A monitoring-based abnormal behavior detection method

Abstract

The application provides a kind of based on monitoring abnormal behavior detection method, comprising: detecting the gradient change range of the edge light intensity blur of highlight feature missing position, gradient analysis is carried out according to change range to obtain the distribution characteristics of edge blur area;According to the distribution characteristics of edge blur area, the local data point of light field sampling point missing is identified, and the filling result of missing data is generated according to local data point using interpolation reconstruction method;According to the filling result of missing data, the transient distortion characteristics caused by the dithering interference of reflective clothing hem edge are analyzed, the transient distortion distribution is obtained, and the dithering frequency and amplitude range are determined according to the transient distortion distribution;The posture change constraint of construction staff is obtained by analyzing dynamic light intensity change and construction staff motion track, and the correction parameter of posture change on light field transient distortion is obtained combined with the optimized light field sampling distribution.

Description

Technical Field

[0001] This invention relates to the field of information technology, and in particular to a method for detecting abnormal behavior based on monitoring. Background Technology

[0002] Ensuring the safety of construction workers during nighttime construction is of paramount importance, especially through intelligent recognition technology to determine whether reflective vests are being worn correctly, thereby reducing the risk of accidents. In complex nighttime lighting conditions, the high-brightness characteristics of reflective vests provide crucial information for safety monitoring; however, the limitations of current technology make it difficult to address the complex challenges in dynamic scenarios.

[0003] Current methods mostly rely on static light field models or single sensor data, which are difficult to adapt to the interference caused by changes in lighting conditions at construction sites, contamination of reflective clothing, and dynamic adjustments in personnel posture. When dealing with local light intensity attenuation on the surface of reflective clothing caused by contamination, these methods often reduce recognition accuracy due to blurred edge light intensity gradients, thus affecting the reliability of safety monitoring.

[0004] The core challenges of light field imaging technology at construction sites at night stem from the lack of data and transient distortion at light field sampling points. Contamination on the surface of reflective clothing weakens its high-brightness features, leading to incomplete sampling point data, especially in edge areas where ambient light scattering further blurs the light intensity distribution, making identification difficult. Simultaneously, the jittering of reflective clothing folds caused by changes in the posture of construction workers exacerbates transient distortion in the light field data, making it difficult for the system to accurately capture the spatial light intensity distribution of the reflective clothing. These two factors are interrelated: data loss weakens the integrity of the light field reconstruction, while fold jitter further disrupts the stability of the transient light field, resulting in decreased feature reconstruction accuracy. Therefore, how to optimize the reconstruction accuracy of the spatial light intensity distribution of reflective clothing under complex nighttime lighting conditions by dynamically adjusting the fusion weights of light field sampling point density and visible light compensation, combined with the spatiotemporal constraints of construction worker posture changes, becomes the key issue of this research. Summary of the Invention

[0005] To address the above problems, this invention provides a monitoring-based method for detecting abnormal behavior, mainly comprising:

[0006] The dynamic light intensity changes and the movement trajectory of construction workers in complex nighttime lighting environments were collected. The ambient light scattering component and the reflective clothing reflection component were separated. The ambient light scattering component and the reflective clothing reflection component were filtered and enhanced to generate a spatial light intensity distribution map.

[0007] The high-brightness attenuation area of ​​the reflective vest is determined based on the ambient light scattering component, and the contamination range of the reflective vest is determined based on the reflective vest reflection component. The correspondence between the high-brightness attenuation area and the contamination range of the reflective vest is determined by combining the spatial light intensity distribution map. The location of missing high-brightness features is determined based on the correspondence.

[0008] The gradient variation range of edge light intensity blurring at the location of missing highlight features is detected, and gradient analysis is performed based on the variation range to obtain the distribution characteristics of the edge blurring region;

[0009] Based on the distribution characteristics of the blurred edge region, identify the missing local data points of the light field sampling points, and use the interpolation reconstruction method to generate the filling result of the missing data based on the local data points;

[0010] Based on the results of filling in the missing data, the transient distortion characteristics caused by the jittering interference of reflective clothing folds are analyzed, the transient distortion distribution is obtained, and the jittering frequency and amplitude range are determined based on the transient distortion distribution;

[0011] Based on the jitter frequency and amplitude range, the density of light field sampling points is dynamically adjusted, and the visible light compensation weights obtained by weighted smoothing and feature extraction of the spatial light intensity distribution map are adaptively allocated to obtain the optimized light field sampling distribution.

[0012] By analyzing the dynamic light intensity changes and the construction worker's movement trajectory, the constraints on the construction worker's posture changes are obtained. Combined with the optimized light field sampling distribution, the correction parameters for the transient distortion of the light field caused by posture changes are obtained.

[0013] Based on the correction parameters for transient distortion of the light field caused by attitude changes and the optimized light field sampling distribution, multi-scale feature fusion processing is performed on the spatial light intensity distribution map to obtain the reflective vest feature distribution map, and the final image recognition result of the reflective vest is output.

[0014] Furthermore, the process involves collecting dynamic light intensity changes and the movement trajectory of construction workers under complex nighttime lighting conditions, separating the ambient light scattering component and the reflective clothing reflection component, filtering and enhancing the ambient light scattering component and the reflective clothing reflection component, and generating a spatial light intensity distribution map, including:

[0015] Light intensity data at the construction site at night is collected, and the time-series coordinates of the construction worker's movement trajectory are obtained using a positioning device. Based on the correspondence between the light intensity data and the movement trajectory, the rate of change of light intensity is determined, and the portion exceeding a preset threshold is classified as the reflection component of the reflective vest, while the portion below the preset threshold is classified as the ambient light scattering component. The ambient light scattering component is processed by median filtering, and the reflective vest's reflection component is processed by Gaussian filtering to generate enhanced scattering and reflection component data. A spatial coordinate grid is established in the construction area, and the combined light intensity value of the enhanced scattering and reflection component data is calculated using a bilinear interpolation algorithm to generate the spatial light intensity distribution map.

[0016] Furthermore, the process of determining the high-brightness attenuation area of ​​the reflective vest based on the ambient light scattering component, determining the contamination range of the reflective vest based on the reflective vest reflection component, determining the correspondence between the high-brightness attenuation area and the contamination range of the reflective vest using a spatial light intensity distribution map, and determining the location of missing high-brightness features based on the correspondence includes:

[0017] Based on the ambient light scattering component, sampling points with scattered light intensity below a preset threshold are identified, the distribution density is calculated, and areas with density exceeding the threshold are marked as the high-brightness attenuation region. Based on the boundary coordinates of the high-brightness attenuation region, the reflective component data of the reflective clothing is extracted, the reflection intensity difference is calculated, the contamination points are determined, and the contamination range is summarized. Through spatial superposition of the high-brightness attenuation region and the contamination range, the location of the missing high-brightness feature is determined.

