Power transmission channel hidden danger risk assessment method, system, device and medium

Abstract

This invention belongs to the field of power transmission channel hazard monitoring technology, specifically disclosing a method, system, equipment, and medium for power transmission channel hazard risk assessment. The method includes: using a target detection model to identify hazards in acquired power transmission channel images and outputting detection boxes for hazard targets; fusing the category confidence heatmap and local contrast saliency map of each detection box to obtain a joint saliency map corresponding to each detection box; calculating an adaptive threshold for the current detection box; performing connected component analysis on pixels within the detection box whose joint saliency value is greater than the adaptive threshold to obtain candidate regions for the highest point; determining the highest point of the hazard target within the current detection box from the candidate regions for the highest point; calculating the clearance distance between the highest point and the transmission line; and determining the risk level of the hazard target. This invention can achieve accurate detection of the highest point of the hazard target and precise identification of hazard risk without changing existing hardware.

Description

Technical Field

[0001] This invention belongs to the field of power transmission channel hidden danger monitoring technology, and in particular relates to a method, system, equipment and medium for assessing the hidden danger risks of power transmission channels. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] In the operation and maintenance of power transmission channels, the clearance distance between potential hazards and the conductors is a core indicator for assessing safety risks. Currently, a large number of monocular vision monitoring devices are still deployed in power transmission channels. While these devices can identify the types of potential hazards, they cannot directly output the actual spatial location of the hazards. Taking the entry of large machinery such as cranes and excavators into power transmission channels as an example, after receiving an alarm, maintenance personnel need to log into the platform to view on-site images, manually calculate the relative distance between the large machinery and the transmission line, or go to the site for verification. This process is time-consuming and makes accurate early warning difficult.

[0004] While existing target detection methods can identify the type and detection bounding box of potential hazards, they cannot pinpoint the highest point that poses the greatest threat, nor can they calculate its true geographical coordinates and the clearance distance to the power line. This results in the inability to quantify alarm thresholds, requiring manual verification on-site. Some methods rely on labeled data of the highest point or additional training of segmentation network models, which leads to high labeling costs, poor generalization ability, and difficulty in adapting to complex and ever-changing power transmission scenarios.

[0005] While distance measurement can be achieved using lidar or binocular vision devices, additional equipment needs to be deployed, resulting in high modification costs and long deployment cycles, making it difficult to promote on a large scale in the current scenarios of monocular vision surveillance equipment. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a method, system, device, and medium for risk assessment of potential hazards in power transmission channels. Without additional data annotation or changes to hardware configuration, it can accurately obtain the physical coordinates of the highest point of the potential hazard target, accurately calculate the clearance distance between the potential hazard target and the power transmission line, and then assess the risk of the potential hazard target based on the clearance distance.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: The first aspect of this invention provides a method for assessing potential risks in power transmission channels, comprising: The target detection model is used to identify potential hazards in the acquired power transmission channel images, and the detection bounding boxes of potential hazards are output. The category confidence heatmap and local contrast saliency map of each detection box are fused to obtain the joint saliency map corresponding to each detection box; Using the mean of the joint saliency map as a benchmark, and the product of the standard deviation and a preset adjustment coefficient as a dynamic adjustment term, the adaptive threshold corresponding to the current detection box is calculated; connected component analysis is performed on the pixels in the detection box whose joint saliency value is greater than the adaptive threshold to obtain the candidate region of the highest point; Determine the highest point of the potential hazard target within the current detection frame from the candidate region of the highest point; Calculate the clearance distance between the highest point and the transmission line to determine the risk level of the potential hazard.

[0008] This invention utilizes a lightweight target detection model, eliminating the need for peak point annotation data or additional segmentation network training, thus adapting to complex and ever-changing power transmission scenarios. By fusing a category confidence heatmap and a LC saliency map to obtain a joint saliency map, it achieves complementary enhancement of semantic information and visual saliency, effectively suppressing background noise interference with peak point localization. Through adaptive threshold segmentation and connected component analysis of the joint saliency map, a unique peak point is selected based on physical peak point rules. The entire process requires no changes to existing hardware, has extremely low computational overhead, and is suitable for the real-time operation needs of current low-computing-power edge devices.

[0009] This invention dynamically calculates an adaptive threshold based on the statistical distribution of the joint saliency map within the current detection box. It can automatically adjust according to the image content, eliminating the need for manual parameter tuning for different scenarios and fully adapting to complex and ever-changing power transmission scenarios.

[0010] As a further step, the process of generating the category confidence heatmap for each detection box is as follows: Obtain the last layer feature map in the target detection model before entering the detection head, and calculate the average gradient of each channel in the feature map as the weight of the corresponding channel; wherein, the average gradient is obtained by global average pooling based on the partial derivative of the target category score at each position in the feature map; Based on the weights, the feature maps of each channel are weighted and summed, and negative values ​​are removed by ReLU activation to obtain the original heatmap. The original heatmap is upsampled to the same size as the original power transmission channel image and normalized to obtain a category confidence heatmap.

[0011] As a further step, the calculation process for the local contrast saliency map of each detection box is as follows: The original grayscale values ​​of the detection box are uniformly compressed and quantized, and the weight coefficient of each quantization level after compression is calculated; wherein, the weight coefficient is calculated based on the difference of pixel grayscale values ​​between the target quantization level and other quantization levels; For each pixel within the detection box, the corresponding weight coefficient is taken as the local contrast saliency value of the pixel according to its quantization level; finally, the obtained local contrast saliency map is normalized to obtain the final local contrast saliency map of the detection box.

