High-altitude operation risk early warning method and system based on camera image recognition

Abstract

The invention provides a high-altitude operation risk early warning method and system based on camera image recognition, and relates to the technical field of computer vision, and the method comprises the steps: firstly collecting a video image sequence of a high-altitude operation scene, and generating a fusion feature map containing environment and operation main body features through multi-level feature extraction; performing spatial dimension segmentation and regional feature comparative analysis on the fusion feature map to obtain a spatial risk distribution map containing risk region identification information, processing the spatial risk distribution map of continuous frames based on a time sequence feature fusion rule to generate a dynamic risk evolution map, and calling a risk decision model to perform mode recognition to obtain a dynamic risk evolution map; and generating a risk level classification result and a risk position coordinate set according to the risk level classification result and the risk position coordinate set, and finally generating a risk early warning signal and sending the risk early warning signal to the monitoring terminal, thereby comprehensively, accurately and dynamically monitoring the high-altitude operation risk, and improving the accuracy and timeliness of risk early warning.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and more specifically, to a method and system for high-altitude operation risk warning based on camera image recognition. Background Technology

[0002] In the field of high-altitude operations, ensuring the safety of workers is paramount. As high-altitude work scenarios become increasingly complex, the risk factors encountered during operations are also becoming more diverse, such as severe weather conditions, unstable working environments, and worker violations. Currently, monitoring of risks in high-altitude operations mainly relies on manual inspections and data collection from some fixed sensors. Manual inspections suffer from low efficiency and difficulty in providing real-time, comprehensive coverage of the work area, making it impossible to promptly identify potential risks. While fixed sensors can acquire some environmental data, they cannot comprehensively and accurately monitor the behavior of the workers or the dynamic interaction between the environment and the workers. Furthermore, most existing risk warning methods are based on single-dimensional data analysis, lacking the fusion and dynamic analysis of multi-dimensional data. This results in insufficient accuracy and timeliness of risk warnings, making it difficult to effectively prevent high-altitude work accidents and posing a serious threat to the lives of workers. Summary of the Invention

[0003] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide a high-altitude operation risk early warning method based on camera image recognition, the method comprising: A video image sequence of a high-altitude operation scene is acquired, and multi-level feature extraction processing is performed on the video image sequence to generate a fused feature map containing environmental feature set and operation subject feature set; The fused feature map is segmented spatially to obtain multiple local region feature sub-maps. Regional feature comparison analysis is performed on each local region feature sub-map to generate a spatial risk distribution map containing risk area identification information. Based on preset temporal feature fusion rules, temporal feature fusion processing is performed on the spatial risk distribution map of consecutive frames in the video image sequence to generate a dynamic risk evolution map. The risk decision-making model is invoked to perform pattern recognition processing on the dynamic risk evolution map, generating a risk level classification result and a set of risk location coordinates; A risk warning signal is generated based on the risk level classification result and the risk location coordinate set, and the risk warning signal is sent to the monitoring terminal.

[0004] In another aspect, embodiments of the present invention also provide a high-altitude operation risk warning system based on camera image recognition, including a processor and a machine-readable storage medium connected to the processor. The machine-readable storage medium is used to store programs, instructions, or code, and the processor is used to execute the programs, instructions, or code in the machine-readable storage medium to implement the above-described method.

[0005] Based on the above, this embodiment of the invention collects video image sequences of high-altitude operation scenarios and performs multi-level feature extraction processing to generate a fused feature map that integrates environmental features and the features of the operation subject. The fused feature map is then spatially segmented and regional feature comparison analysis is performed to generate a spatial risk distribution map containing risk area identification information. This allows for precise location of potential risk areas in the operation scenario. Based on preset temporal feature fusion rules, the spatial risk distribution map of consecutive frames is processed to generate a dynamic risk evolution map, effectively capturing the trend of risk changes over time. A risk decision model is invoked to perform pattern recognition on the dynamic risk evolution map, generating risk level classification results and a set of risk location coordinates. This achieves accurate risk classification and location. Finally, based on this information, a risk warning signal is generated and sent to the monitoring terminal, enabling monitoring personnel to promptly understand the risk status of the operation site and take corresponding measures. This greatly improves the accuracy and timeliness of high-altitude operation risk warnings, effectively reduces the probability of accidents, and protects the lives of operators. Attached Figure Description

[0006] Figure 1 This is a schematic diagram of the execution flow of the high-altitude operation risk warning method based on camera image recognition provided in an embodiment of the present invention.

[0007] Figure 2 This is a schematic diagram of exemplary hardware and software components of a high-altitude operation risk warning system based on camera image recognition provided in an embodiment of the present invention. Detailed Implementation

[0008] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 This is a flowchart illustrating a high-altitude operation risk warning method based on camera image recognition, according to an embodiment of the present invention. The high-altitude operation risk warning method based on camera image recognition will be described in detail below.

[0009] Step S110: Collect video image sequences of high-altitude operation scenarios, perform multi-level feature extraction processing on the video image sequences, and generate a fused feature map containing environmental feature sets and operation subject feature sets.

[0010] In high-altitude work scenarios, to accurately and comprehensively monitor the risks during the operation, it is necessary to collect relevant video image sequences. This can be achieved, for example, by strategically deploying multiple image acquisition devices in the work area. These devices should possess wide-angle shooting and high-definition imaging capabilities, covering the main areas of the high-altitude work and ensuring that the acquired video image sequences contain sufficient information. The acquired video image sequence is a series of continuous image frames, each containing specific instantaneous information about the high-altitude work scenario.

[0011] Next, multi-level feature extraction processing is performed on the video image sequence, specifically including the following steps: Step S111: Perform edge enhancement and morphological closing operations on single-frame images in the video image sequence to generate an enhanced target region segmentation map.

[0012] For each single frame in the acquired video image sequence, edge enhancement processing is performed first. Edge enhancement can highlight the edge information of objects in the image. In high-altitude operation scenarios, the outlines of workers' bodies, safety equipment, and key objects in the surrounding environment can be more clearly displayed through edge enhancement. This helps to more accurately identify and segment target areas subsequently.

[0013] Edge enhancement can be achieved using various methods, such as using gradient operators (e.g., Sobel operator, Prewitt operator) to calculate the image gradient; regions with larger gradients are considered edge regions. Assume a single-frame image is I(x, y), where x and y represent the coordinates of pixels in the image. When using the Sobel operator for edge enhancement, the gradients Gx and Gy in the x and y directions are calculated respectively. Then, the gradient magnitude G is calculated using the formula G = √(Gx² + Gy²), yielding the enhanced image I_edge(x, y).

