustrates the specific process of performing security target classification and localization based on the corrected target detection features.

[0072] After feature correction is completed in step S4, the target detection network obtains the corrected features. ( l =1,2,…,L); correction characteristics at each scale. This constitutes the corrected multi-scale target detection features. The detection head receives the corrected multi-scale target detection features and, at each scale, applies the corresponding corrected features. Classification and positioning of security targets.

[0073] The detection head adopts a decoupled structure, including two parallel modules: a classification branch and a regression branch.

[0074] The classification branch receives the corrected features at each scale and maps them to confidence scores for each category. For each candidate target location, the classification branch outputs a probability distribution vector, representing the probability that the location belongs to each preset category (such as personnel, safety helmet, work clothes, safety belt).

[0075] Since the features input to the classification branch have undergone reflection interference suppression and compensation correction in step S4, the feature responses located within the reflection interference region have been effectively suppressed, while the features of the real target located within the candidate target protection region have been preserved. Therefore, the classification branch's determination result is not easily affected by the high-brightness reflection of the metal surface—fixed reflective bright spots cannot pass the classification threshold because their features are suppressed, while the real target outputs a high-confidence classification result because its features are preserved.

[0076] The regression branch also receives the same correction features to predict the positional parameters of the target bounding box. Specifically, the regression branch predicts the center coordinate offset (Δx, Δy) of the target bounding box relative to a preset reference point (such as the center of the feature map grid or a preset anchor box), as well as the width w and height h of the bounding box.

[0077] After decoding the predicted offsets and mapping them back to the original image coordinate system, the precise bounding box of the target in the image is obtained, denoted as ( ).

[0078] The classification branch's class confidence score is merged with the regression branch's bounding box position to form a set of candidate detection boxes with class labels and confidence scores. The detection head typically outputs a large number of candidate detection boxes, many of which contain overlapping boxes and low-confidence boxes. Non-maximum suppression (NMS) is used to filter the candidate boxes. Sort all candidate boxes according to classification confidence; retain the candidate box with the highest confidence and delete other candidate boxes whose intersection-union ratio exceeds a preset threshold (e.g., 0.45); Repeat the above steps until all candidate boxes have been processed.

[0079] After NMS filtering, only one optimal detection box is retained for each target.

[0080] After NMS, the object detection network outputs object detection results for the current frame image. Each detection result includes three types of information: Target category (e.g., "personnel", "safety helmet") Bounding box position (e.g., [ ]) Confidence score (reflects the reliability of the test result) The above test results can be output directly.

[0081] It should be noted that the implementation of the detection head in the object detection network is not limited to the decoupled structure described above. In other implementations, the detection head can adopt the following variations: Anchor frame method: Preset anchor frames of different sizes and aspect ratios at multiple scales, and use regression branches to predict the position offset and size scaling of the anchor frames; Anchor-free methods (CenterNet, FCOS, etc.): Directly predict the center point location and bounding box size of the target without pre-setting anchor boxes, resulting in less computation and more flexible deployment; Classification and regression ensemble (YOLO series): The classification branch and regression branch predict in parallel on the same feature map, and the class probability and bounding box parameters are directly output at each grid position of the feature map; Transformer methods (DETR, etc.): Directly output a fixed number of target detection results through a self-attention mechanism, without the need for NMS post-processing steps.

[0082] All of the above variants can perform classification and localization functions based on the corrected target detection features, and all fall within the protection scope of this invention.

[0083] Example 6: Deployment and adaptation.

[0084] The equipment materials, lighting conditions, and reflection distribution vary at different monitoring points in a power plant. Training or fine-tuning the target detection model separately for each point would lead to low deployment efficiency and difficulty in ensuring performance consistency. To address the problem of poor deployment flexibility across different points, this invention also includes a deployment adaptation step. This embodiment illustrates the deployment adaptation process that separates the basic model from the point adaptation parameters.

[0085] The scene constraint baselines are spatially converged to obtain a point scene vector that characterizes the overall scene characteristics of the current monitoring point. Specifically, the material type, reflection confidence, common reflection periods, and working area of ​​each grid are converged and combined with the camera installation height, pitch angle, and target pixel scale to form the point scene vector. This vector represents information such as the material ratio, surface ratio, reflection area distribution, backlight period, camera pose, and target pixel scale of the point.

