Method and system for monitoring illegal information based on Beidou positioning combined with digital simulation

Abstract

The application relates to the technical field of illegal information identification, in particular to a Beidou positioning and digital simulation-based illegal information monitoring method and system, which comprises the following steps: firstly, Beidou positioning data is used to strip a personnel mask area in a field monitoring video to determine a background area, and the background pixel gradient is quantified to determine the environmental interference complexity; then, the video is input into a convolution network to extract features, the feature channel mask rate is dynamically calculated according to the environmental interference complexity, and a binary mask is generated in combination with the activation intensity of a standard illegal feature; the mask is used to perform bitwise multiplication filtering on the field video features and the standard illegal features, and the similarity of the filtered feature vectors is calculated to determine the illegal type. The application solves the misidentification problem caused by dust and light fluctuation in a complex work site, and realizes high-precision automatic illegal identification.

Description

Technical Field

[0001] This application relates to the field of violation information identification technology, and in particular to a method and system for monitoring violation information based on BeiDou positioning combined with digital simulation. Background Technology

[0002] With the intelligent transformation of industrial scenarios such as power infrastructure and large-scale engineering construction, the demand for safety supervision at construction sites is becoming increasingly urgent. In order to prevent safety accidents caused by personnel violations, the use of computer vision technology for real-time monitoring and automatic identification of violations has become a core requirement of the industry.

[0003] To meet safety supervision requirements, the commonly used technical approach is deep learning-based video analysis. This method utilizes pre-trained target detection models to extract the posture or body features of people in real time from on-site surveillance videos and directly compares these features with preset typical violation action features. When the similarity between the extracted features and the template features exceeds a preset threshold, a violation warning signal is triggered, thus identifying the violation.

[0004] However, in actual construction sites, environmental conditions are often extremely complex and variable. Due to frequent sudden dust storms, smoke generated by machinery operation, and severe light fluctuations, a large amount of high-frequency noise and pixel-level interference will be generated in the background of the monitoring screen. The global feature alignment used in the feature comparison process cannot effectively shield these randomly generated environmental interference items, making it easy to generate false alarms or missed alarms due to noise interference in harsh environments. Summary of the Invention

[0005] To address the technical problem of low accuracy in violation identification caused by interference from complex environments, this application provides a method and system for monitoring violation information based on BeiDou positioning combined with digital simulation.

[0006] In a first aspect, this application provides a method for monitoring violation information based on BeiDou positioning combined with digital simulation. The monitoring method includes: determining a personnel masking region in the on-site monitoring video based on BeiDou positioning data of the construction site, and using the region after removing the personnel masking region from the on-site monitoring video as the background region; calculating the pixel gradient difference of the background region in two adjacent frames, and calculating the environmental interference complexity based on the pixel gradient difference; inputting the on-site monitoring video into a preset convolutional network for dimensionality reduction processing to obtain on-site video features; calculating the feature channel masking rate based on the environmental interference complexity, and generating a binary dimensionality reduction comparison mask based on the feature channel masking rate and the activation intensity of the standard violation features; using the binary dimensionality reduction comparison mask to perform bitwise multiplication filtering on the on-site video features and the standard violation features to obtain filtered on-site video feature vectors and filtered standard violation feature vectors; calculating the feature similarity of the two filtered feature vectors, and determining the target violation type existing at the construction site when the feature similarity is greater than a preset similarity threshold.

[0007] By introducing environmental interference complexity to drive the adaptive construction of feature masks, the monitoring process can dynamically adjust the feature filtering intensity according to the visual quality of the scene. This allows the model to focus on key action features in harsh environments, overcoming the visual interference of dust and light fluctuations on violation judgment and reducing the false alarm rate of behavior in complex operating scenarios.

[0008] Preferably, calculating the pixel gradient difference between two adjacent frames of the background region and calculating the environmental interference complexity based on the pixel gradient difference includes: using an edge detection operator to calculate the gradient magnitude of each pixel in two adjacent frames of the background region; and determining the environmental interference complexity by the ratio of the sum of the absolute values ​​of the gradient magnitude differences between each pixel in two adjacent frames to the product of the total number of pixels in the background region and the maximum allowed gradient magnitude.

[0009] Preferably, before acquiring the on-site monitoring video, the method further includes: performing image registration and image fusion on monitoring images from different perspectives at the construction site to obtain panoramic video frames; and combining the panoramic video frames in a time sequence to obtain the on-site monitoring video.