[0014] Next, morphological closing operations are performed. Morphological closing is an image processing method based on morphological operations, consisting of dilation and erosion operations. Dilation expands the object region in an image, while erosion shrinks it. First, dilation is performed on the edge-enhanced image to extend the object's edges outward, filling in small holes and gaps, making the object region more continuous and complete. Then, erosion is performed to remove some of the excess parts caused by dilation, restoring the object's approximate shape. Morphological closing removes noise and small interfering objects from the image while preserving the main features of the object. Assuming the structuring element of dilation is SE1 and the structuring element of erosion is SE2, the edge-enhanced image I_edge(x, y) is first dilated to obtain I_dilate(x, y), and then eroded to obtain I_close(x, y), ultimately generating the enhanced target region segmentation map.

[0015] Step S112: Based on the preset safety equipment semantic segmentation model, perform multi-category region division processing on the target region segmentation map to separate the seat belt wearing region, the self-locking device connection region and the speed difference device suspension region.

[0016] The pre-defined semantic segmentation model for security equipment is a deep learning model trained on a large amount of labeled data, such as semantic segmentation models based on convolutional neural networks (CNNs), like U-Net and DeepLab. The function of this semantic segmentation model is to perform multi-category region division on the enhanced target region segmentation map.

[0017] The target region segmentation map is input into a pre-defined safety equipment semantic segmentation model. The model classifies and labels different regions in the image based on image features and learned patterns. In high-altitude work scenarios, the focus is on the safety belt wearing area, the self-locking device connection area, and the speed-differential device suspension area. The safety belt wearing area is where the worker wears the safety belt; the self-locking device connection area is where the self-locking device connects to the safety belt or other fixed points; and the speed-differential device suspension area is where the speed-differential device is suspended. The safety equipment semantic segmentation model accurately separates these regions from the image based on their features and feature patterns learned from the training data. For example, the model identifies the safety belt wearing area based on the color, shape, and texture features of the safety belt; determines the self-locking device connection area based on the unique appearance and connection method of the self-locking device; and finds the speed-differential device suspension area based on the shape and suspension position features of the speed-differential device. The final output is a multi-category region segmentation result, dividing the target region segmentation map into the safety belt wearing area, the self-locking device connection area, the speed-differential device suspension area, and other background areas.

[0018] Step S113: Perform anchor point coordinate detection processing on the seat belt wearing area to extract the spatial position parameters of the main anchor point and auxiliary anchor point of the seat belt.

[0019] After separating the seatbelt wearing area, it is necessary to perform anchor point coordinate detection processing on this area to extract the spatial position parameters of the main anchor point and auxiliary anchor point of the seatbelt. The main anchor point is the primary location for fixing the seatbelt, while the auxiliary anchor point is the location that provides auxiliary fixing.

[0020] Anchor point coordinate detection can be performed using a method based on feature matching and geometric calculation. First, points with distinctive features are identified within the seatbelt wearing area, such as the connection points between the seatbelt and a fixed object, and the buckle positions. These points can serve as candidate anchor points. Then, based on the seatbelt's structure and usage rules, the characteristics and positional relationships of the primary and secondary anchor points are determined. For example, the primary anchor point is typically the most important connection between the seatbelt and the fixed structure, while the secondary anchor points are relatively less important. By filtering and evaluating the candidate anchor points, the primary and secondary anchor points of the seatbelt are determined.

[0021] After determining the main and auxiliary anchor points, their spatial position parameters need to be extracted. This can be achieved using the image's coordinate system. Assume the image's origin is at the top left corner, the x-axis is to the right, and the y-axis is downwards. For the main and auxiliary anchor points, record their pixel coordinates (x1, y1) and (x2, y2) in the image, respectively. This coordinate information can be further converted into actual spatial position parameters. For example, using the camera's calibration parameters and known scene geometry, the pixel coordinates can be converted into coordinates in three-dimensional space, thus obtaining the accurate spatial position parameters of the seatbelt's main and auxiliary anchor points.

[0022] Step S114: Perform bi-branch contour analysis on the self-locking device connection area to separate the self-locking device body contour and the connecting rope contour, and calculate the minimum bounding box center distance parameter of the self-locking device body contour and the connecting rope contour.

[0023] For the self-locking device connection area, a bi-branch contour analysis method is used to separate the contour of the self-locking device body from the contour of the connecting rope. The bi-branch contour analysis can be performed based on the edge information of the image and the shape features of the object.

[0024] First, edge detection is performed on the image of the self-locking device's connection area to obtain its edge information. Then, based on the shape characteristics of the self-locking device and the connecting rope, the edge information is classified and grouped. The self-locking device typically has a relatively regular shape, such as a rectangle or circle, while the connecting rope exhibits a slender, linear characteristic. Through analysis and judgment of the edge information, the edges belonging to the self-locking device body and the edges belonging to the connecting rope can be extracted separately, thus separating the contours of the self-locking device body and the connecting rope.

[0025] After separating the contours, it is necessary to calculate the minimum bounding box center distance parameter between the self-locking device body contour and the connecting rope contour. The minimum bounding box is the smallest rectangle that can completely enclose the object contour. For the self-locking device body contour and the connecting rope contour, their minimum bounding boxes are calculated separately. Assuming the center coordinates of the minimum bounding box of the self-locking device body contour are (x3, y3) and the center coordinates of the minimum bounding box of the connecting rope contour are (x4, y4), then by calculating the distance between the two points using the formula d=√((x4-x3)²+(y4-y3)²), the minimum bounding box center distance parameter between the self-locking device body contour and the connecting rope contour can be obtained, which can reflect the connection state and relative positional relationship between the self-locking device and the connecting rope.

[0026] Step S115: Perform suspension point trajectory prediction processing on the suspension area of ​​the speed difference device, and extract the instantaneous position coordinates of the suspension device in the current frame and its offset angle parameters with respect to the preset baseline.

[0027] The purpose of predicting the suspension point trajectory in the suspension area of ​​the speed difference device is to understand the motion state and positional changes of the suspension device. Suspension point trajectory prediction can be achieved based on the continuity of video image sequences and the motion patterns of objects.

[0028] First, the suspension area of ​​the speed difference device is tracked in consecutive video image frames. Target tracking algorithms, such as KCF (Kernelized Correlation Filters) and CSRT (Discriminative Correlation Filter with Chanel and Spatial Reliability), can be used to accurately track the position of the suspension device. By tracking multiple frames, the position information of the suspension device at different times can be obtained.