[0012] As a further step, the weighting coefficient for each quantization level is as follows: Count the number of pixels at each quantization level within the compressed detection box; For a target quantization level q, iterate through each quantization level r, calculate the sum of the products of the absolute difference between the center grayscale of the target quantization level q and the quantization level r and the number of pixels in the corresponding quantization level r, and use this sum as the weight coefficient of the target quantization level q.

[0013] As a further solution, the original grayscale values ​​of the detection box are uniformly compressed and quantized, specifically as follows: The original grayscale levels from 0 to 255 are uniformly compressed and quantized to 0 to 31.

[0014] To address the computational efficiency limitations on resource-constrained devices, this invention balances accuracy and computational load when calculating the LC saliency map. It uniformly compresses and quantizes the original grayscale values ​​and uses the compressed grayscale levels to calculate the LC saliency map, which greatly reduces the computational load and improves computational efficiency.

[0015] As a further step, the category confidence heatmap and local contrast saliency map of each detection box are fused together, specifically: Detection box i The heatmap value of each pixel in the category confidence heatmap is compared with the detection box. i The detection box is obtained by multiplying the saliency values ​​of corresponding pixels in the local contrast saliency map. i The joint saliency value corresponding to each pixel in the detection box is used to obtain the detection box. i The corresponding joint saliency map.

[0016] As a further solution, connected component analysis is performed on pixels within the detection box whose joint saliency value is greater than the adaptive threshold to obtain candidate regions for the highest point, specifically: All pixels within the detection box whose joint saliency value is greater than the adaptive threshold constitute a binary mask. Connectivity analysis is performed on the binary mask to remove isolated pixels, and one or more connected regions are obtained as candidate regions for the highest point.

[0017] As a further solution, the highest point of the potential hazard target within the current detection frame is determined from the candidate region of the highest point, specifically as follows: Using the top left corner of the image as the origin, select the pixel with the smallest ordinate from all candidate regions of the highest point. If there is only one pixel with the smallest ordinate, then that pixel is taken as the highest point of the hazard target in the current detection box; if there are multiple pixels with the smallest ordinate, then the pixel with the largest joint significance value is taken as the highest point of the hazard target in the current detection box.

[0018] As a further option, it also includes: If the highest point candidate region cannot be obtained, then a region with a ratio of γ is selected at the top of the detection frame. The height of the region is γ times the height of the detection frame, and the width is the same as the width of the detection frame; where γ is a preset value. The pixel with the largest joint saliency value within the area is selected as the highest point of the potential hazard target within the current detection frame.

[0019] As a further measure, the clearance distance between the highest point and the transmission line is calculated to determine the risk level of the potential hazard, specifically: Using the pre-calibrated camera intrinsic parameter matrix K, extrinsic parameter rotation matrix R, and translation vector T, combined with the ground elevation model, the pixel coordinates of the highest point are back-projected onto the camera normalization plane, then transformed to the world coordinate system, and intersected with the ground plane to finally obtain the true three-dimensional coordinates of the highest point. Based on the pre-established three-dimensional spatial model of the transmission line conductor, the three-dimensional real coordinates of the highest point and the shortest Euclidean distance of the transmission line are calculated, which is the clearance distance from the hidden danger target to the transmission line. Based on the aforementioned clearance distance, the risk level of the potential hazard is determined.

[0020] A second aspect of the present invention provides a power transmission channel hazard risk assessment system, comprising: The hazard target identification module is configured to use a target detection model to identify hazards in the acquired power transmission channel images and output the detection bounding box of the hazard target; The joint saliency map construction module is configured to fuse the category confidence heatmap and local contrast saliency map of each detection box to obtain the joint saliency map corresponding to each detection box. The highest point detection module is configured to use the mean of the joint saliency map as a benchmark and the product of the standard deviation and a preset adjustment coefficient as a dynamic adjustment term to calculate the adaptive threshold corresponding to the current detection box; perform connected component analysis on the pixels in the detection box whose joint saliency value is greater than the adaptive threshold to obtain the candidate region of the highest point; and determine the highest point of the hidden danger target in the current detection box from the candidate region of the highest point. The hazard risk assessment module is configured to calculate the clearance distance between the highest point and the transmission line to determine the risk level of the hazard target.

[0021] As a further solution, the system also includes: The supplementary positioning module is configured to select a region with a ratio of γ at the top of the detection frame when the candidate region of the highest point cannot be obtained. The height of the region is γ times the height of the detection frame, and the width is the same as the width of the detection frame; where γ is a preset value. The pixel with the largest joint saliency value within the area is selected as the highest point of the potential hazard target within the current detection frame.

[0022] As a further solution, the hazard risk assessment module includes: The coordinate transformation unit is configured to use the pre-calibrated camera intrinsic parameter matrix K, extrinsic parameter rotation matrix R, and translation vector T, combined with the ground elevation model, to back-project the pixel coordinates of the highest point onto the camera normalization plane, then transform it to the world coordinate system, and intersect it with the ground plane to finally obtain the three-dimensional true coordinates of the highest point. The hazard identification unit is configured to calculate the shortest Euclidean distance between the three-dimensional real coordinates of the highest point and the transmission line based on a pre-established three-dimensional spatial model of the transmission line conductor, which is the clearance distance from the hazard target to the transmission line; and determine the risk level of the hazard target based on the clearance distance.

[0023] A third aspect of the present invention provides a power transmission channel hazard target monitoring device, which includes a processor and a memory. The processor is used to implement instructions; the memory is used to store multiple instructions, which are adapted to be loaded and executed by the processor to perform the above-described power transmission channel hazard risk assessment method.

[0024] A fourth aspect of the present invention provides a chip including the aforementioned power transmission channel hazard monitoring device.