[0010] Preferably, inputting the on-site monitoring video into a preset convolutional network for dimensionality reduction processing to obtain on-site video features includes: extracting features from the on-site monitoring video using multiple convolutional layers and multiple pooling layers to obtain multidimensional original features; and performing global average pooling processing on the multidimensional original features to obtain on-site video features.

[0011] Preferably, the feature channel mask rate satisfies the following relationship:

[0012] ;

[0013] In the formula, For the feature channel mask rate, Based on the base mask rate, This is the mask sensitivity coefficient. This refers to the complexity of environmental interference.

[0014] By establishing an adaptive mapping relationship between the feature channel mask rate and the complexity of environmental interference, the adaptive adjustment of the filtering intensity is realized, effectively avoiding the identification error of violation information caused by the accumulation of environmental noise.

[0015] Preferably, generating a binary dimensionality reduction comparison mask based on the feature channel mask rate and the activation intensity of the standard violation feature includes: obtaining the activation intensity of the standard violation feature in each feature channel, and sorting each feature channel from low to high activation intensity; forcibly setting the mask bits corresponding to the feature channels at the bottom of the sort to 0, and forcibly setting the mask bits corresponding to the remaining feature channels to 1, to obtain the binary dimensionality reduction comparison mask, wherein the preset ratio is the feature channel mask rate.

[0016] By using the activation intensity of standard violation features to force the mask bits to zero, feature dimension filtering is achieved, ensuring that the retained channels all have high semantic representativeness of violations.

[0017] Preferably, calculating the feature similarity between two filtered feature vectors includes: calculating the inner product of the two filtered feature vectors, and determining the ratio of the inner product to the product of the magnitudes of the two filtered feature vectors as the feature similarity.

[0018] Preferably, after determining the type of violation, the method further includes: mapping the BeiDou positioning data to the virtual coordinate system of a pre-constructed three-dimensional simulation model to obtain a mapped coordinate point; constructing a spatial region with the mapped coordinate point as the center and a safety distance threshold as the radius, and using the spatial region as the location of the violation.

[0019] Preferably, the method for obtaining the standard violation features is as follows: acquiring historical violation monitoring videos of the construction site under conditions without environmental interference as historical data; inputting the historical data into the preset convolutional network for dimensionality reduction processing, extracting video features of each violation scene in the historical data; calculating the average value of the video features under each violation scene to obtain the standard violation features of each violation scene.

[0020] In a second aspect, this application also provides a traffic violation information monitoring system based on BeiDou positioning combined with digital simulation, including a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement the traffic violation information monitoring method based on BeiDou positioning combined with digital simulation according to the first aspect of this application.

[0021] The technical solution of this application has the following beneficial technical effects:

[0022] By using BeiDou positioning data to distinguish between personnel and background areas, and by quantifying the complexity of environmental interference using background pixel gradients, the feature channel mask rate is dynamically adjusted to achieve adaptive suppression of noise in real scenes. This effectively reduces environmental errors caused by scene noise between on-site monitoring video features and standard violation features. Furthermore, by combining binarized mask filtering with high-reliability feature comparison, the semantics of violation behavior can be accurately extracted from cluttered visual backgrounds, realizing an automated monitoring closed loop from environmental perception to behavior judgment and then to physical location mapping. Attached Figure Description

[0023] Figure 1 This is a flowchart of a method for monitoring traffic violations based on BeiDou positioning and digital simulation, according to an embodiment of this application.

[0024] Figure 2 This is a structural block diagram of a traffic violation information monitoring system based on BeiDou positioning and digital simulation, according to an embodiment of this application. Detailed Implementation

[0025] This application provides a method for monitoring traffic violations based on BeiDou positioning combined with digital simulation. Figure 1 This is a flowchart of a method for monitoring traffic violations based on BeiDou positioning combined with digital simulation, according to an embodiment of this application. Figure 1 As shown, the method for monitoring traffic violations based on BeiDou positioning and digital simulation includes steps S101 to S105, which are described in detail below.

[0026] S101, Based on the BeiDou positioning data of the construction site, determine the personnel masking area in the on-site monitoring video, and take the area after removing the personnel masking area from the on-site monitoring video as the background area; calculate the pixel gradient difference of the background area in two adjacent frames, and calculate the environmental interference complexity based on the pixel gradient difference.