[0018] Furthermore, the gradient change range of the edge light intensity blurring at the detected bright feature missing location, and the distribution characteristics of the edge blurring region obtained by gradient analysis based on the change range, include:

[0019] Edge sampling points at the locations where the highlighted features are missing are detected. The light intensity difference between the points and the sampling points in the eight adjacent directions is calculated, and the maximum gradient value is determined as the gradient feature value. Points below a preset threshold are marked as blurred edge points. Based on the gradient feature values ​​of the blurred edge points, the Sobel operator is used to calculate the horizontal and vertical gradient components to determine the edge gradient intensity and the width of the blurred region. Based on the edge gradient intensity, the ratio of the light intensity sequence difference to the width of the blurred region is calculated to generate the distribution characteristics of the blurred edge region.

[0020] Furthermore, the step of identifying missing local data points of the light field sampling points based on the distribution characteristics of the blurred edge region, and generating a filling result for the missing data using an interpolation reconstruction method based on the local data points, includes:

[0021] Based on the distribution characteristics of the blurred edge region, the light field sampling points are scanned, and sampling points with light intensity values ​​lower than the average of the surrounding non-blurred regions are marked as missing sampling points. Based on the coordinates of the missing sampling points, surrounding valid sampling points are searched, and a bicubic interpolation method is used to calculate the interpolated light intensity of the missing sampling points based on the light intensity values ​​and distance weights of the valid sampling points, thereby generating the filling result.

[0022] Furthermore, based on the filling results of missing data, the transient distortion characteristics caused by the jitter interference of reflective clothing folds are analyzed to obtain the transient distortion distribution. The jitter frequency and amplitude range are then determined based on the transient distortion distribution, including:

[0023] Based on the filling results, the light intensity change data of the folded region is extracted, the coordinate difference between adjacent time moments and the light intensity fluctuation amplitude are calculated, the time correlation is calculated, and a transient distortion feature set is formed. Based on the transient distortion feature set, the distortion degree value is calculated, the sampling points exceeding the threshold are marked, and a transient distortion distribution map is generated through neighborhood connectivity. Based on the transient distortion distribution map, a Kalman filter and Fourier transform are used to determine the jitter frequency and amplitude range.

[0024] Furthermore, the step of dynamically adjusting the light field sampling point density based on the jitter frequency and amplitude range, weighting and smoothing the spatial light intensity distribution map, extracting visible light compensation weights, and adaptively allocating them to obtain an optimized light field sampling distribution includes:

[0025] Based on the jitter frequency and amplitude range, the sampling demand index for each spatial region is determined by multiplying the ratio of the jitter frequency to the preset reference frequency by the ratio of the amplitude range to the preset reference amplitude. The sampling point density is adjusted according to the sampling demand index, increasing the density of regions with a sampling demand index greater than 1. Light intensity data from the spatial light intensity distribution map is collected based on the adjusted sampling point density. Gaussian weighted smoothing is used to calculate the ratio of the average light intensity of each region to the overall average light intensity, determining the visible light compensation weight. Sampling points are allocated according to the visible light compensation weight, adjusting the number of sampling points in each region to generate the optimized light field sampling distribution.

[0026] Furthermore, by analyzing the dynamic light intensity changes and the construction worker's movement trajectory to obtain the constraints on the construction worker's posture changes, and combining this with the optimized light field sampling distribution, correction parameters for the transient distortion of the light field caused by posture changes are obtained, including:

[0027] Based on the dynamic light intensity change and the construction worker's movement trajectory, the angle between the movement direction and the angle between the two points is calculated, the inflection point exceeding the threshold is marked, the light intensity difference is extracted, and the posture type is determined. Combined with the optimized light field sampling distribution, a four-dimensional spatiotemporal data point set is constructed, the transformation matrix is ​​calculated using the least squares method, the translation component and the scaling and rotation component are extracted, and the correction parameters are generated.

[0028] Furthermore, based on the correction parameters for transient distortion of the light field caused by attitude changes and the optimized light field sampling distribution, multi-scale feature fusion processing is performed on the spatial light intensity distribution map to obtain the reflective vest feature distribution map, and the final image recognition result of the reflective vest is output, including:

[0029] Based on the correction parameters, the light intensity values ​​and coordinates of the spatial light intensity distribution map are compensated, and the light intensity data is decomposed into multi-scale levels using wavelet transform to determine the fusion weights and generate the reflective clothing feature distribution map. Based on the feature intensity of the reflective clothing feature distribution map, the regions exceeding the threshold are marked as reflective clothing candidate regions, and the contours and positions are determined through connected component analysis to output the image recognition results.

[0030] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects:

[0031] This invention discloses a monitoring-based method for detecting abnormal behavior. By collecting dynamic nighttime light intensity changes and the movement trajectory of construction workers, the method separates the ambient light scattering and reflective clothing reflection components, and performs filtering and enhancement processing to obtain a spatial light intensity distribution map. Based on the high-brightness attenuation region and the contamination range, the method determines the location of missing features, analyzes the features of blurred edge regions, and fills in the missing data. By analyzing the transient distortion caused by edge jitter interference, the method dynamically adjusts the light field sampling point density, combines the constraints of the construction worker's posture changes, and uses spatiotemporal joint modeling technology to obtain correction parameters. Finally, multi-scale feature fusion processing is performed on the spatial light intensity distribution map to obtain the reflective clothing feature distribution map, and the final recognition result is output. This invention effectively solves the problem of reflective clothing recognition under complex lighting conditions, improving recognition accuracy and robustness. Attached Figure Description

[0032] Figure 1 This is a flowchart of an abnormal behavior detection method based on monitoring according to the present invention.

[0033] Figure 2 This is a schematic diagram of an abnormal behavior detection method based on monitoring according to the present invention. Detailed Implementation

[0034] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this specification, and not all embodiments. Based on the embodiments in this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this specification.

[0035] like Figure 1-2 This embodiment of an abnormal behavior detection method based on monitoring may specifically include:

[0036] S101. Collect dynamic light intensity changes and construction worker movement trajectories under complex nighttime lighting conditions, separate the ambient light scattering component and the reflective clothing reflection component, filter and enhance the ambient light scattering component and the reflective clothing reflection component, and generate a spatial light intensity distribution map.

[0037] Real-time light intensity data and worker location coordinates are collected at the nighttime construction site. A photoelectric sensor array records light intensity values ​​at different locations, while a positioning device acquires worker movement trajectory data. Based on the time series of trajectory points and the corresponding light intensity changes, a raw light intensity dataset containing spatiotemporal information is obtained. For the light intensity value of each sampling point in the raw light intensity dataset, its component type is determined based on the rate of change of light intensity. When the rate of change of light intensity between adjacent sampling points exceeds a preset threshold, that portion of the light intensity is classified as the reflective component, and the portion with a rate of change below the preset threshold is classified as the ambient light scattering component, thus achieving separation of the two components. Median filtering is used to remove noise interference from the separated ambient light scattering component, Gaussian filtering is used to smooth the reflective component, histogram equalization is used to improve contrast of the scattering component, and peak detection is used to amplify the characteristic signal of the reflection component, resulting in enhanced scattering and reflection component data. Based on the enhanced scattering and reflection component data, a spatial coordinate grid is established in the construction area. The scattering and reflection component values ​​of each grid point are calculated using a bilinear interpolation algorithm. By superimposing the two component values, the comprehensive light intensity value of each grid point is obtained, generating a spatial light intensity distribution map covering the entire construction area.