[0025] A fifth aspect of the present invention provides a computer-readable storage medium storing a plurality of instructions adapted to be loaded and executed by a processor of a terminal device, the above-described method for assessing potential risks in power transmission channels.

[0026] A sixth aspect of the present invention provides a computer program product, including a computer program or instructions that, when executed by a processor, implement the above-described method for assessing potential risks in power transmission channels.

[0027] Compared with the prior art, the beneficial effects of the present invention are: This invention uses the category confidence heatmap generated during the inference process of a lightweight object detection network as a semantic prior, and performs element-wise weighted fusion with the LC (Local Contrast) saliency map that does not require training to construct a joint saliency map; it achieves complementary enhancement of semantic information and visual saliency, and effectively suppresses the interference of background noise on the localization of the highest point without adding extra annotation and training.

[0028] This invention performs adaptive threshold segmentation on the joint saliency map. The adaptive threshold is dynamically calculated based on the statistical distribution of the joint saliency map within the current detection box. It can be automatically adjusted according to the image content, eliminating the need for manual parameter tuning for different scenarios and adapting to the complex and ever-changing characteristics of power transmission channel scenarios.

[0029] This invention combines connected component analysis and target feature region selection at a predetermined ratio. Without changing existing hardware, it can accurately detect the highest point of a potential hazard and accurately calculate the clearance distance between the potential hazard and the transmission line. It is particularly suitable for existing monocular vision monitoring equipment in power transmission channels. Based on the clearance distance, it classifies the risk level of potential hazards and generates differentiated alarm information according to the different risk levels, which significantly improves the engineering practicality of the monocular vision solution. Attached Figure Description

[0030] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0031] Figure 1 This is a flowchart of the risk assessment method for hidden dangers in power transmission channels in an embodiment of the present invention; Figure 2 This is a schematic diagram of the joint saliency map generation process in an embodiment of the present invention; Figure 3 This is an example of an input test image in an embodiment of the present invention; Figure 4 This is a schematic diagram of the target detection bounding box obtained based on the test image; Figure 5 This is a heatmap of category confidence obtained based on the test image; Figure 6 The LC significance map is for the leftmost detection box. Figure 7 The joint saliency map is calculated for the leftmost detection box; Figure 8 This is a schematic diagram of the candidate region for the highest point obtained by performing connected component analysis on the leftmost detection box based on the joint saliency map; Figure 9 A schematic diagram to determine the highest point of the potential hazard target within the current detection frame; Figure 10 This is a schematic diagram showing the clearance distance measured based on the highest point. Detailed Implementation

[0032] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0033] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0034] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0035] Example 1 The front-end monitoring equipment targeted in this embodiment is mainly the existing monocular monitoring equipment already installed on transmission towers. These devices cannot achieve accurate positioning and ranging like binocular vision or lidar, so they cannot directly locate the highest point of the potential hazard, and therefore cannot accurately calculate the clearance distance between the potential hazard and the transmission line, thus making it impossible to conduct a hazard risk assessment.

[0036] Based on this, in one or more embodiments, a method is disclosed that can accurately output the clearance distance between the potential hazard target and the conductor and achieve hazard risk assessment without requiring additional hardware deployment, combined with... Figure 1 Specifically, it includes the following process: S101: Use the target detection model to identify potential hazards in the acquired power transmission channel images, and output the detection box of the potential hazard target and its category confidence heatmap.

[0037] In this embodiment, the RGB images of the power transmission channel are acquired by a front-end monitoring device installed on the power transmission tower at a set cycle.

[0038] The acquired images are fed into a lightweight target detection model, which can be a lightweight deep learning network, such as the YOLOv4 series. A pre-built dataset of potential hazards in power transmission channels is used to train the target detection model, enabling it to accurately identify common potential hazards such as cranes, excavators, and tower cranes.

[0039] The acquired images are processed by a trained object detection model, which can output the bounding box location and category of each potential hazard. Simultaneously, a feature map is extracted from the last convolutional layer before the network detection head of the object detection model; this feature map is denoted as... Its dimensions are ,in For the number of channels, , These represent the length and width of the feature map, respectively.

[0040] This embodiment calculates the category confidence heatmap corresponding to each detection box based on the feature map. The specific process is as follows: S1011: Obtain the last layer feature map A in the target detection model before entering the detection head, and calculate the average gradient of each channel in feature map A as the weight of the corresponding channel; For each hazard target category, the average gradient of each channel in the feature map is calculated as the weight of that channel. The average gradient of each channel is obtained by global average pooling of the partial derivative of the target category score at each position in the feature map. The specific formula is as follows: ; in, Let c be the weight of the c-th channel. The target category score output by the network. Refers to the first Location on each channel eigenvalues. This indicates the sensitivity of the target category score Y to this feature value. The weight of the c-th channel is obtained by summing and averaging over all positions. .

[0041] S1012: Weighted summation of the feature maps of each channel, followed by ReLU activation to remove negative values, yields the original heatmap; specifically: ; in, This represents the original heatmap. Indicates the first The feature map of each channel, with a size of H'×W', contains high-level semantic features for identifying potential hazards. The result of a weighted summation of the feature maps of all C channels is a two-dimensional matrix of size H'×W'. The value at each position in the matrix is ​​the weighted sum of the responses of different channel features at that position. This is the activation function, which sets all negative values ​​in the weighted summation result to 0 and retains only positive values.

[0042] S1013: Upsample the original heatmap to the same size as the original power transmission channel image. The data is then normalized to the 0-1 range to obtain the category confidence heatmap.