[0027] In one embodiment, to obtain a high-quality monitoring view and eliminate the impact of personnel movement on the quantification of environmental complexity, multiple monitoring sources need to be aligned before acquiring the on-site monitoring video. Before acquiring the on-site monitoring video, the process further includes: image registration and image fusion of monitoring images from different perspectives at the construction site to obtain panoramic video frames; and combining the panoramic video frames in a time sequence to obtain the on-site monitoring video. Specifically, using pre-calibrated camera intrinsic and extrinsic parameters, feature matching points between images from different perspectives are found, and an affine transformation matrix is ​​calculated to uniformly map the images captured by multiple cameras to a single global pixel coordinate system. A weighted average fusion algorithm is used to eliminate stitching artifacts in overlapping areas, thereby generating panoramic video frames with a wide field of view. Arranging the continuously generated panoramic video frames in a time sequence with a frame rate of 25 frames per second constitutes the on-site monitoring video. This ensures that the monitoring view can completely cover every corner of the work site, eliminating blind spots.

[0028] After acquiring the on-site monitoring video, the real-time coordinates of the on-site personnel are obtained using a BeiDou positioning terminal to obtain the BeiDou positioning data. Since the BeiDou positioning data represents physical space coordinates, it needs to be mapped to the two-dimensional pixel coordinate system of the on-site monitoring video using a pre-calibrated perspective transformation matrix. Within each frame, a rectangular area covering the human body contour is constructed centered on the mapped pixel coordinates and combined with a preset proportion of human body size, serving as the personnel mask area; subsequently, all pixel areas in the on-site monitoring video other than the personnel mask area are defined as the background area.

[0029] Subsequently, an objective quantitative analysis of the environmental state of the background area is performed. To accurately remove inherent static structural interference from the scene, the pixel gradient difference of the background area in two adjacent frames is calculated, and the environmental interference complexity is calculated based on the pixel gradient difference, including: using an edge detection operator to calculate the gradient magnitude of each pixel in the background area in two adjacent frames; the ratio of the absolute value of the gradient magnitude difference of each pixel in two adjacent frames to the product of the total number of pixels in the background area and the physically allowed maximum gradient magnitude is determined as the environmental interference complexity.

[0030] It should be noted that the Sobel or Prewitt operators are used to perform convolution operations on the background regions of two adjacent frames to obtain the gradient magnitude of each pixel. Since the personnel masking region has been removed, and the complex background of the construction site, such as machinery and scaffolding, is stationary between adjacent frames, their corresponding gradient magnitudes cancel each other out in the difference calculation. Therefore, the difference in gradient magnitude between two adjacent frames only reflects the dynamic visual changes caused by smoke, dust, rain, snow, or light fluctuations, which have temporal randomness and high-frequency spatial abrupt changes.

[0031] Furthermore, the complexity of environmental disturbances satisfies the following relationship:

[0032]

[0033] In the formula, To account for the complexity of environmental interference; For the pixels in the background region of the current frame The gradient magnitude; This is the gradient magnitude at the same pixel location in the previous frame; This is the set of pixel coordinates for the background region. This represents the total number of pixels contained within the background region of the current frame. To allow the maximum gradient amplitude. Since personnel cannot fill the entire on-site monitoring video, therefore... The value of is a positive number greater than 0.

[0034] Understandably, the environmental interference complexity is reduced by subtracting adjacent frames to eliminate interference from the static background; the higher the sum of the absolute values ​​of the differences, the more severe the dynamic meteorological noise or smoke obstruction at the scene. By calculating the ratio of the actual total dynamic gradient jump to the theoretical maximum capacity, the degree of dynamic environmental interference can be quantified with extreme precision, eliminating the interference of static inherent textures on environmental state assessment.

[0035] It should be noted that the time interval between two adjacent frames can be set to 1 second to ensure that there is enough time to distinguish between a static background and dynamic visual changes that are in a high-frequency jumping state.

[0036] In this way, by stripping personnel targets from BeiDou positioning data and combining it with the inter-frame gradient difference mechanism, the accurate extraction and measurement of dynamic environmental interference at the work site were achieved, providing basic data for subsequent adaptive adjustment of the mask rate.

[0037] S102, the on-site monitoring video is input into a preset convolutional network for dimensionality reduction processing to obtain on-site video features.