[0038] Specifically, the lighting environment at night construction sites is extremely complex, containing both stable ambient light from fixed light sources and dynamic reflected light from workers moving in reflective clothing. The arrangement of the photoelectric sensor array needs to fully consider the spatial characteristics of the construction area, typically employing a grid layout. The sensor spacing is determined based on the size of the construction site, generally between 5 and 10 meters. Each sensor node records not only the light intensity value but also the acquisition timestamp, forming time-series data. The worker's positioning device can utilize ultra-wideband positioning technology or the BeiDou Navigation Satellite System to acquire the worker's three-dimensional coordinates in real time. This synchronous acquisition mechanism ensures the spatiotemporal correspondence between changes in light intensity and changes in position.

[0039] In one possible implementation, the original light intensity dataset contains rich information dimensions. Each data point includes attributes such as timestamp, spatial coordinates, light intensity value, and sensor number. The rate of change of light intensity is calculated using the difference method between adjacent time points, that is, the difference between the light intensity value at the current moment and the light intensity value at the previous moment divided by the time interval. The ambient light scattering component usually exhibits a slow change characteristic with a small rate of change, while the reflective component of reflective clothing, due to the high reflectivity of the reflective material, will produce instantaneous light intensity abrupt changes. The determination of the preset threshold needs to be calibrated according to the actual construction environment. By collecting background light intensity change data when there is no human activity, the distribution of its rate of change is statistically analyzed, and the upper limit of the distribution is selected as the separation threshold.

[0040] It's important to note that median filtering and Gaussian filtering each have their advantages when processing different types of light intensity components. Median filtering is particularly effective at removing impulse noise from ambient light scattering components. It replaces the current value with the median value within a sliding window, preserving the edge features of light intensity changes while eliminating outliers. Gaussian filtering is suitable for processing the reflection component of reflective clothing. It uses a Gaussian kernel function to perform a weighted average of the light intensity values ​​in the neighborhood, making the distribution of reflected light smoother and more continuous. Histogram equalization adjusts the distribution of light intensity values, making the scattering component data, which was originally concentrated in a certain interval, more evenly distributed, thus enhancing the contrast under different lighting conditions. Peak detection algorithms identify local maxima in the reflection component. By amplifying the amplitude of these feature points, the location information of the reflective clothing becomes more apparent.

[0041] Specifically, the generation of the spatial light intensity distribution map involves complex data fusion. The construction area is divided into a regular grid, with each grid point representing a spatial location. A bilinear interpolation algorithm uses the light intensity values ​​of four known data points surrounding a grid point and calculates an estimated value for that point through linear weighting. For the scattering and reflection components, interpolation calculations are performed separately, and then superimposed at each grid point. This superposition is not a simple addition, but considers the physical characteristics of the two components: the scattering component serves as the base background light intensity, and the reflection component is superimposed as a dynamic increment. The generated spatial light intensity distribution map not only reflects the overall lighting conditions of the construction site but also highlights the activity areas of the construction workers, providing an intuitive visual basis for nighttime construction safety management.

[0042] S102. Determine the high-brightness attenuation area of ​​the reflective clothing based on the ambient light scattering component, determine the contamination range of the reflective clothing based on the reflective clothing reflection component, determine the correspondence between the high-brightness attenuation area and the contamination range of the reflective clothing by combining the spatial light intensity distribution map, and determine the location of missing high-brightness features based on the correspondence.

[0043] Based on ambient light scattering component data, sampling points with scattered light intensity below a preset scattering intensity threshold are identified. The spatial distribution density of these sampling points is calculated. When the density of low-intensity sampling points in a continuous area exceeds a preset density threshold, the continuous area is marked as a high-brightness attenuation area for the reflective clothing. Based on the boundary coordinates of the high-brightness attenuation area, the reflective component data of the reflective clothing at the corresponding locations is extracted. The difference between the reflective intensity of each sampling point within the area and the normal reflective intensity outside the area is calculated. When the difference exceeds a preset difference threshold, the sampling point is identified as a contamination point. The spatial range of all contamination points is summarized to obtain the contamination range of the reflective clothing. The coordinate information of each sampling point within the contamination range is obtained and spatially superimposed with the high-brightness attenuation area of ​​the reflective clothing. Sampling points that simultaneously meet the criteria of being located within both the high-brightness attenuation area and the contamination range are identified, and the locations of these sampling points are determined as locations with missing high-brightness features.

[0044] Specifically, ambient light scattering component data reflects the background lighting characteristics of the construction site. When the surface of reflective clothing is contaminated or worn, its reflectivity decreases significantly, resulting in a significantly lower scattered light intensity in that area compared to normal levels. The determination of the preset scattering intensity threshold needs to be based on the baseline value of scattered light intensity under standard lighting conditions for clean reflective clothing, typically taking 60% to 70% of the baseline value as the judgment standard. After identifying low-intensity sampling points, the spatial distribution density is calculated using a sliding window method, counting the ratio of the number of low-intensity sampling points to the total number of sampling points within each window.

[0045] In one possible implementation, the continuous region is defined using the eight-neighbor connectivity criterion, meaning that the adjacent points in the top, bottom, left, right, and four diagonal directions of a given sampling point are all included in the connectivity assessment. When multiple low-density sampling points are interconnected through neighborhood relationships to form a patchy distribution, and the density of low-intensity points within this patchy region exceeds 80%, the entire continuous region is marked as a reflective clothing high-brightness attenuation region. This density-based region identification method can effectively distinguish between local contamination and large-area attenuation, avoiding misclassification of individual abnormal points as attenuation regions.

[0046] It should be noted that there is a close spatial correspondence between the reflective component data of the reflective clothing and the high-brightness attenuation area. The process of extracting the boundary coordinates of the attenuation area involves edge detection calculations, determining the boundary location by identifying abrupt changes in light intensity inside and outside the area. The determination of the normal reflective intensity outside the area uses statistical methods, selecting sampling points within a certain range from the boundary of the attenuation area and calculating the median of their reflective intensity as a reference value. The difference calculation considers not only absolute numerical differences but also relative rates of change; when the reflective intensity decreases by more than 40% of the normal value, it is considered significantly contaminated.

[0047] Specifically, the process of summarizing the spatial extent of contamination points needs to consider the diffusion effect of contamination. The impact of a single contamination point is not limited to its own location but also spreads outwards, forming a contamination halo. Through dilation operations in morphological calculations, each contamination point expands outward by one or two sampling units, forming a contamination-affected area. After multiple contamination-affected areas overlap, they form the complete contamination range of the reflective clothing. This processing method can more accurately reflect the actual impact range of contamination on reflective performance. Spatial superposition operations are a key step in determining the locations of missing high-brightness features. By establishing a unified spatial coordinate system, the high-brightness attenuation area and the contamination range are mapped to the same coordinate frame. For each sampling point, it is determined whether it simultaneously meets two conditions: it is located within the high-brightness attenuation area and within the contamination range. Sampling points that meet both conditions indicate that there is both physical attenuation of the reflective material and surface contamination at that location, resulting in a severe decrease in reflective performance. The set of these locations constitutes the dataset of missing high-brightness feature locations, providing precise location information for subsequent maintenance and replacement of reflective clothing.