[0029] In the current frame, extract the instantaneous position coordinates of the suspension device. Similarly, based on the image's coordinate system, record the pixel coordinates (x5, y5) of the suspension device in the current frame. Then, convert them to actual spatial position coordinates.

[0030] Next, the offset angle parameter between the suspension device and the preset baseline is calculated. The preset baseline is a reference line pre-set according to the safety requirements and scene characteristics of high-altitude operations, such as a vertical baseline. The offset angle parameter can be obtained by calculating the angle between the instantaneous position of the suspension device and the preset baseline. Assuming the equation of the preset baseline is y=kx+b, and the instantaneous position coordinates of the suspension device are (x5, y5), the offset angle parameter between the suspension device and the preset baseline can be obtained through geometric calculation methods, such as vector angle calculation methods. This parameter reflects the stability of the suspension device and whether it meets safety requirements.

[0031] Step S116: Perform channel-specific normalization processing on the spatial position parameters of the main anchor point and auxiliary anchor point of the safety belt, the center distance parameters of the self-locking device body and the connecting rope, and the offset angle parameters of the suspension device to generate a dynamic feature matrix of the safety equipment.

[0032] To facilitate subsequent feature fusion and analysis, it is necessary to perform channel-specific normalization on the extracted spatial position parameters of the main and auxiliary anchor points of the safety belt, the center distance parameters between the self-locking device and the connecting rope, and the offset angle parameters of the suspension device.

[0033] Channel normalization is a process that performs a uniform scaling transformation on different types of parameters, ensuring they have the same value range. For the spatial position parameters of the main and auxiliary anchor points of the seatbelt, assuming the main anchor point's spatial position parameters are (x1, y1, z1) and the auxiliary anchor point's spatial position parameters are (x2, y2, z2), normalization can be performed on the x, y, and z channels respectively. For example, for the x channel, a min-max normalization method can be used to map the values ​​of x1 and x2 to the range [0, 1]. The formula is x_normalized = (x - x_min) / (x_max - x_min), where x_min and x_max are the minimum and maximum values ​​of the x channel, respectively.

[0034] The center distance parameter d between the self-locking device body and the connecting rope is also normalized and mapped to the range [0, 1]. A similar normalization method is used for the offset angle parameter θ of the suspension device.

[0035] After channel-specific normalization, these normalized parameters are combined into a matrix, namely the safety equipment dynamic characteristic matrix. This safety equipment dynamic characteristic matrix contains dynamic characteristic information of safety equipment such as seat belts, self-locking devices, and speed differential devices.

[0036] Step S117: Perform multi-level feature channel fusion processing on the dynamic feature matrix of the safety equipment to generate the main feature set of the operation subject, which includes the distribution of seat belt anchor points, the connection status of self-locking devices and the offset of speed difference devices. Then, perform cross-source splicing with the ambient light gradient parameters extracted by background segmentation to generate the fused feature map.

[0037] Multi-level feature channel fusion processing is performed on the dynamic feature matrix of safety equipment to generate a feature set of the main operation. Various methods can be used for multi-level feature channel fusion processing, such as weighted averaging and feature fusion layers of convolutional neural networks.

[0038] Through multi-level feature channel fusion processing, features from different channels in the dynamic feature matrix of safety equipment are fused to obtain a set of operational features including the distribution of seat belt anchor points, the connection status of the self-locking device, and the offset of the speed-differential device. The distribution of seat belt anchor points reflects the spatial positional relationship and distribution of the main and auxiliary anchor points of the seat belt; the connection status of the self-locking device is reflected by the center distance parameter between the self-locking device body and the connecting rope; and the offset of the speed-differential device is represented by the offset angle parameter of the suspension device.

[0039] Simultaneously, ambient illumination gradient parameters are extracted through background segmentation. Background segmentation can employ image segmentation algorithms such as GrabCut and MeanShift to separate the background and foreground in the image. Then, illumination analysis is performed on the background region to calculate illumination gradient parameters. Illumination gradient parameters reflect changes in ambient illumination and are crucial for assessing environmental conditions during high-altitude operations.

[0040] Finally, the feature set of the task subject and the ambient lighting gradient parameters are concatenated across sources. Cross-source concatenation merges features from different data sources to form a more comprehensive feature representation. The feature set of the task subject and the ambient lighting gradient parameters are concatenated according to a set rule to generate a fused feature map, which contains feature information of both the task subject and the environment.

[0041] Step S120: Perform spatial dimension segmentation processing on the fused feature map to obtain multiple local region feature sub-maps, and perform regional feature comparison analysis processing on each local region feature sub-map to generate a spatial risk distribution map containing risk area identification information.

[0042] After obtaining the fused feature map, spatial segmentation is required to analyze the risks in high-altitude work scenarios in more detail. Spatial segmentation divides the fused feature map into multiple local regions, each corresponding to a specific part of the high-altitude work scenario.

[0043] Spatial dimensional segmentation can be performed using either uniform segmentation or adaptive segmentation. Uniform segmentation divides the fused feature map into regions of fixed size and shape, such as dividing it into rectangular regions of equal size. Adaptive segmentation, on the other hand, divides the map based on the feature distribution of the fused feature map and the actual conditions of the scene, ensuring that each local region has similar features or corresponds to a meaningful part of the scene.

[0044] After spatial dimension segmentation, multiple local region feature sub-maps are obtained. Each local region feature sub-map contains feature information of that local region, which is a subset of the features of the corresponding region in the fused feature map.

[0045] Next, a regional feature comparison analysis will be performed on each local feature sub-image, as detailed in the following steps: Step S121: Extract the location coordinates of the safety equipment wearing area of ​​the operator in the feature sub-image of the local area, and generate the safety belt anchor point parameters, climbing self-locking device connection point parameters, and speed difference device suspension point parameters.

[0046] In this embodiment, the local region feature sub-map is a localized feature carrier obtained by spatially segmenting the fused feature map. Each sub-map corresponds to a specific spatial range in the high-altitude operation scenario, which may cover the torso area, limb movement area, or installation and use area of ​​safety equipment of the worker. Since the separation of key areas of safety equipment and the preliminary extraction of feature parameters have been completed in the previous multi-level feature extraction process (steps S111-S117), this step needs to further accurately locate the specific coordinates of the area where the worker's safety equipment is worn based on the spatial focus of the local region, so as to provide refined local feature parameters for subsequent risk comparison analysis.