[0038] In one embodiment, to transform on-site surveillance video containing pixel information into a high-dimensional semantic representation, the on-site surveillance video is input into a preset convolutional network for dimensionality reduction processing to obtain on-site video features. Specifically, multiple convolutional layers and multiple pooling layers are used to extract features from the on-site surveillance video to obtain multidimensional original features; global average pooling is then performed on the multidimensional original features to obtain the on-site video features.

[0039] The preferred pre-defined convolutional network is a deep convolutional network with residual structures, such as ResNet50. During processing, video frames from the on-site monitoring video are uniformly adjusted to a preset size, which can be 224×224. Multiple convolutional layers are used to perform sliding convolution operations on the adjusted video frames, capturing local spatial features such as limb lines and joint connections of personnel during construction through convolutional kernels of different sizes. Subsequently, pooling layers are used to perform downsampling operations, thereby outputting multidimensional original features of a three-dimensional tensor shape.

[0040] To transform the multidimensional original features of a three-dimensional tensor into a globally representative one-dimensional feature vector, global average pooling is performed. This global average pooling process satisfies the following relationship:

[0041]

[0042] In the formula, The first feature vector in the live video The characteristic values ​​of each channel; The height pixel value of the original multidimensional feature within a single channel; The width pixel value of the original multidimensional feature within a single channel; For multidimensional original features in the first The coordinates in each channel are and The value of the feature point at that location.

[0043] Understandably, global average pooling can directly reduce the dimensionality of feature maps to low-dimensional feature vectors, significantly filtering out redundant background textures in the original video. For example, if the multidimensional original features output by the pre-defined convolutional network have 512 channels, and the height and width pixel values ​​of each feature map are both 7, then after global average pooling, the original feature matrix containing 25088 elements is reduced to a live video feature vector containing 512 elements, providing a computational basis for subsequent feature comparison in a lower-dimensional mask space.

[0044] S103, calculate the feature channel mask rate based on the environmental interference complexity, and generate a binarized dimensionality reduction comparison mask based on the feature channel mask rate and the activation intensity of the standard violation feature.

[0045] In one embodiment, the method for obtaining the standard violation features is as follows: acquiring historical violation monitoring videos of the construction site under conditions without environmental interference as historical data; inputting the historical data into the preset convolutional network for dimensionality reduction processing, extracting video features of each violation scene in the historical data; calculating the average value of the video features under each violation scene to obtain the standard violation features of each violation scene.

[0046] Understandably, using the same pre-defined convolutional network to extract features from historical data and calculate the average not only ensures strict alignment of feature dimensions but also eliminates the random errors of individual historical samples. Since the standard violation feature distribution is extremely pure, it can serve as a benchmark template for judging violations; however, because it is unaffected by the current real-world environment, directly comparing it with on-site video features will lead to inaccuracies in the feature space. Therefore, it is necessary to dynamically construct a filtering mask using an environmental perception mechanism.

[0047] Subsequently, the feature channel mask rate is calculated based on the environmental interference complexity. The feature channel mask rate directly determines the proportion of interfered channels that need to be masked subsequently, and this feature channel mask rate satisfies the following relationship:

[0048]

[0049] In the formula, The feature channel mask rate; The base mask rate is preset to 0.05, which is used to maintain the basic generalization ability of the model and filter out normal background noise when environmental interference is extremely low. This is the mask sensitivity coefficient, with a preset optimal value of 0.4, used to control the amplification factor of the intensity of feature discarding caused by environmental interference; This refers to the complexity of environmental interference.

[0050] Understandably, the feature channel mask rate establishes an adaptive feature rejection operation. When severe dust or construction vehicles obstruct the view at the work site, increasing the complexity of environmental interference, the feature channel mask rate increases linearly. By discarding more weak feature channels that may carry noise in harsh environments, the system focuses on the most robust and discriminative core violation features, effectively mitigating the discrepancies between real-time features and standard templates caused by current environmental pollution.

[0051] After determining the feature channel mask rate, the precise construction of the mask vector is performed. Specifically, generating a binary dimensionality reduction comparison mask based on the feature channel mask rate and the activation intensity of the standard violation feature includes: obtaining the activation intensity of the standard violation feature in each feature channel, and sorting each feature channel from high to low activation intensity; forcing the mask bits corresponding to the feature channels at the bottom of the sort to 0, and forcing the mask bits corresponding to the remaining feature channels to 1, to obtain the binary dimensionality reduction comparison mask, where the preset ratio is the feature channel mask rate.