[0048] S103. Detect the gradient change range of the edge light intensity blur at the location where the highlight feature is missing, and perform gradient analysis based on the change range to obtain the distribution characteristics of the edge blur region.

[0049] For each edge sampling point at a location lacking highlight features, the light intensity values ​​of that point and its eight adjacent sampling points in the eight directions are obtained. The light intensity difference between the center point and each adjacent point is calculated and divided by the distance between the sampling points to obtain the light intensity gradient values ​​in the eight directions. The maximum gradient value is selected as the gradient feature value of the edge point. When the gradient feature value is lower than a preset gradient threshold, it is determined to be a blurred edge point. For the gradient feature values ​​of all blurred edge points, the Sobel operator is used to calculate the gradient components in the horizontal and vertical directions respectively. The edge gradient intensity is obtained by taking the square root of the sum of the squares of the gradient components in the two directions. The continuous distribution length of adjacent blurred edge points is counted as the width of the blurred region. Based on the edge gradient intensity and the width of the blurred region, the light intensity sequence from bright to dark in the blurred region is extracted. The light intensity attenuation rate is obtained by calculating the ratio of the difference between the first and last light intensity values ​​of the sequence to the width of the blurred region. The edge gradient intensity, blurred region width, and light intensity attenuation rate of each blurred region are integrated to form the distribution characteristics of the blurred edge region.

[0050] Specifically, edge detection at locations lacking highlight features involves fine-grained analysis of light intensity variations. Edge sampling points are selected based on the boundaries between contaminated and normal areas of the reflective clothing, as light intensity variations at these locations best reflect the degree of reflective degradation. The eight-directional sampling design originates from the eight-neighborhood concept in digital image processing, including four positive directions (up, down, left, right) and four diagonal directions. This omnidirectional sampling can capture light intensity variation characteristics in all directions.

[0051] In one embodiment, the light intensity gradient value is calculated using a differential method. Assuming the light intensity value at the center sampling point is 100 units, the light intensity value at the adjacent point directly above it is 85 units, and the distance between the two points is 1 sampling unit, then the gradient value in that direction is 15. By calculating the gradient values ​​in eight directions, the maximum value is selected as the gradient feature value of the edge point. A preset gradient threshold is typically set to 30% of the normal gradient value at the edge of reflective clothing. When the detected gradient feature value is lower than this threshold, it indicates that the edge has become blurred, and the reflective performance has significantly decreased.

[0052] It should be noted that the Sobel operator has unique advantages in edge detection. This operator calculates the gradient components in the horizontal and vertical directions using a 3×3 convolution kernel. The horizontal convolution kernel is sensitive to vertical edges, and the vertical convolution kernel is sensitive to horizontal edges. The gradient components in the two directions are synthesized through square root operations, and the resulting edge gradient intensity contains both the direction information of the edge and reflects its intensity characteristics. This calculation method can effectively suppress noise interference and improve the accuracy of edge detection. The determination of the blurred region width is based on the principle of continuity. When multiple adjacent sampling points are identified as blurred edge points, these points form a continuous blurred band. By statistically analyzing the span of this blurred band in the direction perpendicular to the edge, the width of the blurred region can be obtained. A wider blurred region usually indicates a more severe degree of contamination of the reflective material or deeper wear of the reflective coating.

[0053] Specifically, calculating the light intensity attenuation rate requires extracting the complete light intensity change sequence within the blurred region. Starting from the bright side boundary of the blurred region, light intensity values ​​are recorded point by point along the direction of decreasing light intensity until the dark side boundary. The difference between the first and last light intensity values ​​of the sequence reflects the total light intensity attenuation, which, when divided by the width of the blurred region, yields the attenuation rate per unit distance. A high attenuation rate indicates a sharp decrease in reflective performance over a short distance, typically occurring at locations where reflective materials are severely damaged. The distribution characteristics of the blurred edge region integrate three key parameters: edge gradient intensity reflects the clarity of the edge, the width of the blurred region characterizes the spatial extent of contamination, and the light intensity attenuation rate quantifies the rate of decline in reflective performance. These three parameters are interrelated and collectively describe the severity of the loss of high-brightness features in reflective clothing. By establishing this multi-dimensional feature description, the actual protective effect of reflective clothing can be comprehensively evaluated, providing accurate quantitative data for construction safety management.

[0054] S104. Based on the distribution characteristics of the blurred edge region, identify the missing local data points of the light field sampling points, and use the interpolation reconstruction method to generate the filling result of the missing data based on the local data points.

[0055] Based on the distribution characteristics of the blurred edge region, all light field sampling points within the blurred region are scanned. When the light intensity value of a sampling point is lower than a preset proportional threshold of the average light intensity value of the surrounding unblurred region, the light intensity data of that sampling point is determined to be missing, and it is marked as a missing sampling point. All missing sampling points are summarized to obtain a set of missing light field points. For each missing sampling point in the set of missing light field points, its spatial coordinate position is determined, and all valid sampling points within a preset radius around the missing point are searched. The missing type is determined based on the number of valid sampling points found. If the number of valid sampling points exceeds a preset threshold, it is marked as a local data point. Based on the spatial distribution of the local data points, a bicubic interpolation method is used to calculate the filling value. The sixteen nearest valid sampling points around the local data point are selected, and the light intensity values ​​of the valid points are weighted and summed using the reciprocal of the distance between these valid points and the local data point as weight coefficients to obtain the interpolated light intensity of the local data point, generating the filling result of the missing data.

[0056] Specifically, the distribution characteristics of blurred edge areas provide crucial information for identifying missing light field data. When the surface of reflective clothing becomes dirty or worn, it not only reduces reflectivity but also prevents photoelectric sensors from acquiring effective light intensity data at certain locations. This data loss is often concentrated within blurred areas because the light intensity signal at these locations is too weak, falling below the sensor's detection threshold. The preset proportional threshold is usually based on experimental calibration, typically set at 20% to 30% of the average light intensity of the surrounding normal area as the judgment standard.

[0057] In one embodiment, the scanning process of the light field sampling points is performed row by row and column by column. For each sampling point located within a blurred region, the system extracts the light intensity values ​​of all sampling points in its eight neighboring unblurred regions and calculates the arithmetic mean of these values ​​as a reference benchmark. When the light intensity value of the target sampling point is lower than this benchmark value multiplied by a preset ratio, the system marks it as a missing sampling point. This relative comparison method can adapt to changes in light intensity under different lighting conditions, improving the accuracy of missing point detection.

[0058] It's important to note that determining local data points involves spatial neighborhood analysis. The preset radius is typically set to 3 to 5 sampling units. This range needs to be large enough to include sufficient valid sampling points, but not so large as to introduce irrelevant data from distant locations. Valid sampling points are those with normal light intensity values ​​located in non-blurred areas. When more than 8 valid sampling points are found, it indicates that there is sufficient data around the missing point to support interpolation reconstruction, and therefore it is marked as a local data point. Conversely, if there are too few valid sampling points, the data loss at that location is too severe to be suitable for interpolation.