[0047] First, for the local feature sub-image, it is necessary to first identify the core activity area of ​​the workers. By combining the extracted features of the main work subject in the sub-image (such as human body contour edge features, color and texture features of safety equipment), the sub-image can be locally spatially aligned with the previously generated target area segmentation image (step S111) through a region matching algorithm, so as to clarify the relative position range of the workers in the sub-image, and then define the approximate spatial boundary of the area where the safety equipment is worn.

[0048] To extract the coordinates of the safety belt wearing area, the coordinate range needs to be further refined in the local feature sub-image based on the safety belt wearing area features separated in step S112. Pixel-level coordinate positioning determines the coverage area of ​​the safety belt on the worker's torso, including key contact points of the shoulder straps and waist belt. Using the image coordinate system (origin at the top left corner of the sub-image, x-axis to the right, y-axis downwards) as a reference, the set of pixel coordinates of the boundary of this area is recorded. Based on this, and combining the spatial position parameters of the main and auxiliary anchor points of the safety belt extracted in step S113, the specific pixel coordinates of the main anchor point (usually the core node connecting the safety belt to the fixed structure, such as the back buckle connection point) and auxiliary anchor points (such as spare connection points on both sides of the waist) are accurately identified in the sub-image. These coordinates are then integrated to generate safety belt anchor point parameters containing the coordinates of the main anchor point (xa1, ya1), the coordinates of the auxiliary anchor point (xa2, ya2), and the anchor point spacing.

[0049] During the generation of climbing self-locking device connection point parameters, the physical location of the climbing self-locking device needs to be located in the local region feature sub-map based on the self-locking device connection area features separated in step S114. Since the self-locking device is usually connected between the worker's safety harness and the climbing rope, the outline boundary of the self-locking device body needs to be identified in the sub-map first, and then the positions of the tethered end (the end connected to the climbing rope) and the human connection end (the end connected to the D-ring of the safety harness) need to be further distinguished. Feature points at both ends are extracted using an edge detection algorithm, and the pixel coordinates (xb1, yb1) of the tethered end and the pixel coordinates (xb2, yb2) of the human connection end in the sub-map are recorded. At the same time, the straight-line distance between the two ends in the sub-map plane is calculated to form climbing self-locking device connection point parameters that include the coordinates of both ends and the initial spacing.

[0050] For the suspension point parameters of the speed-differential device, it is necessary to combine the suspension area features of the speed-differential device obtained in step S115 with the instantaneous position coordinates of the suspension device to accurately locate the suspension point of the speed-differential device in the local area feature sub-map. The suspension point of the speed-differential device is usually fixed to a stable structure above the working area (such as scaffold crossbars or steel beams). It is necessary to identify the contour features of the suspension device (such as hooks or rings) in the sub-map, record the center point position (xc, yc) of the suspension point in pixel coordinates, and extract the local features of the fixed structure around the suspension point (such as structural edges and colors) as auxiliary verification information to ensure the accuracy of the suspension point positioning. Finally, the suspension point coordinates and auxiliary verification features are integrated to generate the suspension point parameters of the speed-differential device.

[0051] Step S122: Perform spatial projection matching processing on the anchor point parameters of the seat belt and the preset human torso reference coordinate system, and calculate the contour overlap between the fixed area of ​​the seat belt and the reference area.

[0052] The preset human torso reference coordinate system is a coordinate system pre-set according to the structure and posture of the human body. It can be used as a reference frame to evaluate whether the seat belt is fixed in the correct position.

[0053] The seatbelt anchor point parameters are spatially projected and matched with a preset human torso reference coordinate system. This spatial projection matching process involves projecting the spatial position of the seatbelt anchor point onto the human torso reference coordinate system, and then comparing the positional relationship between the seatbelt fixing area and the reference area.

[0054] First, the coordinates of the seatbelt anchor point need to be transformed from the image coordinate system to the human torso reference coordinate system. This can be achieved using a coordinate transformation matrix, which can be calculated based on the camera's calibration parameters and the human's posture information.

[0055] After coordinate transformation, calculate the overlap between the seatbelt fixing area and the reference area. The overlap is the ratio of the area of ​​the overlapping portion to the total area of ​​both areas. This overlap can be obtained by calculating the polygonal areas of the two areas and the area of ​​the overlapping portion. A higher overlap indicates that the seatbelt's fixing position is more in line with safety requirements; conversely, a lower overlap may indicate a safety hazard.

[0056] Step S123: Perform dual-branch feature analysis processing on the connection point parameters of the climbing self-locking device, separate the position coordinates of the rope end and the position coordinates of the human connection end, and calculate the offset ratio of the distance between the two ends to the standard safety distance.

[0057] A two-branch feature analysis process is performed on the connection point parameters of the climbing self-locking device to separate the position coordinates of the tether end and the human connection end. This two-branch feature analysis can be based on the structure and connection method of the climbing self-locking device.

[0058] Climbing self-locking devices typically have a tether end and a body connection end. The tether end is used to connect to the rope, and the body connection end is used to connect to the worker's safety harness or other equipment. By analyzing and processing the connection point parameters of the climbing self-locking device, they can be divided into tether end position coordinates and body connection end position coordinates based on the location and characteristics of the connection points.

[0059] After separating the position coordinates of the two ends, calculate the offset ratio of the distance between the two ends to the standard safety distance. The distance between the two ends refers to the distance between the position coordinates of the rope end and the position coordinates of the human connection end, which can be obtained by calculating the Euclidean distance between the two points. The standard safety distance is a distance value preset according to the safety regulations and requirements for high-altitude operations.

[0060] The offset ratio refers to the ratio of the difference between the distance between the two ends and the standard safety distance to the standard safety distance. By calculating the offset ratio, it can be determined whether the connection status of the climbing self-locking device meets safety requirements. If the offset ratio is small, it indicates that the distance between the two ends is close to the standard safety distance, and the connection status of the climbing self-locking device is good; if the offset ratio is large, there may be safety hazards, and timely adjustment is required.

[0061] Step S124: Input the suspension point parameters of the speed difference device into the three-dimensional spatial trajectory prediction model to generate the dynamic movement trajectory of the suspension device, and calculate the vertical projection deviation between the trajectory end point and the preset high-hanging and low-use safety interval.

[0062] The three-dimensional trajectory prediction model is a model built based on machine learning or deep learning algorithms to predict the motion trajectory of the suspension point of the speed difference device in three-dimensional space. This model learns from a large amount of historical data, including information such as the position, velocity, and acceleration of the suspension point, to establish the trajectory prediction model.