[0086] Point adaptation parameters are generated or selected based on the point scene vector. Specifically, the point adaptation network outputs the channel scaling coefficients and biases of each feature layer based on the vector, or selects a set from multiple pre-trained sets of adaptation parameters. These point adaptation parameters are used to adjust the reflection suppression coefficient of the target detection network. The detection threshold comprises at least one of the following: feature channel scaling factor, feature channel bias, and detection threshold. The detection threshold is the minimum classification confidence threshold required for the candidate detection boxes output by the target detection network to be retained as the final detection result. In the post-processing of the target detection network, candidate detection boxes can be filtered by classification confidence based on the detection threshold, eliminating candidate detection boxes with classification confidence lower than the detection threshold. The range of the detection threshold can be determined based on the degree of reflection interference at the location and the tolerance for missed detections. Higher values ​​can be used for locations with strong reflection interference to reduce false alarms, while lower values ​​can be used for locations with weak reflection interference but many small targets to reduce missed detections. The feature channel scaling factor is used to scale the response amplitude of each feature channel, and the feature channel bias is used to shift the response of each feature channel. The feature channel scaling factor and feature channel bias can be generated based on the location scene vector or selected from multiple pre-trained sets of adaptation parameters. The spatial aggregation method of the point scene vector includes: statistically analyzing the material category ratio, surface ratio, average reflection confidence, frequency of highlight occurrence, distribution of common highlight time periods, and distribution of work area for each image grid, and then splicing or weighting these values ​​with the camera installation height, pitch angle, and target pixel scale to obtain the point scene vector used to characterize the overall scene characteristics of the current monitoring point.

[0087] Using the target detection network as a base model, the base model, the point adaptation parameters, the scene constraint baseline, and the preset task rules are bound together to form a deployable unit associated with the current camera channel. The base model is a target detection network trained in multiple power plant scenes that preserves general target features. The preset task rules include task types and event rules; each task type corresponds to a spatial location constraint relationship, which is contained within the event rules.

[0088] When deploying at new locations, at least some parameters of the base model remain unchanged; only the location adaptation parameters and the scene constraint baseline are updated. That is, for new locations, only a new scene constraint baseline needs to be established and new location adaptation parameters generated, without retraining the base model.

[0089] The deployment record includes at least the camera channel identifier, basic model version, scene constraint baseline version, point adaptation parameter version, work area, task type, and event rules in the preset task rules. After receiving the video stream, the edge computing node first performs lightweight registration and reflection trigger judgment; if no trigger condition is triggered, it uses a buffer mask and infers along a single-frame path; if the trigger condition is met, it only performs multi-exposure fusion and interference feature branch calculation on the work area or the vicinity of the candidate target to control additional computational overhead.

[0090] The deployment and adaptation process aggregates scene constraint baselines into point-scene vectors and generates corresponding point-adaptation parameters. These parameters are then bound to a base model that stores general target features, forming deployable units. This allows for rapid deployment across points by establishing scene constraint baselines and generating adaptation parameters, eliminating the need for retraining or fine-tuning the base model. Simultaneously, since fixed reflection interference at each point is constrained within the scene constraint baselines and point-adaptation parameters, rather than directly participating in the base model's parameter updates, the repeated entry of fixed reflection pseudo-targets at single points into the training set, preventing the degradation of general target features, ensures the base model maintains stable recognition capabilities during continuous operation. Furthermore, the unified base model version and independent management of adaptation parameters provide clearer version control and performance traceability during large-scale multi-channel deployments, offering engineering support for the system's large-scale deployment.

[0091] Example 7: Event determination.

[0092] The target detection network outputs the target category and location in a single frame image, but practical security applications require structured security events verified by spatial logic, temporal consistency, and reflection interference. Therefore, the method of this invention also includes an event determination step. This embodiment illustrates the specific process of generating security events by combining spatial relationships and temporal continuity.

[0093] Personnel tracking results are obtained by associating targets in consecutive frames using a multi-target tracking algorithm. The target detection results are then associated with the personnel tracking results, and spatial relationship verification is performed according to the task type and its corresponding spatial constraints in the preset task rules. The spatial relationship verification includes: the safety helmet target should be located in the personnel's head and shoulder area; the work clothes or reflective strip target should be located in the personnel's torso area; and the safety belt and its attachment points should have a preset spatial relationship with the high-altitude work area and the personnel's torso area.