[0059] Specifically, the bicubic interpolation method has unique advantages in light field data reconstruction. Unlike simple linear interpolation, this method considers the influence of data points over a larger range. A 4×4 grid structure is formed by selecting the sixteen nearest valid sampling points; this layout can capture the variation trend of light intensity in two-dimensional space. The calculation method using the reciprocal of distance as a weighting coefficient reflects the principle of spatial correlation: points that are closer have a greater influence on the interpolation result.

[0060] For example, a point at a distance of 1 unit has a weight of 1, a point at a distance of 2 units has a weight of 0.5, and so on. The weighted summation process is essentially a comprehensive evaluation of the surrounding effective light intensity values. The light intensity value of each effective sampling point is multiplied by its corresponding weight coefficient, all products are added together, and then divided by the sum of the weight coefficients to obtain the final interpolated light intensity. This calculation method ensures the smoothness and continuity of the interpolation results, avoiding data abrupt changes. The generated filling result not only restores the light intensity information of the missing locations but also maintains consistency with the surrounding data, making the entire light field distribution more complete. Through this refined data reconstruction, even when some areas of the reflective clothing are severely damaged, the system can still accurately assess the overall reflective performance distribution.

[0061] S105. Based on the results of filling in the missing data, analyze the transient distortion characteristics caused by the jittering interference of the reflective folds, obtain the transient distortion distribution, and determine the jittering frequency and amplitude range based on the transient distortion distribution.

[0062] Based on the missing data filling results, the areas where wrinkles occur at the edge of reflective clothing are identified as wrinkled areas. Light intensity variation data of these wrinkled areas are extracted over a continuous time series. The spatial coordinate difference between adjacent sampling points at the same time is calculated to obtain the jitter displacement. The difference between the maximum and minimum values ​​of the light intensity sequence at that sampling point is calculated to obtain the light intensity fluctuation amplitude. The correlation coefficient of light intensity values ​​at different time intervals is calculated using an autocorrelation function to obtain the temporal correlation. The displacement, fluctuation amplitude, and correlation coefficient are summarized to form a transient distortion feature set. Based on the three feature parameters of each sampling point in the transient distortion feature set, a weighted sum is calculated as the distortion level value for that point. When the distortion level value exceeds a preset distortion threshold, the sampling point is marked as a distorted point. Adjacent distorted points are connected using a neighborhood connectivity method to obtain a transient distortion distribution map. For the sequence of distorted points in the transient distortion distribution map, a Kalman filter is used to establish a state-space model to track the trajectory of the distorted point positions over time. The dominant frequency component of the trajectory data is identified using Fourier transform to determine the jitter frequency, and the maximum distance of the trajectory from the center position is statistically analyzed to determine the jitter amplitude range.

[0063] Specifically, the identification of folded areas in reflective clothing is based on the spatial distribution characteristics of light intensity data after filling. Folds typically appear in easily deformable locations such as the cuffs, hems, and side seams of reflective clothing, where they periodically fold and unfold during worker activity. By analyzing areas of abrupt changes in light intensity gradient in the filling results, the location of folds can be accurately pinpointed. When the light intensity difference between adjacent sampling points exceeds 50% of the average light intensity, and this difference exhibits a banded distribution, it can be identified as a folded area.

[0064] In one embodiment, the calculation of the jitter displacement involves precise spatial positioning. Each sampling point has definite coordinates in three-dimensional space, and the displacement vector of that point can be obtained by comparing the coordinate changes of the same sampling point at adjacent time points.

[0065] For example, if the coordinates of a sampling point at time t are (100, 200, 50), and at time t+1 they become (102, 198, 51), then the jitter displacement of that point is the square root of the sum of the squares of the displacements in each direction, approximately 3.3 units. The amplitude of light intensity fluctuation reflects the dynamic range of reflective performance. By statistically analyzing the maximum and minimum light intensity values ​​over a period of time, the degree of influence of edge jitter on the reflective effect can be quantified.

[0066] It is important to note that the autocorrelation function plays a crucial role in analyzing temporal correlation. This function reveals the periodicity of the signal by calculating the similarity of light intensity sequences at different time delays. Specifically, the original light intensity sequence is multiplied point-by-point with its own translated sequence and summed to obtain the correlation coefficients corresponding to different delay times. When the correlation coefficient peaks at a specific delay time, it indicates that the signal exhibits periodicity within that time interval. This periodicity is a typical characteristic of jitter. The quantitative assessment of distortion employs a multi-parameter weighted fusion method. Three characteristic parameters reflect transient distortion from different dimensions: jitter displacement reflects spatial variation, light intensity fluctuation amplitude reflects intensity variation, and temporal correlation characterizes the variation pattern. In the weighted sum calculation, the weight of each parameter is determined according to its importance to the overall distortion; typically, displacement has a weight of 0.4, fluctuation amplitude has a weight of 0.4, and correlation has a weight of 0.2. When the weighted sum exceeds a preset threshold, that location is marked as a distortion point. Through a neighborhood connectivity algorithm, spatially adjacent distortion points are connected into regions, forming a complete transient distortion distribution map.

[0067] Specifically, the application of the Kalman filter in time-series tracking is based on the establishment of a state-space model. This filter uses the position and velocity of the distortion point as state variables, achieving accurate tracking of the motion trajectory through two steps: prediction and update. The prediction step estimates the current position based on the state at the previous moment, while the update step corrects the prediction result using actual observations. This recursive processing method effectively suppresses measurement noise, resulting in a smooth motion trajectory. Fourier transform converts the time-domain trajectory data to the frequency domain for analysis. By identifying the dominant frequency components in the spectrum, the fundamental frequency of the edge jitter can be determined. The jitter amplitude range is obtained by statistically analyzing the maximum distance of the trajectory point from the average position. These quantitative parameters provide a scientific basis for evaluating the actual protective effect of reflective vests under dynamic conditions.

[0068] S106. Based on the jitter frequency and amplitude range, dynamically adjust the density of light field sampling points, perform weighted smoothing and feature extraction on the spatial light intensity distribution map to obtain visible light compensation weights, and adaptively allocate them to obtain the optimized light field sampling distribution.

[0069] Based on the jitter frequency and amplitude range, a sampling demand index is calculated for each spatial region. The sampling demand index equals the ratio of the jitter frequency to the preset reference frequency multiplied by the ratio of the amplitude range to the preset reference amplitude. When the sampling demand index is greater than 1, the sampling point density of that region is multiplied by the sampling demand index to obtain the enhanced sampling density; otherwise, the original density remains unchanged, resulting in an adjusted sampling point density distribution. Based on the adjusted sampling point density distribution, spatial light intensity data is acquired at sampling points of the corresponding density. Gaussian weighted smoothing is used to process the light intensity data, and the ratio of the mean light intensity of each region after smoothing to the overall mean light intensity is calculated. When the ratio is less than 1, the degree of insufficient illumination is obtained by subtracting the ratio from 1. The degree of insufficient illumination is then normalized to the interval between 0 and 1 as the visible light compensation weight value. Based on the visible light compensation weight value, calculate the proportion of the weight value of each region to the total weight of all regions, and allocate the preset total number of sampling points to each region according to this proportion. If the number of sampling points allocated to a certain region is less than the minimum number of sampling points required for the area of ​​that region, then sampling points are allocated from the region with the smallest weight value in sequence to supplement it. After the allocation is completed, the optimized light field sampling distribution is obtained.