[0043] The category confidence heatmap in this embodiment can reflect the spatial distribution of the contribution of each spatial location in the image to the model decision. Each pixel value in the heatmap represents the confidence of the model in determining that the pixel belongs to the current hidden danger target. The closer the value is to 1, the greater the contribution of the location to the target classification. It usually corresponds to the key structural area of ​​the target, such as the top of the crane boom or the bucket teeth of the excavator.

[0044] The category confidence heatmap transforms the decision-making basis of the object detection model into a visual spatial attention distribution, thereby making the model's behavior interpretable and verifiable.

[0045] S102: Calculate the local contrast (LC) saliency map for each detection box, and multiply the category confidence heatmap and the LC saliency map element by element to generate the joint saliency map for each detection box.

[0046] In this embodiment, after obtaining the detection box, the image region within each box is processed separately.

[0047] For each detection box, the color image within the box can be converted to grayscale, and then the LC saliency map can be calculated. The formula for calculating the LC saliency map is: ; in, This represents the set of pixels within the current detection box. For target pixels grayscale value, This represents the grayscale value of each pixel in the pixel set.

[0048] However, since the original grayscale level of a grayscale image is 0-255, direct calculation is highly complex and places high demands on the performance of the deployment device. To address the computational efficiency bottleneck of LC saliency calculation on resource-constrained devices, this embodiment adopts an adaptive LC acceleration calculation method for edge deployment. This method reduces the computational load by compressing and quantizing the original grayscale levels. The specific implementation process is as follows: S1021: Compress and quantize the original grayscale value of the detection box uniformly, and calculate the weight coefficient of each quantization level after compression; The weighting coefficient for each quantization level is calculated as follows: The number of pixels in each quantization level within the compressed detection box is counted. For the target quantization level q, each quantization level r is traversed, and the sum of the products of the absolute difference between the center gray level of the target quantization level q and the quantization level r and the number of pixels in the corresponding quantization level r is calculated as the weight coefficient of the target quantization level q.

[0049] S1022: For each pixel within the detection box, the corresponding weight coefficient is taken as the LC saliency value of the pixel according to its quantization level; finally, the obtained LC saliency map is normalized to obtain the final LC saliency map of the detection box.

[0050] As a specific implementation method, if the original grayscale is uniformly quantized to 32 levels, each quantization level covers 8 original grayscale values. The maximum deviation between the center grayscale and the actual grayscale is 4 grayscale levels, and the impact on the positioning of the highest point can be ignored. If uniform quantization is performed to 64 levels, the computational load will increase by nearly double, but the positioning accuracy will only improve by about 0.3% compared to 32 levels, which is not cost-effective. If 16 levels are used, although the calculation is faster, the positioning accuracy will decrease significantly in complex lighting scenes such as evening shadows or backlighting.

[0051] Therefore, considering the balance between accuracy and computational complexity, this embodiment uniformly quantizes the original 0 to 255 gray levels into 32 levels with a quantization step size of 8. The quantization function is denoted as: ; in, For pixels The original grayscale value, Quantized pixels The grayscale value ranges from 0 to 31.

[0052] Each quantization level within the statistical detection frame Number of pixels and define quantization levels The center gray level is .

[0053] For each target quantization level q, the weighting coefficients are pre-calculated: ; in, The weights for the target quantization level q, The center gray level of the target quantization level q, For quantization level The center gray level.

[0054] For each quantization level, the corresponding weighting coefficients are calculated using the method described above.

[0055] Then iterate through each pixel within the detection box, based on its quantization level. Take the corresponding weight coefficient directly. The LC saliency values ​​are used as the saliency values; the LC saliency values ​​of all pixels within the detection box constitute the saliency map; the entire saliency map is normalized to the interval between 0 and 1 to obtain the final LC saliency map.

[0056] In this embodiment, the LC saliency map can reflect pixel-level visual differences. The higher the LC saliency value, the greater the difference between the gray level of the pixel and the gray level of most pixels in the image, and the higher the visual contrast.

[0057] Combination Figure 2 After obtaining the category confidence heatmap and the LC significance map, the normalized category confidence heatmap and the LC significance map are multiplied pixel by pixel to generate a joint significance map, specifically represented as follows: ; in, For pixels The joint significance value, Represents pixels Category confidence heat value, For pixels The LC saliency values ​​of all pixels, and the joint saliency values ​​of all pixels, constitute the joint saliency map.

[0058] In this embodiment, It is obtained by cropping the corresponding region from the full image heatmap and then normalizing it. The full image heatmap refers to the category confidence heatmap output by the object detection network that covers the entire input image. The corresponding region refers to the local area in the full image heatmap that spatially corresponds to a detection box. The cropped local data is normalized to obtain the category confidence heatmap of the corresponding detection box.

[0059] In this embodiment, the category confidence heatmap generated during the inference process of the lightweight object detection network is used as a semantic prior and fused with the LC saliency map which does not require training, element-wise weighted fusion is performed. After element-wise weighted fusion, only those pixels that are both judged as the core structure of the target by the object detection model and have strong visual contrast can obtain higher response values, thereby achieving complementary enhancement of semantic information and visual saliency. Without adding additional annotation and training, the interference of background noise on the localization of the highest point is effectively suppressed.

[0060] S103: Dynamically calculate the adaptive threshold based on the statistical distribution of the joint saliency map within the current detection box, perform connected component analysis on pixels within the detection box whose joint saliency value is greater than the adaptive threshold, obtain the candidate region of the highest point, and determine the highest point of the hidden danger target within the current detection box from the candidate region of the highest point based on the physical highest point rule (the principle of minimum ordinate).