[0094] For safety helmet identification, candidate safety helmet frames are associated with key points on the person's head and shoulders or the upper region of the person's frame. If a bright spot is located only in the metal equipment area, has no personnel trajectory, and remains fixed for a long time, then even if its color or shape is similar to a safety helmet, no "not wearing a safety helmet" event will be generated. If the candidate target protection area P shows evidence of a target presence in that area, then a further determination of whether it is a real target is made based on time continuity verification.

[0095] For work uniform or reflective vest recognition, the candidate area of ​​reflective strips is associated with the torso area of ​​the person. Although reflective strips have the characteristic of high brightness, they should move with the trajectory of the person; strip-shaped reflections fixed on the surface of equipment do not meet the requirements of motion consistency and human spatial relationship, and are therefore subject to reduced event confidence.

[0096] For safety belt identification, the safety belt body, connecting rope, or attachment point is associated with the person's torso, the high-altitude work area, and the preset attachment point area. For safety belt areas that overlap with reflections from metal railings, if the candidate target protection area shows that it moves with the person, the detection result is retained.

[0097] The overlap ratio between the detection bounding box of the target detection result that has passed the spatial relationship verification and the reflection interference region M is calculated. The event confidence score is used to characterize the credibility of the detection result as a security event. The event confidence score can be determined based on at least one of the following: target detection confidence score, spatial relationship verification result, reflection overlap ratio, and temporal continuity.

[0098] When the detection frame highly overlaps with the reflection interference area M (e.g., the overlap ratio is greater than 0.5), maintains a fixed spatial position at multiple times, and lacks personnel movement or infrared contour support, the event confidence of the detection result is reduced.

[0099] When the detection frame overlaps with the reflection interference region M but has continuous human body trajectory or infrared contour support, the detection result is retained using the candidate target protection region P.

[0100] Perform time continuity verification on the retained detection results (e.g., valid for multiple consecutive frames), and generate security events from the detection results that pass the verification.

[0101] This step employs a triple mechanism of spatial relationship verification, reflection overlap verification, and temporal continuity verification to structurally validate the target detection results. Spatial relationship verification filters out detection results with spatial misalignment based on physical logic such as "safety helmet located in the head and shoulder area," "work clothes located in the torso area," and "safety belt associated with the high-altitude work area." Reflection overlap verification uses the reflection interference area M and the candidate target protection area P to perform dual judgment on the detection frame, filtering out false targets that overlap with a fixed reflection height and have no motion or infrared support, while retaining detection results with evidence of genuine targets. Temporal continuity verification requires the target to exist stably across multiple consecutive frames, eliminating false alarms in a single frame caused by instantaneous interference such as welding arc light or flashing lights. The synergistic effect of these three verifications ensures that the security events output by the system have high reliability and operability, avoiding the problem of a large number of false alarms caused by general target detection solutions directly outputting unverified detection results that cannot be directly used for actual alarms.

[0102] Example 8: Feedback Update.

[0103] If false alarms generated on-site are not distinguished between fixed reflection interference and errors in identifying real targets, incremental training can easily lead to repeated learning of the same type of false targets. This not only fails to effectively reduce the false alarm rate but may also damage the general target features already learned by the model, resulting in model degradation. To address the problem of a crude false alarm feedback mechanism, this invention also includes a feedback update step. This embodiment illustrates the specific process of updating the scene constraint baseline or model parameters according to the cause of the false alarm.

[0104] Acquire false alarm and missed alarm information verified manually or by multiple sensors. The system saves the corresponding camera identifier, time, target trajectory, detection box, overlap ratio of reflection interference area, and model version for each event.

[0105] The false alarms and missed alarms are mapped back to the image coordinates of the corresponding cameras to generate an error spatial distribution map. Specifically, the center of the detection box is mapped to the error spatial distribution map and accumulated within a preset time window.

[0106] When an error repeatedly occurs at a fixed location, overlaps with the reflection interference area M to a set condition (e.g., overlap ratio greater than 0.6), and exhibits a regular variation with the duration of sunlight or the status of a fixed light source, the error is classified as a scene reflection error. The frequency of highlight occurrence h(p), material weight coefficient set, preset threshold for reflection confidence, or common highlight periods in the scene constraint baseline are then updated. For example, if false alarms are concentrated in multiple areas with the same material and curvature, the corresponding material weight coefficient set or the weight of the consistency of reflection direction is adjusted.