[0070] Specifically, the calculation of the sampling demand index reflects the design concept of dynamic response. The jitter frequency reflects the speed of movement of the reflective fabric folds, while the amplitude range characterizes the intensity of the movement. The preset reference frequency is typically set to the frequency of clothing movement during normal walking, approximately 2 Hz, while the reference amplitude corresponds to the displacement range of slight shaking, approximately 5 cm. When the actual detected jitter frequency is 4 Hz and the amplitude is 10 cm, the frequency ratio is 2, the amplitude ratio is also 2, and the sampling demand index is 4. This means that the dynamic changes in this area are drastic, requiring more frequent sampling to accurately capture changes in light intensity.

[0071] In one embodiment, the adjustment of the sampling point density follows an adaptive principle. The original sampling density might be 100 sampling points per square meter, and when the sampling demand index is 4, the adjusted density becomes 400 sampling points per square meter. This increase in density is not uniformly distributed, but concentrated in the folded areas where the jitter is most severe. Conversely, for the relatively static back area, since the sampling demand index is less than 1, the sampling density remains unchanged, avoiding waste of computational resources.

[0072] It should be noted that Gaussian weighted smoothing has unique advantages in processing light intensity data. This method uses a Gaussian kernel function to weight the light intensity values ​​within the neighborhood, with the weights decreasing in a Gaussian distribution with distance. The center point has the largest weight, and the weights of surrounding points gradually decrease. This processing method preserves local characteristics while eliminating random noise. The smoothed light intensity data better reflects the overall illumination level of the area, providing a reliable basis for subsequent compensation weight calculations. The quantitative assessment of insufficient illumination is based on the principle of relative comparison. The overall light intensity mean represents the average lighting level of the construction site, while the light intensity mean of each area reflects the actual local situation.

[0073] For example, when the overall mean is 100 lux and the mean of a certain area is only 60 lux, the ratio is 0.6, and the degree of insufficient light is 0.4. This value directly reflects the degree to which that area needs additional compensation. Normalization ensures that the compensation weights of all areas are on a uniform scale, which facilitates subsequent resource allocation.

[0074] Specifically, the adaptive allocation of sampling resources involves a complex optimization process. The preset total number of sampling points is usually determined based on hardware capabilities and real-time requirements, such as 10,000 sampling points. When allocating according to weight ratios, if a region with a weight value of 0.4 accounts for 20% of the total weight, it should be allocated 2,000 sampling points. However, this region may require at least 2,500 sampling points to meet the minimum sampling requirement. In this case, the system will allocate 500 sampling points from the region with the lowest weight. This dynamic allocation mechanism ensures the sampling accuracy of critical areas while maintaining the overall resource balance. The optimized light field sampling distribution exhibits obvious non-uniform characteristics. High-dynamic areas and insufficiently illuminated areas receive more sampling resources, while stable and sufficiently illuminated areas receive less sampling. This adaptive distribution not only improves the monitoring accuracy of critical areas but also optimizes the overall computational efficiency, achieving a reasonable allocation of resources.

[0075] S107. By analyzing the dynamic light intensity changes and the construction worker's movement trajectory, the constraints on the construction worker's posture changes are obtained. Combined with the optimized light field sampling distribution, the correction parameters for the transient distortion of the light field caused by posture changes are obtained.

[0076] By analyzing the temporal correspondence between dynamic light intensity changes and the construction worker's movement trajectory, the angle between the movement direction vectors at adjacent moments is calculated. When the angle exceeds a preset angle threshold, it is marked as a trajectory turning point. The light intensity difference before and after the turning point is extracted as abrupt change features. The posture type is determined based on the direction of the abrupt change: a vertical abrupt change is determined as bending over, and a horizontal abrupt change is determined as turning around, thus obtaining posture change constraints. Combining the posture change constraints and the optimized light field sampling distribution, a four-dimensional spatiotemporal data point set is constructed by combining the spatial coordinates and acquisition time of each sampling point. The transformation matrix of the light field distribution before and after the posture change is calculated using the least squares method. The translation component is extracted from the transformation matrix as the posture offset, and the combined value of the scaling and rotation components is extracted as the light field distortion coefficient. Based on the time series of the posture offset and the light field distortion coefficient, the time difference between the construction worker's posture change moment and the peak moment of the light field response is calculated. By performing cross-correlation operations on the posture signal and the light field signal, the time offset corresponding to the maximum correlation value is found and determined as the time delay factor. The three parameters are combined to form a transient distortion correction parameter set.

[0077] Specifically, the identification of turning points in the movement trajectory is based on the principle of vector angle calculation. The construction worker's movement trajectory can be represented as a series of location points over time, with directional vectors formed between adjacent location points. When the construction worker walks in a straight line, the angle between adjacent vectors is close to zero degrees, while when turning, bending over, or other posture changes occur, the angle increases significantly. The preset angle threshold is usually set at 30 degrees, which can effectively distinguish between normal path deviations and genuine posture changes.

[0078] In one embodiment, the extraction of light intensity abrupt changes involves multi-dimensional analysis. Vertical light intensity abrupt changes are mainly caused by bending over, as the reflection angle of the reflective vest changes, resulting in a sharp drop in the light intensity reflected to the sensor. Horizontal abrupt changes are mostly caused by turning over; different parts of the reflective vest have different reflective properties, and the reflective surface presented to the sensor changes during turning. A change in intensity exceeding 40% of the average light intensity can be considered a valid posture change signal. This directionality-based posture recognition method can accurately distinguish between different types of movements.

[0079] It is important to note that the construction of a four-dimensional spatiotemporal data point set is fundamental to achieving accurate modeling. Each data point contains three spatial coordinates and one temporal coordinate, forming a complete spatiotemporal description. This representation not only records the location information of the sampling points but also preserves the temporal evolution characteristics. During the worker's movement, the same spatial location may have different light intensity values ​​at different times; the four-dimensional representation can fully capture this dynamic change. The least squares method plays a crucial role in calculating the transformation matrix. This method finds the optimal transformation parameters by minimizing the sum of squared errors between corresponding points before and after the attitude change. The transformation matrix contains multiple transformation components, including translation, rotation, and scaling. The translation component directly reflects the displacement of the worker's center of gravity, i.e., the attitude offset. The combined effect of the rotation and scaling components is reflected in the degree of light field distortion; the light field distortion coefficient is obtained by calculating the norms of these components.