[0061] In power transmission channel monitoring scenarios, monitoring cameras are typically mounted horizontally according to line inspection standards, with their optical axes nearly level with the horizontal plane. Therefore, the image's vertical coordinate (y-value) directly corresponds to the target's altitude in physical space. Generally, with the top left corner of the image as the origin, based on perspective imaging principles, points with higher altitudes in physical space are projected higher onto the image plane (with smaller vertical coordinates). This embodiment limits the search range to the candidate region for the highest point. Within this local region, the pixel with the smallest vertical coordinate is selected, i.e., the point closest to the top edge within the region. This point is equivalent to the highest point of the potential hazard target in physical space. This selection rule conforms to the physical laws of imaging geometry and is basically matched with the actual installation and deployment conditions of the power transmission channel monitoring cameras. Therefore, the point with the smallest vertical coordinate is selected as the highest physical point.

[0062] Specifically, after obtaining the joint saliency map, an adaptive threshold is dynamically calculated based on the statistical distribution of the joint saliency map within the current detection box; the calculation process of the adaptive threshold is as follows: Calculate the mean of the joint saliency map within the current detection box. and standard deviation Using the mean as a benchmark, and the product of the standard deviation and a preset adjustment coefficient as a dynamic adjustment term, the adaptive threshold corresponding to the current detection box is calculated. The corresponding formula is: ; in, The adjustment coefficient is determined empirically, and in this embodiment, the preferred value is λ=1.5.

[0063] Fixed thresholds are difficult to account for all situations. If the threshold is too high, it is easy to miss the real high points. If it is too low, it will introduce a lot of noise. It is not suitable for the complex and ever-changing situation of power transmission channels.

[0064] This embodiment employs an adaptive threshold to segment the joint saliency map. The threshold is dynamically calculated based on the statistical distribution of the joint saliency map within the current detection box, automatically adjusting to the image content without requiring manual parameter tuning for different scenes. Furthermore, the joint saliency map is formed by fusing a class confidence heatmap and an LC saliency map; its numerical distribution is scene-dependent. The adaptive threshold naturally matches this distribution characteristic, effectively filtering out isolated noise points with low confidence while preserving the true target area.

[0065] Meanwhile, the adaptive threshold generation method in this embodiment has extremely low computational overhead, making it suitable for the real-time operation requirements of edge devices. Furthermore, the empirical value of λ=1.5 has proven its robustness in numerous experiments. Combined with the γ rule as a fallback when there are no candidates, the overall robustness is even stronger.

[0066] After calculating the adaptive threshold of the current detection box, pixels with a joint saliency value greater than the adaptive threshold are selected to form a binarized mask. Connectivity analysis is performed on the binarized mask to remove isolated pixels, and one or more connected regions are obtained as candidate regions for the highest point.

[0067] For all pixels within the candidate regions of the highest point, the highest point of the hazard target within the current detection box is selected from all candidate regions based on the principle of minimizing the ordinate. The specific selection method is as follows: Select the pixel with the smallest ordinate (i.e., closest to the top edge of the detection box) from all candidate regions of the highest point. If there is only one pixel with the smallest ordinate, then that pixel is taken as the highest point of the hazard target in the current detection box; if there are multiple pixels with the smallest ordinate, then the pixel with the largest joint significance value is taken as the highest point of the hazard target in the current detection box.

[0068] In most power transmission channel hazards, the top of the target is a key reference point for identification or ranging, such as the top of the boom of construction machinery, the top of the tower, or the highest point of the tree canopy. In this embodiment, the pixel with the smallest ordinate is selected as the highest point of the hazard target by combining the joint saliency value. Compared with other complex algorithms, the computational overhead is extremely low and the identification accuracy is high, which is very suitable for real-time processing scenarios on edge devices.

[0069] S104: If the candidate area of ​​the highest point cannot be obtained, a predetermined proportion of the area is selected in the upper part of the detection frame, and the point with the largest joint significance value in the area is selected as the highest point of the hidden danger target in the current detection frame.

[0070] In this embodiment, if the joint saliency value of pixels within the selected detection box is greater than the adaptive threshold, and no connected region is obtained through connected component analysis, or if the joint saliency value of all pixels within the detection box is not greater than the adaptive threshold, then the highest point candidate region cannot be obtained. In this case, the system automatically reverts to the target feature region selection method, that is, automatically selects a predetermined proportion of the area at the top of the detection box, and filters the highest point of the potential hazard target from this area. The specific implementation process is as follows: Let the width of the detection frame be W and the height be H. Take the upper γ-proportion region within the detection frame, whose height is... γH The width is kept at W. This area corresponds to the upper part of the target bounding box, that is, the highest point area where the potential hazard is most likely to occur (such as the top of the crane boom). By taking the maximum value of the joint saliency map within this area, the final pixel coordinates of the highest point can be obtained. .in As a hyperparameter, simulation experiments determined that a value of 0.10 yielded the best results. The above process can be formally expressed as: ; in, , That is the highest point.

[0071] In this embodiment, when candidate regions exist, the point with the smallest ordinate is selected as the physical highest point; when no candidate regions exist, the method automatically backtracks to the target feature region and searches for the point with the maximum joint saliency map in the upper region of the detection box. This method can effectively filter low-confidence noise while ensuring robustness in locating the highest point in extreme cases.

[0072] S105: Calculate the clearance distance between the highest point and the transmission line to determine the risk level of the potential hazard.

[0073] Specifically, obtain the highest pixel in the image. Then, the system uses the pre-calibrated camera intrinsic parameter matrix. extrinsic rotation matrix Translation vector And the ground elevation model, the pixel coordinates are back-projected onto the camera normalized plane, then transformed to the world coordinate system, and intersected with the ground plane to finally obtain the true 3D coordinates of the highest point. The specific transformation relationship is as follows: ; ; in, The direction vector in the camera coordinate system. Used as world coordinates.