[0107] When an error occurs due to movement of personnel or equipment and does not meet the fixed reflection characteristics, the error is classified as a target recognition error, and the corresponding sample is added to the hard sample set. This is used to update the location adaptation parameters or, after centralized verification, to update the base model. Specifically, the location adaptation parameters are updated and verified first; only when multiple locations are verified as having a common target feature problem are the samples used to update the base model.

[0108] The base model uses version number management. New locations added after an update use the new version, while existing locations retain the original version until actively upgraded. The updated version is first tested on a subset of locations. If precision, recall, false alarm rate, or processing latency fails to meet set conditions (failing to meet the preset threshold for reflection confidence), the system reverts to the previous scenario constraint baseline or model version. In other words, the updated version is first tested on a subset of locations, and performance metrics before and after the update are compared. If the performance meets the standards, it is released; otherwise, it is rolled back.

[0109] This step maps false alarms and missed alarms to an error spatial distribution map, and classifies them into "scene reflection errors" and "target recognition errors" based on their spatial distribution characteristics and temporal variation patterns. Scene reflection errors occur repeatedly at fixed locations, overlap with reflection interference areas, and change with illumination patterns. Physical parameters such as the highlight frequency, material weight coefficient set, or preset reflection confidence threshold at the corresponding locations in the scene constraint baseline are updated to make the physical model more accurate. Target recognition errors occur due to movement of personnel or equipment and do not meet fixed reflection characteristics. Corresponding samples are added to a difficult sample set, and point adaptation parameters are updated preferentially. The basic model is only updated when multiple points verify that the problem is a common issue, avoiding the contamination of common target features by single-point fixed reflection pseudo-targets. Different update strategies are used for the two types of errors, achieving iterative optimization of the physical baseline and model parameters. A mechanism of "partial point trial operation—release if performance meets standards, rollback if not" ensures system stability, allowing the system's detection performance to iterate positively without degradation during continuous operation.

[0110] Example 9: A power plant security target detection system based on scene physical constraints, the system being used to execute the above method, the system comprising: The scene constraint modeling module is used to perform step S1; A reflection interference calculation module is used to perform step S2; The acquisition and control module is used to execute step S3; The feature correction and target detection module is used to perform steps S4 and S5.

[0111] The aforementioned modules can be deployed on the camera side, edge computing nodes, edge management servers, or central training platforms. Video images are processed preferentially at edge computing nodes, while the central platform receives event results, difficult samples, and version information. Edge computing nodes can manage multiple video channels simultaneously and process them independently according to the scene constraint baselines of each channel.

Claims

1. A method for detecting security targets in power plants based on scene physical constraints, characterized in that, Includes the following steps: S1. Obtain scene attribute information of fixed monitoring points in the power plant, and establish a scene constraint baseline based on the scene attribute information that reflects the distribution pattern of reflection interference and the pattern of illumination change in the scene; wherein, the scene attribute information includes at least camera pose parameters, surface material reflectivity of objects in the monitored scene, and surface morphology information. S2. Match the current frame image with the scene constraint baseline, and determine the reflection interference area and candidate target protection area in the current frame image based on the matching result; the candidate target protection area is the image area in the current frame image that belongs to the actual security target and is determined from the current frame image based on the evidence of the existence of the target in the current frame image. S3. Based on the reflection interference area and the candidate target protection area, control the camera's exposure parameters to obtain the input image for target detection; S4. Input the target detection input image into the target detection network to extract target detection features; use the reflection interference region to suppress the feature response of the target detection features, and use the candidate target protection region to limit the intensity of the suppression to obtain the corrected target detection features; S5. Based on the corrected target detection features, classify and locate the security target, and output the target detection results.

2. The power plant security target detection method based on scene physical constraints as described in claim 1, characterized in that, The scene constraint baseline includes: the consistency of reflection direction calculated based on the camera's line of sight, the device's surface normal or curvature, the fixed light source direction, or the sunlight direction that changes over time; and the brightness distribution, highlight frequency, chromaticity stability, gradient direction stability, and common highlight periods of each image grid obtained statistically from the reference frame within at least one observation period. The camera pose parameters include camera intrinsic parameters, installation position, installation height, pitch angle, and orientation.