[0063] The suspension point parameters of the speed difference device are input into a three-dimensional spatial trajectory prediction model. Based on the input parameters and the learned motion laws, the model generates a dynamic movement trajectory of the suspension device. The dynamic movement trajectory is a three-dimensional spatial coordinate sequence containing multiple time points, reflecting the movement path of the suspension device over a period of time.

[0064] The preset high-mounted, low-use safety zone is a three-dimensional spatial area pre-defined according to the safety regulations and requirements for high-altitude operations. This zone specifies the range within which the suspension point of the speed difference device should be located during normal operation. After generating the dynamic movement trajectory of the suspension device, it is necessary to calculate the vertical projection deviation between the trajectory's end point and the preset high-mounted, low-use safety zone.

[0065] First, the endpoint of the trajectory is projected onto the plane containing the preset high-mounted, low-used safety zone, resulting in a projection point. Then, the shortest distance between the projection point and the boundary of the safety zone is calculated; this shortest distance is the vertical projection deviation. The vertical projection deviation reflects whether the movement trajectory of the suspension device exceeds the safe range. If the deviation is large, it indicates a potential safety risk to the suspension device.

[0066] Step S125: Standardize the contour overlap, offset ratio and vertical projection deviation to generate a standardized set of safety equipment matching parameters.

[0067] To facilitate comprehensive analysis and comparison of contour overlap, offset ratio, and vertical projection deviation, these parameters need to be standardized. Standardization involves converting parameters with different ranges and dimensions into a unified range and dimension, making them comparable.

[0068] Different standardization methods can be used for contour overlap, offset ratio, and vertical projection deviation. For example, for contour overlap, its value range can be mapped to [0, 1], so that the higher the overlap, the closer the standardized value is to 1; for offset ratio, its value range can be mapped to [0, 1], so that the smaller the offset ratio, the closer the standardized value is to 1; for vertical projection deviation, its value range can be mapped to [0, 1], so that the smaller the deviation, the closer the standardized value is to 1.

[0069] After standardization, the contour overlap, offset ratio, and vertical projection deviation are combined into a parameter set, namely the standardized safety equipment matching parameter set, which can comprehensively reflect the usage and matching degree of safety equipment.

[0070] Step S126: Input the standardized safety equipment matching parameter group into the multi-level risk decision tree model to generate risk area identification information including seat belt failure identification, self-locking device misalignment identification and speed difference violation identification, and map the risk area identification information to the spatial grid coordinate system of the local area feature sub-map to generate the spatial risk distribution map.

[0071] The multi-level risk decision tree model is a model built on the decision tree algorithm, used for risk assessment and decision-making based on input parameters. This model establishes risk decision rules by learning from a large amount of historical data, including sets of security equipment matching parameters and corresponding risk situations.

[0072] The standardized safety equipment matching parameter set is input into a multi-level risk decision tree model. This model can then assess and classify risks based on the input parameters and learned decision rules. The generated risk area identification information includes seatbelt failure indicators, self-locking device misalignment indicators, and speed difference violation indicators. These risk area indicators clearly indicate the type of risk present in that local area.

[0073] After generating risk area identification information, it is mapped to the spatial grid coordinate system of the local area feature submap. The spatial grid coordinate system is a coordinate system that divides the local area feature submap into multiple small grids, each grid corresponding to a specific spatial location. By mapping the risk area identification information to the spatial grid coordinate system, the risk situation at each location in the local area can be visually displayed.

[0074] Finally, a spatial risk distribution map is generated based on the mapping results. The spatial risk distribution map graphically displays the risk type and risk level at various locations within a local area.

[0075] Step S130: Based on the preset temporal feature fusion rules, perform temporal feature fusion processing on the spatial risk distribution map of consecutive frames in the video image sequence to generate a dynamic risk evolution map.

[0076] During high-altitude operations, the risks may change over time. To gain a more comprehensive understanding of the risk evolution process, it is necessary to perform temporal feature fusion processing on the spatial risk distribution maps of consecutive frames in a video image sequence.

[0077] The pre-defined time-series feature fusion rules are rules pre-set based on the evolutionary patterns and characteristics of risks, used to guide the fusion process of time-series features. These rules can consider multiple factors, such as the continuity of risks, changing trends, and propagation speed.

[0078] First, time-series features are extracted from the spatial risk distribution map of consecutive frames. Time-series features refer to the changes in risk characteristics at different points in time, such as changes in the size, location, and type of risk areas. These time-series features can be extracted by comparing and analyzing the spatial risk distribution map of consecutive frames.

[0079] Then, according to the preset time series feature fusion rules, the extracted time series features are fused. Various methods can be used for fusion processing, such as weighted average and dynamic programming. Through fusion processing, risk features at different time points can be integrated to obtain a comprehensive risk feature representation.

[0080] Finally, a dynamic risk evolution map is generated based on the fused risk characteristics. This map graphically illustrates the evolution of risk over time, including its occurrence, development, spread, and decline. It helps monitoring personnel understand risk trends and take appropriate measures for risk warning and control.

[0081] Step S131: Extract the time series features from the spatial risk distribution map of consecutive frames to generate the risk state change trajectory.

[0082] To generate a risk state change trajectory, time-series features need to be extracted from the spatial risk distribution map of consecutive frames. These time-series features reflect the state changes of the risk at different points in time.

[0083] First, the spatial risk distribution map of consecutive frames is analyzed frame by frame. For each frame's spatial risk distribution map, key information is extracted, such as the location, size, and type of risk area. This key information constitutes the risk status characteristics of that frame.

[0084] Then, the risk state features of consecutive frames are arranged in chronological order to form a time series. In this time series, each time point corresponds to a risk state feature.

[0085] Next, the time series is processed to generate a risk state change trajectory. Interpolation, fitting, and other methods can be used to smooth the time series, making the risk state change trajectory more continuous and accurate. The risk state change trajectory can be represented by a multi-dimensional vector sequence, where each vector corresponds to the risk state at a given time point.

[0086] Step S132: Perform motion vector analysis on the trajectory of the risk state change to generate a risk diffusion direction prediction vector.

[0087] Motion vector analysis is a method used to analyze the direction and velocity of an object's motion. In this step, it is applied to the trajectory of risk state changes to generate a risk diffusion direction prediction vector.

[0088] First, the trajectory of risk state changes is differentially processed. Differential processing calculates the difference in risk state characteristics between adjacent time points to obtain the amount of change in risk state. Through differential processing, the risk state change vector at each time point can be obtained.