[0080] Specifically, cross-correlation calculations have a unique advantage in determining the time delay factor. This calculation calculates the similarity between the two signals at different time offsets by sliding the attitude change signal sequence and the optical field response signal sequence relative to each other on the time axis. When a construction worker completes a bending motion, the optical field response is not immediate but delayed. This delay stems from the light propagation time and the sensor response time. The cross-correlation function reaches its maximum value at a specific time offset, and this offset is the time delay factor. The three components of the correction parameter set are interrelated and together describe the mechanism by which attitude changes affect the optical field. The attitude offset quantifies the change in physical position, the optical field distortion coefficient reflects the degree of deformation in the light intensity distribution, and the time delay factor characterizes the dynamic response characteristics of the system. Accurate acquisition of these parameters provides a quantitative basis for subsequent transient distortion compensation, ensuring that the protective performance of reflective vests can still be accurately evaluated in dynamic environments.

[0081] S108. Based on the correction parameters for transient distortion of the light field due to attitude change and the optimized light field sampling distribution, perform multi-scale feature fusion processing on the spatial light intensity distribution map to obtain the reflective clothing feature distribution map and output the final image recognition result of the reflective clothing.

[0082] Based on correction parameters composed of attitude offset, light field distortion coefficient, and time delay factor, compensation calculations are performed on each sampling point in the spatial light intensity distribution map. The original light intensity value is multiplied by the reciprocal of the distortion coefficient for intensity compensation. Spatial coordinates are adjusted according to the offset, and time alignment is performed according to the delay factor to obtain the corrected light intensity distribution map. Combining the corrected light intensity distribution map and the optimized light field sampling distribution, wavelet transform is used to decompose the light intensity data into three scale levels: coarse scale for extracting the overall contour, medium scale for extracting regional features, and fine scale for extracting edge details. The fusion weight is determined based on the information content ratio of each scale, and a reflective vest feature distribution map is generated by layer-by-layer stacking. Based on the feature intensity values ​​at each position in the reflective vest feature distribution map, a reflective vest recognition threshold is set. Regions with feature intensity exceeding the threshold are marked as reflective vest candidate regions. Adjacent candidate regions are merged through connected component analysis to determine the complete contour and center position of the reflective vest. The image recognition result, including position coordinates, contour boundaries, and average reflection intensity, is output.

[0083] Specifically, the compensation calculation for the correction parameters embodies a systematic error correction approach. The light field distortion coefficient reflects the degree of distortion in the light intensity distribution caused by attitude changes. When the coefficient is 0.8, it indicates that the actual light intensity is only 80% of the ideal value. By multiplying by the reciprocal of the distortion coefficient, 1.25, the original light intensity level can be restored. The spatial coordinate adjustment of the attitude offset involves coordinate transformation. For example, if a construction worker bends over, causing the reflective vest to shift downwards by 20 centimeters, the vertical coordinates of all relevant sampling points need to be adjusted upwards accordingly to ensure that the light intensity data corresponds to the actual position. The alignment processing of the time delay factor is more refined. When the delay is 0.1 seconds, the timestamp of the light intensity data needs to be shifted forward by the corresponding time to achieve spatiotemporal synchronization.

[0084] In one embodiment, the multi-scale decomposition of wavelet transform possesses unique time-frequency localization characteristics. This transform decomposes the original signal into different frequency components through a series of filtering and downsampling operations. The coarse-scale level preserves low-frequency information, reflecting the overall shape and approximate location of the reflective garment; this information corresponds to the approximation coefficients of the wavelet decomposition. The meso-scale level contains mid-frequency components, capturing the main structural features of the reflective garment, such as the boundaries between the sleeves and the body. The fine-scale level preserves high-frequency details, including the edges of reflective stripes and the fine textures created by wrinkles.

[0085] It should be noted that the determination of fusion weights is based on the assessment of information content at each scale. Information content is derived by calculating the energy proportion of the coefficients at each scale; higher energy indicates more effective information contained at that scale. Typically, the weight is set to 0.5 for the coarse scale, 0.3 for the mesoscale, and 0.2 for the fine scale. This allocation ensures the accuracy of the overall outline without sacrificing detailed features. During the layer-by-layer stacking process, the coarse-scale features are first used as the foundation, followed by the sequential stacking of mesoscale and fine-scale features. Each stacking is weighted and summed according to its respective weight. The calculation of feature intensity values ​​integrates information from multiple dimensions. In the reflective clothing feature distribution map, the feature intensity at each location is obtained by weighted summing of the response values ​​at each scale. High feature intensities typically appear in areas with concentrated reflective material, which exhibit significant reflective properties at multiple scales. The reflective clothing recognition threshold is set using an adaptive method, selecting a boundary value that effectively distinguishes the reflective clothing from the background by analyzing the histogram distribution of feature intensities.

[0086] Specifically, connected component analysis plays a crucial role in determining the complete outline of reflective vests. Based on the eight-neighborhood connectivity criterion, this analysis groups spatially adjacent pixels with feature intensities exceeding a threshold into the same connected component. A single construction worker's reflective vest may be segmented into multiple initial connected components due to occlusion or posture changes. By analyzing the distance and feature similarity between these components, components belonging to the same reflective vest are merged. The final identification result contains rich information: location coordinates are accurate to the centimeter level, the outline boundary is represented by a sequence of polygonal vertices, and the average reflective intensity reflects the overall protective performance of the reflective vest. This comprehensive identification result provides reliable data support for construction safety management.

[0087] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A monitoring-based method for detecting abnormal behavior, characterized in that, The method includes: S10 collects dynamic light intensity changes and the movement trajectory of construction workers under complex nighttime lighting conditions, separates the ambient light scattering component and the reflective clothing reflection component, filters and enhances the ambient light scattering component and the reflective clothing reflection component, and generates a spatial light intensity distribution map; S20 determines the high-brightness attenuation area of ​​the reflective clothing based on the ambient light scattering component, and determines the contamination range of the reflective clothing based on the reflective clothing reflection component. Combining the spatial light intensity distribution map, it determines the correspondence between the high-brightness attenuation area and the contamination range of the reflective clothing, and determines the location of missing high-brightness features based on the correspondence; S30 detects the gradient change range of the edge light intensity blurring at the location of missing high-brightness features, and performs gradient analysis based on the change range to obtain the distribution characteristics of the edge blurring area; S40, based on the distribution characteristics of the edge blurring area, identifies the missing local data points of the light field sampling points, and generates the missing data filling result using the interpolation reconstruction method based on the local data points; S50, based on the missing data filling result, analyzes the transient distortion characteristics caused by the jitter interference of the reflective clothing folds, obtains the transient distortion distribution, and determines the jitter frequency and amplitude range based on the transient distortion distribution. S60 dynamically adjusts the density of light field sampling points based on the jitter frequency and amplitude range, performs weighted smoothing and visible light compensation weights obtained from feature extraction on the spatial light intensity distribution map, and adaptively allocates them to obtain the optimized light field sampling distribution; S70 obtains the worker's posture change constraints by analyzing dynamic light intensity changes and the worker's movement trajectory, and obtains the correction parameters for the transient distortion of the light field caused by posture changes by combining the optimized light field sampling distribution; S80 performs multi-scale feature fusion processing on the spatial light intensity distribution map based on the correction parameters for the transient distortion of the light field caused by posture changes and the optimized light field sampling distribution to obtain the reflective vest feature distribution map, and outputs the final image recognition result of the reflective vest.