[0052] It should be noted that obtaining the activation intensity of the standard violation feature in each feature channel includes: extracting the feature value of the standard violation feature corresponding to each feature channel in a one-dimensional vector. Since the standard violation feature is output by a preset convolutional network after processing with a nonlinear activation function and global average pooling, the feature value of each channel is a non-negative real number. The physical meaning of this feature value directly represents the intensity of the response of the convolutional kernel to the violation action in a specific dimension. Therefore, the feature value on each feature channel is directly determined as the activation intensity of the corresponding feature channel; the larger the feature value, the higher the contribution of the feature channel to the expression of the violation semantics, and the higher its ranking in the order from high to low. Thus, the corresponding mask bit in the binarization dimensionality reduction comparison mask is forcibly set to 1 and retained; conversely, channels with smaller values ​​and ranked at a predetermined proportion are regarded as redundant or easily interfered with channels and are forcibly set to 0 and masked.

[0053] Understandably, since standard violation features are fixed feature vectors obtained offline from historical data, the activation intensity of each channel and the ranking results from high to low are fixed for a specific violation type. Based on the feature channel mask rate calculated from real-time on-site monitoring video, the truncation ratio for last-place elimination is dynamically determined; the more severe the environmental interference, the greater the dynamic truncation ratio, and the closer the retained feature channels are to the top-ranked core semantic channels.

[0054] For example, assume that the total dimension of the feature channels output by the pre-defined convolutional network is 512. The standard violation features, after averaging, are a fixed 512-dimensional vector, where the feature value of the 512th channel is 9.2, the 2nd channel is 5.8, and the 1st channel is 0.1. The feature values ​​of each channel are sorted from high to low as activation intensities, with the 512th channel ranked first, the 2nd channel ranked second, and the 1st channel ranked last. This ranking constitutes a fixed priori ranking table for this violation type. When the construction site is in clear, windless weather, the calculated environmental interference complexity is extremely low, resulting in a feature channel masking rate of 0.1. Based on a fixed prior ranking table, the mask bits corresponding to the 51 weakest feature channels (ranked in the bottom 10%) are forcibly set to 0. When severe dust occurs, the calculated environmental interference complexity is extremely high, resulting in a feature channel masking rate of 0.4. Again, based on the fixed prior ranking table, the mask bits corresponding to the 204 feature channels (ranked in the bottom 40%) are forcibly set to 0, retaining only the mask bits corresponding to the top 308 core skeleton channels as 1. The generated binarized dimensionality reduction comparison mask is a one-dimensional feature selection switch composed of values ​​0 and 1, perfectly matching the feature dimensions of the on-site video, ensuring that subsequent comparisons only occur on the strongest performing channels.

[0055] In this way, by introducing environmental interference complexity to quantify the mask discard ratio and using the activation intensity of standard violation features to determine the specific zeroing channel, adaptive filtering of the comparison space is achieved, laying a clean data foundation for subsequent accurate similarity calculation.

[0056] S104, the binarized dimensionality reduction comparison mask is used to perform bitwise multiplication filtering on the on-site video features and the standard violation features to obtain the filtered on-site video feature vector and the filtered standard violation feature vector.

[0057] In one embodiment, in order to perform baseline alignment within a clean feature subspace and eliminate the interference of environmental noise on subsequent violation determination, the binarized dimensionality reduction comparison mask is used to perform bitwise multiplication filtering on the on-site video features and the standard violation features to obtain the filtered on-site video feature vector and the filtered standard violation feature vector.

[0058] Specifically, the binarized dimensionality reduction comparison mask is used as a feature selection operator, and is multiplied bitwise by both the on-site video features and the standard violation features. The bitwise multiplication filtering satisfies the following relationship:

[0059]

[0060]

[0061] In the formula, The first element in the filtered live video feature vector The values ​​of each feature channel; The first element in the filtered standard violation feature vector The values ​​of each feature channel; The first feature in the on-site video The values ​​of each feature channel; The first of the standard characteristics of violations The values ​​of each feature channel; For binarization dimensionality reduction comparison mask, the first The mask value is a bit, and its value can only be a constant 0 or a constant 1; This is the index number of the feature channel, and its maximum value is equal to the total dimension of the feature channel.