[0089] Then, statistical analysis is performed on these risk state change vectors to calculate their average value and direction. The average value reflects the overall trend of risk state changes, and the direction indicates the direction of risk diffusion.

[0090] Finally, based on the statistical analysis results, a risk diffusion direction prediction vector is generated. This vector is a vector with direction and magnitude, representing the potential direction and trend of risk diffusion in the future. By using this risk diffusion direction prediction vector, the direction of risk diffusion can be predicted in advance.

[0091] Step S133: Based on the risk evolution patterns in the historical risk event database, perform pattern matching processing on the risk diffusion direction prediction vector to generate risk trend prediction results.

[0092] A historical risk event database is a database that stores a large amount of information about historical risk events, including the time, location, type, and evolution of each event. Analysis of this database allows us to summarize the evolution patterns of different types of risks.

[0093] Pattern matching of the risk diffusion direction prediction vector involves comparing and matching the predicted vector with risk evolution patterns in a historical risk event database. Specifically, a similarity calculation method can be used to calculate the similarity between the risk diffusion direction prediction vector and each historical risk evolution pattern.

[0094] Based on the similarity calculation results, the historical risk evolution pattern with the highest similarity is selected as the matching pattern. Then, based on the risk evolution trend of the matching pattern, a risk trend prediction result is generated. The risk trend prediction result can include information such as the direction of risk development, the speed of development, and the possible scope of impact.

[0095] Pattern matching can be used to predict current risk trends by leveraging historical experience, thereby improving the accuracy and reliability of risk warnings.

[0096] Step S134: Map the risk state change trajectory to the time dimension channel of the spatial grid coordinate system, convert the risk diffusion direction prediction vector into a direction encoding matrix and superimpose it with the spatial grid coordinate system, generate a time decay weight coefficient based on the risk trend prediction result, and generate a dynamic risk evolution map by splicing the spatiotemporal channels.

[0097] To generate a dynamic risk evolution map, it is necessary to integrate the risk state change trajectory, the risk diffusion direction prediction vector, and the risk trend prediction results.

[0098] First, the trajectory of risk state changes is mapped to the time dimension channel of the spatial grid coordinate system. The spatial grid coordinate system is a three-dimensional coordinate system, where two dimensions represent spatial location and one dimension represents time. By mapping the trajectory of risk state changes to the time dimension channel in chronological order, the risk state at each point in time corresponds to a location in the spatial grid coordinate system.

[0099] Then, the risk diffusion direction prediction vector is converted into a direction encoding matrix. The direction encoding matrix is ​​a two-dimensional matrix where each element represents the risk diffusion direction at a spatial location. It can be converted into a direction encoding matrix based on the direction and magnitude of the risk diffusion direction prediction vector.

[0100] Next, the direction encoding matrix is ​​superimposed on the spatial grid coordinate system. The superimposed matrix contains information on spatial location, time, and the direction of risk diffusion.

[0101] A time decay weighting coefficient is generated based on the risk trend prediction results. The time decay weighting coefficient is a time-related coefficient that indicates the gradual weakening of the risk's impact over time. It can be generated based on the risk's development speed and potential impact range in the risk trend prediction results.

[0102] Finally, a dynamic risk evolution map is generated by spatiotemporal channel stitching. Spatiotemporal channel stitching involves combining information including spatial location, time, risk diffusion direction, and time decay weighting coefficients to form a four-dimensional dynamic risk evolution map. The dynamic risk evolution map graphically illustrates the spatial and temporal evolution of risk.

[0103] Step S140: Call the risk decision model to perform pattern recognition processing on the dynamic risk evolution map to generate risk level classification results and risk location coordinate set.

[0104] The risk decision model is a trained artificial intelligence model used for pattern recognition and risk assessment of dynamic risk evolution maps. This model can learn the characteristics and patterns of different risk models and, based on the input dynamic risk evolution map, output risk level classification results and a set of risk location coordinates.

[0105] Step S141: Extract key node features and edge connection features from the risk transmission path graph of the dynamic risk evolution graph to generate a key risk feature graph.

[0106] For example, step S1411: extract the time series risk parameters and three-dimensional spatial coordinate parameters of each key node in the risk transmission path map, and generate the spatiotemporal feature vector of the node.

[0107] Each key node in the risk transmission path map possesses specific risk status changes over time, as well as location information in three-dimensional space. By combining the time-series risk parameters (reflecting changes in risk intensity, duration, etc.) and three-dimensional spatial coordinate parameters of each key node, a node spatiotemporal feature vector is formed. This node spatiotemporal feature vector accurately describes the characteristics of the key node in both time and space.

[0108] Step S1412: Extract the weight parameters and direction encoding parameters from the edge connection features to generate the edge propagation strength matrix.

[0109] Edge connectivity features reflect the propagation of risk across different key nodes. The weight parameters reflect the probability and intensity of risk propagation along the edge connectivity, while the direction encoding parameters indicate the direction of risk propagation. Combining the weight and direction encoding parameters forms an edge propagation strength matrix, which comprehensively describes the propagation characteristics of risk along each edge connectivity.

[0110] Step S1413: Perform mean-shift clustering on the time series risk parameters in the spatiotemporal feature vector of the nodes to identify the stability index of risk nodes.

[0111] Mean-shift clustering is an unsupervised learning algorithm that processes the time-series risk parameters in the spatiotemporal feature vectors of nodes to group nodes with similar risk states together. Based on the clustering results, a stability index for risky nodes can be identified. If a node's risk state exhibits relatively stable characteristics during the clustering process, i.e., it remains in the same cluster category for a longer period, then the node has a high stability index; conversely, its stability index is low.

[0112] Step S1414: Perform spatial grid discretization mapping processing on the three-dimensional spatial coordinate parameters to generate a spatial distribution density index of nodes.

[0113] Spatial grid discretization mapping processes map continuous three-dimensional spatial coordinate parameters to a discrete spatial grid. By statistically analyzing the number of key nodes in each spatial grid, a spatial distribution density index of nodes can be generated. This index reflects the spatial distribution density of key nodes; if a spatial grid has a large number of key nodes, the corresponding node spatial distribution density index is high; conversely, it is low.

[0114] Step S1415: Perform channel-specific normalization on the weight parameters and direction encoding parameters in the edge propagation strength matrix to generate a normalized edge propagation vector.