3. The power plant security target detection method based on scene physical constraints as described in claim 2, characterized in that, The methods for determining the reflection interference area include: Extract feature information from the current frame image relative to the scene constraint baseline, the feature information including at least one of highlight level, low chromaticity level, brightness residual and bright spot gradient direction; Based on the feature information, the reflection confidence of each image location is obtained by weighting the calculation according to the material weight coefficient set corresponding to the surface material reflectivity of the corresponding location. The reflection confidence score is compared with a preset threshold for reflection confidence score, and image locations that reach or exceed the preset threshold for reflection confidence score are segmented by thresholding to form initial reflection regions; Based on the consistency of reflection direction in the scene constraint baseline, the specular reflection direction is calculated as the predicted reflection direction by combining the surface normal and the light source direction, and the initial reflection area is directionally expanded along the predicted reflection direction to obtain the reflection interference area.

4. The power plant security target detection method based on scene physical constraints as described in claim 1, characterized in that, The method for determining the candidate target protection area includes: Obtain evidence of target presence in the current frame image. The evidence of target presence includes at least one of the following: target motion consistency information obtained through inter-frame motion detection, human contour information quickly obtained on a low-resolution image by a lightweight human detector or head and shoulder detector independent of the target detection network, and infrared contour information obtained through the infrared imaging channel. Compare the evidence of the existence of the target with a set protection threshold; When the evidence of the target reaches or exceeds the set protection threshold, the corresponding image region is determined as the candidate target protection region.

5. The power plant security target detection method based on scene physical constraints as described in claim 1, characterized in that, Step S3 includes: S31. Obtain the determined candidate target protection area; S32. Calculate the overexposure ratio of the high brightness of the reflection interference area and the underexposure of the candidate target protection area within the working area. S33. When neither the overexposure ratio nor the underexposure level reaches the trigger condition, single-frame processing is adopted, and the static reflection area of ​​the corresponding time period is read from the scene constraint baseline as the calculation prior for subsequent feature correction; the static reflection area is the area recorded in the scene constraint baseline that has fixed reflection characteristics in the corresponding time period; the trigger condition includes the overexposure ratio threshold and the underexposure level threshold. When at least one of the overexposure ratio and the underexposure degree reaches the trigger condition, multi-exposure acquisition and fusion processing is performed: The exposure order and exposure amount of short exposure frames, medium exposure frames, and long exposure frames are determined based on the scene constraint baseline. The exposure amount of the short exposure frame is determined based on the upper quantile of the brightness of the reflection interference region, so that the upper quantile of the brightness of the reflection interference region is lower than the overexposure threshold. The exposure amount of the long exposure frame is determined based on the lower quantile of the brightness of the candidate target protection region, so that the lower quantile of the brightness of the candidate target protection region reaches the range of extractable contours, and the long exposure time is limited according to the target movement speed. Static background registration is performed using the aforementioned exposure frame as a geometric reference, and local registration of the moving region is performed using the target optical flow in the following manner; For each exposure frame, a basic fusion weight is generated, consisting of exposure suitability, local contrast, and registration reliability. Using the candidate target protection region as a spatial prior, the fusion weights are adjusted according to the following rules: (1) For the image region within the candidate target protection area, increase the fusion weight of long exposure frames or medium exposure frames; wherein, when there is a moving target in the region, the fusion weight of medium exposure frames is increased first, and when the region is a static background, the fusion weight of long exposure frames is increased first. (2) For image regions within the reflection interference area that do not belong to the candidate target protection area, increase the fusion weight of short exposure frames and decrease the fusion weight of long exposure frames; (3) For the underexposed areas of the candidate target protection area, increase the fusion weight of medium exposure frames or long exposure frames; (4) For the motion region, reduce the fusion weight of the exposure frame with insufficient registration reliability; when the candidate target protection region overlaps with the reflection interference region, the weight adjustment rule of the candidate target protection region shall prevail in the overlapping region, that is, mode (1) shall be executed first, and the operation of increasing the weight of short exposure frame and decreasing the weight of long exposure frame in mode (2) shall not be executed or the execution of mode (2) shall be restricted. After normalizing the weights of each exposure frame, they are fused to generate a fused image, which is used as the input image for target detection.