[0074] By combining the equations of the ground plane and solving them simultaneously, the three-dimensional coordinates of the highest point in the world coordinate system can be obtained. .

[0075] After obtaining the three-dimensional coordinates of the highest point in the world coordinate system, the system acquires the parametric representation of the transmission line conductor in space based on a pre-established three-dimensional spatial model of the conductor. The three-dimensional spatial model of the transmission line conductor depends on the design parameters of the transmission line, including span length, sag, and the three-dimensional coordinates of the suspension points at both ends in the world coordinate system. These parameters can be obtained in two ways: directly from the transmission line design drawings, or by using previously acquired laser point cloud data for catenary fitting and calibration. For any point on the conductor... Let its arc length parameter be . Then the three-dimensional coordinates of the point in the world coordinate system = It can be uniquely determined by the catenary equation.

[0076] In this embodiment, the clearance distance Defined as the highest point The shortest Euclidean distance to the conductor is calculated using the following formula: ; In practical engineering, to reduce computational overhead, a method combining discrete sampling and local parabolic fitting is adopted. Specifically, 50–100 points are sampled at equal intervals along the arc length of the conductor, the distance from the highest point to each sampled point is calculated, and the interval containing the minimum value is selected. Then, cubic spline interpolation is performed on the conductor within this interval, and the nearest point is accurately determined using the gradient descent method to obtain the final clearance distance. .

[0077] This embodiment is based on clearance distance. Two levels of risk were identified: If d is less than or equal to a preset safety threshold (e.g., 10 meters), the corresponding hidden danger target is determined to be high-risk, and an emergency alarm message is generated. The emergency alarm message includes: the category of the hidden danger target, the actual coordinates of the highest point, the clearance distance, and the original image, which is uploaded to the monitoring platform via the network.

[0078] If d is greater than the preset safety threshold, the corresponding hidden danger target is determined to be low-risk, and a low-risk alarm message is generated. The low-risk alarm message includes: the category of the hidden danger target and the original image, which is uploaded to the monitoring platform via the network.

[0079] Of course, the specific risk level can also be divided into three or more levels according to actual needs.

[0080] This embodiment determines the risk level of a potential hazard by comparing the calculated clearance distance with a preset threshold. The smaller the clearance distance between the potential hazard and the transmission line, the higher its risk level. For high-risk potential hazards, refined alarm information is generated to help maintenance personnel accurately perceive the risk; for low-risk potential hazards, simplified alarm information is generated for maintenance personnel to refer to.

[0081] As a specific example, Figure 3 The image shown is used as the input image. A target detection model is used to identify potential hazards in the acquired power transmission channel image, outputting detection boxes (green boxes) for three potential hazards and corresponding category confidence heatmaps, as shown below. Figure 4 and Figure 5 As shown; taking the leftmost crane as the subsequent processing object, the LC saliency map of its corresponding detection box is obtained as follows. Figure 6 As shown, the LC saliency map of the detection box is multiplied element-wise with the class confidence heatmap to generate the joint saliency map corresponding to the detection box, as shown below. Figure 7 As shown; based on the statistical distribution of the joint saliency map within the detection box, an adaptive threshold is dynamically calculated. Connectivity analysis is then performed on pixels within the detection box whose joint saliency value is greater than this adaptive threshold, yielding... Figure 8The highest point candidate area is shown; the highest point of the hazard target within the current detection frame is determined from the highest point candidate area based on the principle of minimizing the ordinate, such as... Figure 9 As shown, Figure 9 The red dot in the upper right corner of the green target box represents the highest pixel; the Euclidean distance between the coordinates of this pixel and the manually calibrated coordinates of the highest pixel is less than 5 pixels, meeting the test accuracy requirements; after obtaining the highest point, calculate the clearance distance from the highest point to the conductor as follows: Figure 10 As shown, the final clearance distance was 87 meters, which was determined to be low risk. A low-risk alarm was generated and uploaded to the monitoring platform.

[0082] The method of this embodiment was verified on test images covering typical hidden danger targets such as cranes and tower cranes. The clearance distance obtained by the method of this embodiment was compared with the actual measured value on site. The ranging error was sufficient to meet the alarm judgment requirements under the 10-meter safety threshold, and no additional hardware cost was required. At the same time, the ranging results can provide a quantitative basis for the classification of hidden danger risk levels, which significantly improves the engineering practicality of the monocular vision solution.

[0083] This embodiment is based on a high-point detection method for potential hazards in power transmission channels, which integrates monocular vision and saliency analysis. It can accurately generate the physical coordinates of the highest point without changing the existing hardware. The entire process does not require manual annotation of the highest point data or training of additional segmentation networks. The test results can meet the accuracy requirements of on-site testing, greatly improving the accuracy of risk assessment for potential hazards. It can be run in real time on existing monocular vision monitoring equipment on towers.

[0084] Example 2 In one or more embodiments, a power transmission channel hazard risk assessment system is disclosed, specifically including: The hazard target identification module is configured to use a target detection model to identify hazards in the acquired power transmission channel images and output the detection bounding box of the hazard target; The joint saliency map construction module is configured to fuse the category confidence heatmap and local contrast saliency map of each detection box to obtain the joint saliency map corresponding to each detection box. The highest point detection module is configured to use the mean of the joint saliency map as a benchmark and the product of the standard deviation and a preset adjustment coefficient as a dynamic adjustment term to calculate the adaptive threshold corresponding to the current detection box; perform connected component analysis on the pixels in the detection box whose joint saliency value is greater than the adaptive threshold to obtain the candidate region of the highest point; and determine the highest point of the hidden danger target in the current detection box from the candidate region of the highest point. The hazard risk assessment module is configured to calculate the clearance distance between the highest point and the transmission line to determine the risk level of the hazard target.