2. The abnormal behavior detection method based on monitoring according to claim 1, characterized in that, The process involves collecting dynamic light intensity changes and the movement trajectory of construction workers under complex nighttime lighting conditions, separating the ambient light scattering component and the reflective clothing reflection component, filtering and enhancing these components, and generating a spatial light intensity distribution map, including: Light intensity data at the construction site at night is collected, and the time-series coordinates of the construction worker's movement trajectory are obtained using a positioning device. Based on the correspondence between the light intensity data and the construction worker's movement trajectory, the rate of change of light intensity is determined, and the portion exceeding a preset threshold is classified as the reflective component of the reflective vest, while the portion below the preset threshold is classified as the ambient light scattering component. The ambient light scattering component is processed by median filtering, and the reflective component of the reflective vest is processed by Gaussian filtering to generate enhanced scattering and reflection component data. A spatial coordinate grid is established in the construction area, and the combined light intensity value of the enhanced scattering and reflection component data is calculated using a bilinear interpolation algorithm to generate the spatial light intensity distribution map.

3. The abnormal behavior detection method based on monitoring according to claim 1, characterized in that, The process of determining the high-brightness attenuation area of ​​the reflective vest based on the ambient light scattering component, determining the contamination range of the reflective vest based on the reflective component, determining the correspondence between the high-brightness attenuation area and the contamination range of the reflective vest using a spatial light intensity distribution map, and determining the location of missing high-brightness features based on the correspondence includes: Based on the ambient light scattering component, sampling points with scattered light intensity below a preset threshold are identified, the distribution density is calculated, and areas with density exceeding the threshold are marked as the high-brightness attenuation region. Based on the boundary coordinates of the high-brightness attenuation region, the reflective component data of the reflective clothing is extracted, the reflection intensity difference is calculated, the contamination points are determined, and the contamination range is summarized. Through spatial superposition of the high-brightness attenuation region and the contamination range, the location of the missing high-brightness feature is determined.

4. The abnormal behavior detection method based on monitoring according to claim 1, characterized in that, The gradient change range of the edge light intensity blurring at the detected high-brightness feature missing location is used to perform gradient analysis based on the change range to obtain the distribution characteristics of the edge blurring region, including: Edge sampling points at the locations where the highlighted features are missing are detected. The light intensity difference between the points and the sampling points in the eight adjacent directions is calculated, and the maximum gradient value is determined as the gradient feature value. Points below a preset threshold are marked as blurred edge points. Based on the gradient feature values ​​of the blurred edge points, the Sobel operator is used to calculate the horizontal and vertical gradient components to determine the edge gradient intensity and the width of the blurred region. Based on the edge gradient intensity, the ratio of the light intensity sequence difference to the width of the blurred region is calculated to generate the distribution characteristics of the blurred edge region.

5. The abnormal behavior detection method based on monitoring according to claim 1, characterized in that, The step of identifying missing local data points in the light field sampling points based on the distribution characteristics of the blurred edge region, and generating a filling result for the missing data using an interpolation reconstruction method based on the local data points, includes: Based on the distribution characteristics of the blurred edge region, the light field sampling points are scanned, and sampling points with light intensity values ​​lower than the average of the surrounding non-blurred regions are marked as missing sampling points. Based on the coordinates of the missing sampling points, surrounding valid sampling points are searched, and a bicubic interpolation method is used to calculate the interpolated light intensity of the missing sampling points based on the light intensity values ​​and distance weights of the valid sampling points, thereby generating the filling result.

6. The abnormal behavior detection method based on monitoring according to claim 1, characterized in that, The process involves analyzing the transient distortion characteristics caused by the jitter interference of reflective fabric folds based on the filling results of missing data, obtaining the transient distortion distribution, and determining the jitter frequency and amplitude range based on the transient distortion distribution, including: Based on the filling results, the light intensity change data of the folded region is extracted, the coordinate difference between adjacent time moments and the light intensity fluctuation amplitude are calculated, the time correlation is calculated, and a transient distortion feature set is formed. Based on the transient distortion feature set, the distortion degree value is calculated, the sampling points exceeding the threshold are marked, and a transient distortion distribution map is generated through neighborhood connectivity. Based on the transient distortion distribution map, a Kalman filter and Fourier transform are used to determine the jitter frequency and amplitude range.

7. The abnormal behavior detection method based on monitoring according to claim 1, characterized in that, The process of dynamically adjusting the light field sampling point density based on the jitter frequency and amplitude range, performing weighted smoothing and feature extraction on the spatial light intensity distribution map to obtain visible light compensation weights, and adaptively allocating them to obtain an optimized light field sampling distribution includes: S61 determines the sampling demand index for each spatial region by multiplying the ratio of the jitter frequency to the preset reference frequency by the ratio of the amplitude range to the preset reference amplitude, based on the jitter frequency and amplitude range. S62 adjusts the sampling point density according to the sampling demand index, and increases the density of the area where the sampling demand index is greater than 1; S63 collects light intensity data of the spatial light intensity distribution map according to the adjusted sampling point density, calculates the ratio of the average light intensity of each region to the average light intensity of the whole region through Gaussian weighted smoothing, and determines the visible light compensation weight; allocates sampling points according to the visible light compensation weight, adjusts the number of sampling points in each region, and generates the optimized light field sampling distribution.

8. The abnormal behavior detection method based on monitoring according to claim 1, characterized in that, The process involves analyzing dynamic light intensity changes and the worker's movement trajectory to obtain constraints on the worker's posture changes. Combined with an optimized light field sampling distribution, correction parameters for the transient distortion of the light field caused by posture changes are acquired, including: Based on the dynamic light intensity change and the construction worker's movement trajectory, the angle between the movement direction and the angle between the two points is calculated, the inflection point exceeding the threshold is marked, the light intensity difference is extracted, and the posture type is determined. Combined with the optimized light field sampling distribution, a four-dimensional spatiotemporal data point set is constructed, the transformation matrix is ​​calculated using the least squares method, the translation component and the scaling and rotation component are extracted, and the correction parameters are generated.

9. The abnormal behavior detection method based on monitoring according to claim 7, characterized in that, The process involves correcting the transient distortion of the light field based on attitude changes and optimizing the light field sampling distribution. Multi-scale feature fusion processing is then performed on the spatial light intensity distribution map to obtain the reflective vest feature distribution map. Finally, the image recognition result of the reflective vest is output, including: Based on the correction parameters, the light intensity values ​​and coordinates of the spatial light intensity distribution map are compensated, and the light intensity data is decomposed into multi-scale levels using wavelet transform to determine the fusion weights and generate the reflective clothing feature distribution map. Based on the feature intensity of the reflective clothing feature distribution map, the regions exceeding the threshold are marked as reflective clothing candidate regions, and the contours and positions are determined through connected component analysis to output the image recognition results.

Images