[0115] Since the weight parameters and directional coding parameters may have different value ranges, they need to be normalized channel-by-channel to make them comparable. The weight parameters and directional coding parameters are normalized separately so that their values ​​are all within a uniform range. Then, the normalized parameters are combined into a standardized side propagation vector for subsequent feature processing.

[0116] Step S1416: Perform feature dimension alignment and splicing processing on the stability index, spatial distribution density index and standardized edge propagation vector to generate a fused node edge feature tensor.

[0117] Stability metrics, spatial distribution density metrics, and normalized edge propagation vectors describe the characteristics of key nodes and edge connections from different perspectives. To integrate these features, feature dimension alignment and concatenation are required. These metrics and vectors are concatenated according to predefined rules to ensure dimensionality matching, forming a fused node-edge feature tensor that integrates multiple feature information from key nodes and edge connections.

[0118] Step S1417: Perform time-dimension slicing processing on the edge feature tensor of the fused node based on a preset sliding time window to generate a multi-time-slice feature sequence.

[0119] The preset sliding time window is a fixed-size time interval that slides along the time dimension of the fused node edge feature tensor. Each slide captures a time slice. By continuously sliding the window, the fused node edge feature tensor is sliced ​​along the time dimension, resulting in a series of multi-time-slice feature sequences. These multi-time-slice feature sequences can reflect the changes in the features of key nodes and edge connections over time.

[0120] Step S1418: Perform bidirectional temporal convolution processing on the multi-time-slice feature sequence to extract risk transmission pattern features across time slices and generate the key risk feature map.

[0121] Bidirectional temporal convolution processing can fully utilize the information from consecutive time slices in a multi-time-slice feature sequence, performing convolution operations on the sequence to extract risk transmission pattern features across time slices. These risk transmission pattern features reflect the transmission rules and patterns of risk between different time slices. Integrating the extracted risk transmission pattern features generates a key risk feature map, which comprehensively describes the key feature information in the risk transmission path map.

[0122] Step S142: Input the key risk feature map into the pre-trained risk decision model, which includes a feature convolution module, an attention allocation module, and a classification output module.

[0123] The pre-trained risk decision-making model is a deep learning-based model consisting of a feature convolution module, an attention allocation module, and a classification output module.

[0124] The key risk feature map is input into the feature convolution module of the risk decision-making model. The feature convolution module is a convolutional neural network that performs local pattern extraction on the key risk feature map. Through convolution operations, the feature convolution module can extract features at different scales and levels from the key risk feature map, generating a multi-scale risk feature set.

[0125] The multi-scale risk feature set contains risk feature information at different scales and levels, which is crucial for accurately determining the risk level and location.

[0126] Step S143: The key risk feature map is processed by local pattern extraction through the feature convolution module to generate a multi-scale risk feature set.

[0127] The feature convolution module adopts the structure of a convolutional neural network and performs local pattern extraction processing on the key risk feature map through operations such as convolutional layers and pooling layers.

[0128] The convolutional layer is the core of the feature convolution module, containing multiple convolutional kernels. Each kernel can be viewed as a small filter that slides across the key risk feature map to extract local features. Different types of local features, such as edges and textures, can be extracted using different kernels.

[0129] During the convolution operation, the convolution kernel is multiplied element-wise with a local region of the key risk feature map and then summed to obtain the convolution result. The convolution result can be regarded as the feature representation of the key risk feature map in that local region.

[0130] Pooling layers are used to downsample the convolution results, reducing the dimensionality of features while retaining important feature information. Common pooling operations include max pooling and average pooling. Max pooling selects the maximum value in each local region as the pooling result, while average pooling calculates the average value in each local region as the pooling result.

[0131] Through multiple convolution and pooling operations, the feature convolution module can extract features at different scales and levels from the key risk feature map, generating a multi-scale risk feature set. This multi-scale risk feature set contains various risk feature information from local to global perspectives.

[0132] Step S144: The attention allocation module performs weight allocation processing on the multi-scale risk feature set to generate a focused risk feature set.

[0133] The attention allocation module performs weight allocation on the multi-scale risk feature set, enabling the model to focus more on important risk features.

[0134] The attention allocation module can employ various mechanisms, such as channel attention and spatial attention. Channel attention weights each channel in the multi-scale risk feature set, allowing the model to focus more on important channels. Spatial attention weights each spatial location in the multi-scale risk feature set, allowing the model to focus more on important spatial regions.

[0135] In this step, the attention allocation module calculates the importance weight of each feature based on the feature information of the multi-scale risk feature set. Features with higher importance weights are those that are more important for risk level classification and risk location, and the model will pay more attention to these features.

[0136] Based on the calculated importance weights, the multi-scale risk feature set is weighted to generate a focused risk feature set. This focused risk feature set is a subset of the multi-scale risk feature set and contains the risk features that the model considers most important. By focusing on this risk feature set, the model can more accurately classify risk levels and locate risks.

[0137] Step S145: Perform probability mapping processing on the focused risk feature set through the classification output module to generate the risk level classification result and the risk location coordinate set.

[0138] The classification output module is the last module of the risk decision-making model. Its function is to perform probability mapping processing on the set of focused risk features to generate risk level classification results and risk location coordinates.

[0139] The classification output module can use operations such as fully connected layers and softmax functions to implement probability mapping. The fully connected layer maps the feature vectors of the risk feature set to a higher-dimensional space, and then the softmax function transforms the mapped vectors into a probability distribution.

[0140] A probability distribution represents the likelihood of each risk level, and the final risk level classification can be determined based on the probability distribution. For example, the risk level with the highest probability is selected as the final classification result.

[0141] Simultaneously, the classification output module can also predict the location coordinates of risks based on the feature information of the focused risk feature set. By performing regression analysis or other methods on the feature vectors, the set of risk location coordinates can be obtained.

[0142] Finally, the classification output module outputs the risk level classification results and the risk location coordinate set. The risk level classification results enable monitoring personnel to quickly understand the severity of the risks in the current high-altitude operation scenario, while the risk location coordinate set can help them accurately locate the specific location where the risk occurs so that appropriate measures can be taken in a timely manner.

[0143] Step S150: Generate a risk warning signal based on the risk level classification result and the risk location coordinate set, and send the risk warning signal to the monitoring terminal.

[0144] After obtaining the risk level classification results and the set of risk location coordinates, a corresponding risk warning signal is generated. A risk warning signal is a signal that conveys risk information, which may include key information such as the risk level and location.