[0085] As a further solution, the system also includes: The supplementary positioning module is configured to select a region with a ratio of γ at the top of the detection frame when the candidate region of the highest point cannot be obtained. The height of the region is γ times the height of the detection frame, and the width is the same as the width of the detection frame; where γ is a preset value. The pixel with the largest joint saliency value within the area is selected as the highest point of the potential hazard target within the current detection frame.

[0086] As a further solution, the hazard risk assessment module specifically includes: The coordinate transformation unit is configured to use the pre-calibrated camera intrinsic parameter matrix K, extrinsic parameter rotation matrix R, and translation vector T, combined with the ground elevation model, to back-project the pixel coordinates of the highest point onto the camera normalization plane, then transform it to the world coordinate system, and intersect it with the ground plane to finally obtain the three-dimensional true coordinates of the highest point. The hazard identification unit is configured to calculate the shortest Euclidean distance between the three-dimensional real coordinates of the highest point and the transmission line based on a pre-established three-dimensional spatial model of the transmission line conductor, which is the clearance distance from the hazard target to the transmission line; and determine the risk level of the hazard target based on the clearance distance.

[0087] It should be noted that the specific implementation methods of the above modules are the same as those in Example 1, and will not be described in detail again.

[0088] Example 3 In one or more embodiments, a power transmission channel hazard target monitoring device is disclosed, which includes a processor and a memory. The processor is used to implement instructions; the memory is used to store multiple instructions, which are adapted to be loaded by the processor and executed by the power transmission channel hazard risk assessment method described in Embodiment 1.

[0089] In other embodiments, a chip is disclosed that includes the aforementioned power transmission channel hazard monitoring device.

[0090] It should be understood that in this embodiment, the processor can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any other conventional processor.

[0091] Memory may include read-only memory and random access memory, and provides instructions and data to the processor. A portion of the memory may also include non-volatile random access memory. For example, memory may also store information about the device type.

[0092] In the implementation process, each step of the above method can be completed by the integrated logic circuits in the processor hardware or by software instructions.

[0093] Example 4 In one or more embodiments, a computer-readable storage medium is disclosed, wherein a plurality of instructions are stored, the instructions being adapted to be loaded by a processor of a terminal device and executed by the power transmission channel hazard risk assessment method of Embodiment 1.

[0094] In other embodiments, a computer program product is disclosed, including a computer program or instructions that, when executed by a processor, implement the power transmission channel hazard risk assessment method in Embodiment 1.

[0095] In such an embodiment, the computer program can be downloaded and installed from a network via a communication component, and / or installed from a removable medium. When the computer program is executed by the central processing unit, it performs the various functions defined in the steps of Embodiment 1. Alternatively, the computer program instructions can also be stored in a computer-readable storage medium capable of directing a computer or other programmable data processing device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium implement the working process of Embodiment 1.

[0096] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for assessing potential risks in power transmission channels, characterized in that, include: The target detection model is used to identify potential hazards in the acquired power transmission channel images, and the detection bounding boxes of potential hazards are output. The category confidence heatmap and local contrast saliency map of each detection box are fused to obtain the joint saliency map corresponding to each detection box; Using the mean of the joint saliency map as a benchmark, and the product of the standard deviation and the preset adjustment coefficient as a dynamic adjustment term, the adaptive threshold corresponding to the current detection box is calculated. Connectivity analysis is performed on pixels within the detection box whose joint significance value is greater than the adaptive threshold to obtain candidate regions for the highest point; Determine the highest point of the potential hazard target within the current detection frame from the candidate region of the highest point; Calculate the clearance distance between the highest point and the transmission line to determine the risk level of the potential hazard.

2. The method of claim 1, wherein, The specific process for generating the category confidence heatmap for each detection box is as follows: Obtain the last layer feature map in the target detection model before entering the detection head, and calculate the average gradient of each channel in the feature map as the weight of the corresponding channel; wherein, the average gradient is obtained by global average pooling based on the partial derivative of the target category score at each position in the feature map; Based on the weights, the feature maps of each channel are weighted and summed, and negative values ​​are removed by ReLU activation to obtain the original heatmap. The original heatmap is upsampled to the same size as the original power transmission channel image and normalized to obtain a category confidence heatmap.

3. The method of claim 1, wherein, The calculation process for the local contrast saliency map of each detection box is as follows: The original grayscale values ​​of the detection box are uniformly compressed and quantized, and the weight coefficient of each quantization level after compression is calculated; wherein, the weight coefficient is calculated based on the difference of pixel grayscale values ​​between the target quantization level and other quantization levels; For each pixel within the detection box, the corresponding weight coefficient is taken as the local contrast saliency value of the pixel according to its quantization level; finally, the obtained local contrast saliency map is normalized to obtain the final local contrast saliency map of the detection box.

4. The power transmission corridor risk assessment method of claim 3, wherein, The specific weighting coefficients for each quantization level are as follows: Count the number of pixels at each quantization level within the compressed detection box; For a target quantization level q, iterate through each quantization level r, calculate the sum of the products of the absolute difference between the center grayscale of the target quantization level q and the quantization level r and the number of pixels in the corresponding quantization level r, and use this sum as the weight coefficient of the target quantization level q.

5. The method of claim 3, wherein, The original grayscale values ​​of the detection box are uniformly compressed and quantized, specifically as follows: The original grayscale levels from 0 to 255 are uniformly compressed and quantized to 0 to 31.