6. The power plant security target detection method based on scene physical constraints as described in claim 5, characterized in that, Step S4 includes: S41. Input the target detection input image into the target detection network, and extract multi-scale target features through the target feature branch of the target detection network. , l =1,2,3,…,L, where L is the number of scales; S42. Generate a material category code based on the surface material reflectivity in the scene attribute information, and generate a brightness residual based on the brightness deviation of the target detection input image relative to the scene constraint baseline; S43. Input the reflection interference area, the material category code, and the brightness residual into the interference feature branch of the target detection network to extract interference features; S44. On the multi-scale feature layer, the reflection interference region is scaled to the corresponding scale, and the reflection suppression coefficient at each scale is determined according to the material category code. ; S45, Regarding the first Target features at each scale The following formula is used to obtain the first... Correction features at each scale : ； in, This represents the reflection interference area at the corresponding scale. This represents the protected area for candidate targets at the corresponding scale. The compensation features are extracted from non-highlight textures and brightness residuals by the interference feature branch. This is the compensation coefficient; S46. Based on the correction characteristics of each scale The corrected target detection features F are obtained, and the classification and localization of security targets are performed.

7. The power plant security target detection method based on scene physical constraints as described in claim 3, characterized in that, It also includes deployment and adaptation steps: Spatially converge the scene constraint baselines to obtain a point scene vector that characterizes the overall scene characteristics of the current monitoring point. Point adaptation parameters are generated or selected based on the point scene vector; the point adaptation parameters are used to adjust at least one of the reflection suppression coefficient, feature channel scaling coefficient, feature channel bias, and detection threshold of the target detection network. Using the target detection network as the basic model, the basic model, the point adaptation parameters, the scene constraint baseline, and the preset task rules are bound together to form a deployable unit associated with the current camera channel; the preset task rules include task types and event rules; the task type corresponds to a spatial position constraint relationship, and the spatial position constraint relationship is contained in the event rules; When deploying at new locations, at least some parameters of the base model remain unchanged, and only the location adaptation parameters and the scene constraint baseline are updated.

8. The power plant security target detection method based on scene physical constraints as described in claim 7, characterized in that, It also includes the event determination step: The personnel tracking results are obtained by associating targets in consecutive frames using a multi-target tracking algorithm. Spatial relationship verification is performed according to the task type and its corresponding spatial position constraint relationship in the preset task rules. The target detection results are associated with the personnel tracking results, and the target detection results are verified according to the spatial position constraint relationship corresponding to the task type. The spatial relationship verification includes: the safety helmet target should be located in the head and shoulder area of ​​the personnel, the work clothes or reflective strip target should be located in the torso area of ​​the personnel, and the safety belt and its attachment point should have a preset spatial relationship with the high-altitude work area and the torso area of ​​the personnel. Calculate the overlap ratio between the detection bounding box of the target detection result that passed the spatial relationship verification and the reflection interference area; When the detection frame highly overlaps with the reflection interference area, maintains a fixed spatial position at multiple times, and lacks personnel movement or infrared contour support, the event confidence of the detection result is reduced. When the detection frame overlaps with the reflection interference area, but has continuous human body trajectory or infrared contour support, the detection result is retained using the candidate target protection area; The retained detection results are checked for temporal continuity, and security events are generated from the detection results that pass the check.

9. The power plant security target detection method based on scene physical constraints as described in claim 7, characterized in that, It also includes feedback update steps: Acquire false alarm and missed alarm information that has been manually or multi-sensor verified, map the false alarm and missed alarm back to the image coordinates of the corresponding camera, and generate an error spatial distribution map; When an error repeatedly occurs at a fixed location, the overlap ratio with the reflection interference area reaches a set condition, and the error changes regularly with the sunshine period or the fixed light source status, the error is classified as a scene reflection error, and the frequency of bright appearance, material weight coefficient set, reflection confidence preset threshold, or common bright period in the corresponding position in the scene constraint baseline is updated. When an error occurs as a person or equipment moves and does not meet the fixed reflection characteristics, the error is classified as a target recognition error, and the corresponding sample is added to the hard sample set to update the location adaptation parameters or to update the basic model after centralized verification. The basic model is managed by version number. After an update, new locations will use the new version, while existing locations will retain the original version until they are actively upgraded. The updated version will be tested at some locations first. If the accuracy, recall, false alarm rate, or processing delay does not meet the set conditions, the model will be restored to the scene constraint baseline or model version before the update.

10. A power plant security target detection system based on scene physical constraints, the system being used to execute the power plant security target detection method based on scene physical constraints as described in any one of claims 1 to 9, characterized in that, The system includes: The scene constraint modeling module is used to perform step S1; A reflection interference calculation module is used to perform step S2; The acquisition and control module is used to execute step S3; The feature correction and target detection module is used to perform steps S4 and S5.

Images