[0062] Understandably, since the binarization dimensionality reduction comparison mask only contains the values ​​0 and 1, bitwise multiplication essentially performs a masking mechanism. When the mask bit is 0, the feature channel values ​​corresponding to the on-site video features and standard violation features at that position are forced to zero, blocking channel features that are severely affected by environmental interference or have low contribution to the semantics of the violation. When the mask bit is 1, the feature channel values ​​corresponding to that position are retained proportionally without loss, ensuring that the retained non-zero elements are fully focused on the most discriminative features.

[0063] Thus, the binarization mask provides the most discriminative feature for subsequent similarity calculation.

[0064] S105, calculate the feature similarity between the two filtered feature vectors, and when the feature similarity is greater than a preset similarity threshold, determine the type of target violation existing at the construction site.

[0065] In one embodiment, to assess the degree of suspicion of violation in on-site operations, it is necessary to calculate the feature similarity between the two filtered feature vectors. Calculating the feature similarity between the two filtered feature vectors includes: calculating the inner product of the two filtered feature vectors, and determining the feature similarity as the ratio of the inner product to the product of the magnitudes of the two filtered feature vectors.

[0066] Feature similarity satisfies the following relationship:

[0067]

[0068] In the formula, This is the feature similarity, and its value ranges from 0 to 1; The first element in the filtered live video feature vector The values ​​of each feature channel; The first element in the filtered standard violation feature vector The values ​​of each feature channel; This represents the total dimension of the feature vectors, and its values ​​are consistent with the dimension after global average pooling. The value is a preset positive number, and in this embodiment it is set to 1.

[0069] Understandably, the feature vectors after binarization and dimensionality reduction, along with mask filtering, have eliminated most of the dimensions contaminated by environmental noise in the space. By calculating the cosine of the angle between the two vectors in the clean dimension, pixel value fluctuations caused by absolute differences in environmental brightness or the superposition of background clutter can be eliminated, focusing instead on the person's actions and postures. The closer the two vectors are to each other, the closer the feature similarity is to 1, indicating a higher degree of matching between the on-site behavior and the standard violation template.

[0070] After obtaining the feature similarity, if the feature similarity is greater than a preset similarity threshold, the target violation type existing at the construction site is determined, where the preset similarity threshold is 0.85. Since the standard violation features include label indexes for multiple different violation scenarios during construction, all labels that meet the similarity standard are used as target violation types and output. That is, a person may have multiple target violation types at the same time, such as illegally entering a high-voltage area without wearing a safety helmet.

[0071] After determining a violation, to achieve precise on-site intervention, following the identification of the target violation type, the process includes: mapping the BeiDou positioning data to a pre-constructed virtual coordinate system of a 3D simulation model to obtain mapped coordinate points; constructing a spatial region centered on the mapped coordinate points and with a safety distance threshold as the radius, and using this spatial region as the location of the violation. Here, the personnel pixel coordinates are the center point of the bounding box output by the target detection network; using a pre-calibrated perspective transformation matrix, the 2D pixels are restored to their true 3D coordinates, thereby accurately guiding management personnel to quickly locate the source of the violation.

[0072] In this way, by using quantified feature similarity to achieve automatic identification and classification of violations, and by accurately guiding managers to quickly locate the source of violations through physical location mapping, the safety control intensity of construction sites has been greatly improved.

[0073] According to a second aspect of this application, this application also provides a violation information monitoring system based on BeiDou positioning combined with digital simulation. Figure 2 This is a structural block diagram of a traffic violation information monitoring system based on BeiDou positioning and digital simulation, according to an embodiment of this application. Figure 2 As shown, the system 50 includes a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement the violation information monitoring method based on BeiDou positioning combined with digital simulation according to the first aspect of this application. The system also includes other components well-known to those skilled in the art, such as a communication bus and communication interface. Their configuration and functions are known in the art and will not be described further here.

[0074] It should be noted that, for those skilled in the art, several modifications and improvements can be made without departing from the concept of this application, and these all fall within the scope of protection of this application.