6. The method of power transmission right-of-way hazard risk assessment of claim 1, wherein, The category confidence heatmap and local contrast saliency map of each detection box are fused together, specifically as follows: multiply the heat value of each pixel in the category confidence heat map of the detection box i with the saliency value of the corresponding pixel in the local contrast saliency map of the detection box i to obtain the joint saliency value corresponding to each pixel in the detection box i , and further obtain the joint saliency map corresponding to the detection box i .

7. The power transmission right-of-way liability risk assessment method of claim 1, wherein, Connectivity analysis is performed on pixels within the detection box whose joint saliency value is greater than the adaptive threshold to obtain candidate regions for the highest point, specifically: All pixels within the detection box whose joint saliency value is greater than the adaptive threshold constitute a binary mask. Connectivity analysis is performed on the binary mask to remove isolated pixels, and one or more connected regions are obtained as candidate regions for the highest point.

8. The power transmission right-of-way liability risk assessment method of claim 1, wherein, The highest point of the potential hazard target within the current detection frame is determined from the candidate region of the highest point, specifically as follows: Using the top left corner of the image as the origin, select the pixel with the smallest ordinate from all candidate regions of the highest point. If there is only one pixel with the smallest ordinate, then that pixel is taken as the highest point of the hazard target in the current detection box; if there are multiple pixels with the smallest ordinate, then the pixel with the largest joint significance value is taken as the highest point of the hazard target in the current detection box.

9. The power transmission right-of-way liability risk assessment method of claim 1, wherein, Also includes: If the highest point candidate region cannot be obtained, then a region with a ratio of γ is selected at the top of the detection frame. The height of the region is γ times the height of the detection frame, and the width is the same as the width of the detection frame; where γ is a preset value. The pixel with the largest joint saliency value within the area is selected as the highest point of the potential hazard target within the current detection frame.

10. The method for assessing potential hazards in power transmission channels as described in claim 1, characterized in that, Calculate the clearance distance between the highest point and the transmission line to determine the risk level of the potential hazard, specifically as follows: Using the pre-calibrated camera intrinsic parameter matrix K, extrinsic parameter rotation matrix R, and translation vector T, combined with the ground elevation model, the pixel coordinates of the highest point are back-projected onto the camera normalization plane, then transformed to the world coordinate system, and intersected with the ground plane to finally obtain the true three-dimensional coordinates of the highest point. Based on the pre-established three-dimensional spatial model of the transmission line conductor, the three-dimensional real coordinates of the highest point and the shortest Euclidean distance of the transmission line are calculated, which is the clearance distance from the hidden danger target to the transmission line. Based on the aforementioned clearance distance, the risk level of the potential hazard is determined.

11. A risk assessment system for hidden dangers in power transmission channels, characterized in that, include: The hazard target identification module is configured to use a target detection model to identify hazards in the acquired power transmission channel images and output the detection bounding box of the hazard target; The joint saliency map construction module is configured to fuse the category confidence heatmap and local contrast saliency map of each detection box to obtain the joint saliency map corresponding to each detection box. The highest point detection module is configured to use the mean of the joint saliency map as a benchmark and the product of the standard deviation and the preset adjustment coefficient as a dynamic adjustment term to calculate the adaptive threshold corresponding to the current detection box. Connectivity analysis is performed on pixels within the detection box whose joint significance value is greater than the adaptive threshold to obtain candidate regions for the highest point; Determine the highest point of the potential hazard target within the current detection frame from the candidate region of the highest point; The hazard risk assessment module is configured to calculate the clearance distance between the highest point and the transmission line to determine the risk level of the hazard target.

12. The power transmission channel hidden danger risk assessment system as described in claim 11, characterized in that, Also includes: The supplementary positioning module is configured to select a region with a ratio of γ at the top of the detection frame when the candidate region of the highest point cannot be obtained. The height of the region is γ times the height of the detection frame, and the width is the same as the width of the detection frame; where γ is a preset value. The pixel with the largest joint saliency value within the area is selected as the highest point of the potential hazard target within the current detection frame.

13. The power transmission channel hidden danger risk assessment system as described in claim 11, characterized in that, The hazard risk assessment module includes: The coordinate transformation unit is configured to use the pre-calibrated camera intrinsic parameter matrix K, extrinsic parameter rotation matrix R, and translation vector T, combined with the ground elevation model, to back-project the pixel coordinates of the highest point onto the camera normalization plane, then transform it to the world coordinate system, and intersect it with the ground plane to finally obtain the three-dimensional true coordinates of the highest point. The hazard identification unit is configured to calculate the shortest Euclidean distance between the three-dimensional real coordinates of the highest point and the transmission line based on a pre-established three-dimensional spatial model of the transmission line conductor, which is the clearance distance from the hazard target to the transmission line; and determine the risk level of the hazard target based on the clearance distance.

14. A monitoring device for potential hazards in power transmission channels, comprising a processor and a memory, wherein the processor is used to implement instructions; and the memory is used to store multiple instructions, characterized in that, The instructions are adapted to be loaded by a processor and executed by the power transmission channel hazard risk assessment method according to any one of claims 1-10.

15. A chip, characterized in that, Includes the power transmission channel hidden danger target monitoring equipment as described in claim 14.

16. A computer-readable storage medium storing a plurality of instructions, characterized in that, The instructions are adapted to be loaded by the processor of the terminal device and executed by the power transmission channel hidden danger risk assessment method according to any one of claims 1-10.

17. A computer program product comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed by the processor, they implement the power transmission channel hazard risk assessment method according to any one of claims 1-10.

Images