Claims

1. A method for monitoring traffic violations based on BeiDou positioning combined with digital simulation, characterized in that, The monitoring method includes: Based on the BeiDou positioning data at the construction site, the personnel masking area is determined in the on-site monitoring video, and the area after removing the personnel masking area from the on-site monitoring video is used as the background area. Calculate the pixel gradient difference of the background region in two adjacent frames, and calculate the environmental interference complexity based on the pixel gradient difference; The on-site monitoring video is input into a preset convolutional network for dimensionality reduction processing to obtain the on-site video features; The feature channel mask rate is calculated based on the environmental interference complexity, and a binarized dimensionality reduction comparison mask is generated based on the feature channel mask rate and the activation intensity of the standard violation feature. The binarized dimensionality reduction comparison mask is used to perform bitwise multiplication filtering on the on-site video features and the standard violation features to obtain the filtered on-site video feature vector and the filtered standard violation feature vector. Calculate the feature similarity between the two filtered feature vectors, and when the feature similarity is greater than a preset similarity threshold, determine the type of target violation present at the construction site; the feature channel mask rate satisfies the following relationship: ； In the formula, For the feature channel mask rate, Based on the base mask rate, This is the mask sensitivity coefficient. To mitigate environmental interference complexity, the generation of a binary dimensionality reduction comparison mask based on the feature channel mask rate and the activation intensity of the standard violation feature includes: obtaining the activation intensity of the standard violation feature in each feature channel and sorting each feature channel from low to high activation intensity; forcibly setting the mask bits corresponding to the feature channels at the bottom of the sort to 0, and forcibly setting the mask bits corresponding to the remaining feature channels to 1, thereby obtaining the binary dimensionality reduction comparison mask, where the preset ratio is the feature channel mask rate.

2. The method for monitoring traffic violations based on BeiDou positioning combined with digital simulation according to claim 1, characterized in that, Calculating the pixel gradient difference of the background region in two adjacent frames, and calculating the environmental interference complexity based on the pixel gradient difference, includes: The gradient magnitude of each pixel in two adjacent frames within the background region is calculated using edge detection operators. The environmental interference complexity is determined by the ratio of the absolute value of the difference between the gradient magnitudes of each pixel in two adjacent frames to the product of the total number of pixels in the background region and the maximum allowed gradient magnitude.

3. The method for monitoring traffic violations based on BeiDou positioning combined with digital simulation according to claim 1, characterized in that, Before obtaining on-site surveillance video, the following steps are also included: Image registration and fusion are performed on monitoring images from different perspectives at the construction site to obtain panoramic video frames; The panoramic video frames are combined in time sequence to obtain the on-site monitoring video.

4. The method for monitoring traffic violations based on BeiDou positioning combined with digital simulation according to claim 1, characterized in that, The on-site monitoring video is input into a preset convolutional network for dimensionality reduction processing, resulting in on-site video features including: Multiple convolutional layers and multiple pooling layers are used to extract features from the on-site monitoring video to obtain multidimensional original features; The original multidimensional features are subjected to global average pooling to obtain the on-site video features.

5. The method for monitoring traffic violations based on BeiDou positioning combined with digital simulation according to claim 1, characterized in that, Calculating the feature similarity between the two filtered feature vectors includes: Calculate the inner product of the two filtered feature vectors, and determine the feature similarity as the ratio of the inner product to the product of the magnitudes of the two filtered feature vectors.

6. The method for monitoring traffic violations based on BeiDou positioning combined with digital simulation according to claim 1, characterized in that, After determining the type of traffic violation, the following is also included: The BeiDou positioning data is mapped onto the virtual coordinate system of a pre-constructed three-dimensional simulation model to obtain the mapped coordinate points; A spatial region is constructed with the mapped coordinate point as the center and the safety distance threshold as the radius, and the spatial region is used as the location of the violation.

7. The method for monitoring traffic violations based on BeiDou positioning combined with digital simulation according to claim 1, characterized in that, The method for obtaining the standard violation features is as follows: acquiring historical violation monitoring videos of the construction site under conditions of no environmental interference as historical data; inputting the historical data into the preset convolutional network for dimensionality reduction processing, and extracting video features of each violation scene in the historical data; Calculate the average value of video features for each violation scenario to obtain the standard violation features for each scenario.

8. A violation information monitoring system based on BeiDou positioning combined with digital simulation, characterized in that, It includes a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the method for monitoring traffic violations based on BeiDou positioning combined with digital simulation according to any one of claims 1 to 7 is implemented.

Images