[0145] The process of generating a risk warning signal can involve encoding and integrating risk level classification results and risk location coordinate sets. For example, different risk levels can be represented by specific codes, risk location coordinates can be recorded in a specific format, and then this information can be combined into a complete risk warning signal.

[0146] After a risk warning signal is generated, it needs to be sent to a monitoring terminal. The monitoring terminal is typically a device located in a security monitoring center; it can receive and display risk warning signals, and sending these signals can be achieved through wired or wireless communication networks.

[0147] When sending risk warning signals, it is essential to ensure their accuracy and timeliness. Data verification and retransmission mechanisms can be used to guarantee accuracy, while real-time communication protocols can ensure timeliness. Upon receiving a risk warning signal, the monitoring terminal can display the risk information to monitoring personnel in a visually intuitive way, such as marking the risk location on a map or using different colors to represent different risk levels, enabling monitoring personnel to take timely measures to address the risk.

[0148] Furthermore, the method may also include: Step S210: Monitor the update status of the video image sequence of the high-altitude operation scene in real time. When the change in the position of the operator in the video image sequence exceeds the first preset threshold, identify the outline features of the operator in the current video image sequence. When the clarity index of the outline features of the operator is lower than the second preset threshold, calculate the expected set of position coordinates for future moments according to the operator's motion trajectory prediction algorithm.

[0149] To ensure the accuracy and timeliness of risk warnings, it is necessary to monitor the update status of video image sequences of high-altitude operation scenarios in real time. For example, the update status can be monitored by periodically checking information such as the timestamp or version number of the video image sequence.

[0150] When a change in the position of a worker in a video image sequence exceeds a first preset threshold, it indicates a significant shift in the worker's location, potentially impacting risk assessment and early warning. In this case, it's necessary to identify the worker's silhouette features in the current video image sequence. Image segmentation and edge detection techniques can be used to identify these silhouette features. First, an image segmentation algorithm separates the worker from the background; then, an edge detection algorithm extracts the worker's silhouette.

[0151] After identifying the outline features of the workers, their sharpness index needs to be calculated. The sharpness index can be obtained by calculating parameters such as the edge strength and contrast of the outline. When the sharpness index of the workers' outline features is lower than a second preset threshold, it indicates that the outline of the workers in the image is not sharp enough, which may affect subsequent analysis and processing.

[0152] At this point, it is necessary to calculate the expected set of coordinates for future positions based on a worker trajectory prediction algorithm. This algorithm can predict future movement trends based on information such as the worker's trajectory and speed over a past period. For example, Kalman filtering and particle filtering algorithms can be used for trajectory prediction. By observing and analyzing the worker's past positions, a motion model is established, and then the expected set of coordinates for future positions is predicted based on this model.

[0153] Step S211: Generate focal length adjustment parameters based on the expected position coordinate set, and control the corresponding image acquisition device to perform multi-stage focusing operation.

[0154] After obtaining the set of expected future location coordinates of the workers, focus adjustment parameters need to be generated based on these coordinate sets. Focus adjustment parameters are used to control the focal length of image acquisition devices (such as cameras), enabling the image acquisition devices to clearly capture images of the workers.

[0155] The process of generating focal length adjustment parameters can be performed based on the expected set of position coordinates and the parameters of the image acquisition device. For example, a suitable focal length adjustment value can be calculated based on information such as the distance between the operator and the image acquisition device, and the focal length range of the image acquisition device.

[0156] After generating the focus adjustment parameters, the corresponding image acquisition device is controlled to perform a multi-stage focusing operation. Multi-stage focusing refers to dividing the focusing process into multiple stages, gradually adjusting the focus to achieve the best focusing effect. For example, a coarse focus can be performed first to quickly adjust the focus to a roughly suitable range, followed by a fine focus to further refine the focus and make the image clearer for the operator.

[0157] During multi-stage focusing, focus parameters can be adjusted by sending control commands to the image acquisition device. Upon receiving the control command, the image acquisition device adjusts the focus parameters according to the focal length specified in the command until the optimal focus effect is achieved.

[0158] Step S212: After completing the multi-stage focusing operation, detect the light intensity distribution characteristics in the re-acquired image. When an overexposed or underexposed area is detected in the re-acquired image, start dynamic light equalization processing to obtain the equalized image.

[0159] After completing the multi-stage focusing operation, the image is reacquired. Then, the illumination intensity distribution characteristics in the reacquired image need to be detected. Detecting the illumination intensity distribution characteristics can be achieved by calculating the illumination intensity value of each pixel in the image and statistically analyzing its distribution.

[0160] When overexposed or underexposed areas are detected in the re-acquired image, it indicates uneven lighting conditions, which may affect subsequent analysis and processing. In this case, dynamic illumination equalization (LIFE) needs to be activated. LIFE is a technique used to adjust the light intensity of an image, making the lighting in the image more uniform.

[0161] Dynamic illumination equalization can employ various methods, such as histogram equalization and adaptive histogram equalization. Histogram equalization is a global illumination equalization method that adjusts the image's histogram to make the brightness distribution more uniform. Adaptive histogram equalization is a local illumination equalization method that adjusts the illumination based on the local features of the image, making the illumination more uniform in local areas as well.

[0162] Dynamic illumination equalization processing yields an equalized image. The equalized image has more uniform illumination and better reflects the actual conditions of the workers and their surrounding environment.

[0163] Step S213: Perform color space conversion on the equalized image to generate a standardized luminance channel and chrominance channel, and perform noise suppression processing on the standardized luminance channel using a bilateral filtering algorithm to generate an optimized luminance channel.

[0164] Color space conversion is performed on an equalized image to transform it from one color space to another more suitable one for processing. A common color space conversion is from RGB to YUV. In the YUV color space, the Y channel represents luminance information, while the U and V channels represent chrominance information.

[0165] Through color space conversion, normalized luminance and chrominance channels are generated. The normalized luminance channel contains the image's luminance information and has undergone normalization to ensure the luminance values ​​are within a suitable range. The chrominance channel contains the image's color information.

[0166] Then, noise suppression is performed on the normalized luminance channel using a bilateral filtering algorithm. Bilateral filtering is a non-linear filtering algorithm that can smooth an image while preserving its edge information. When performing bilateral filtering on the normalized luminance channel, the algorithm determines the filtering weights based on the spatial distance and brightness difference between pixels. Pixels that are close together and have small brightness differences are given larger weights to smooth the image; pixels that are far apart or have large brightness differences are given smaller weights to preserve the image's edge information.

[0167] Noise suppression using a bilateral filtering algorithm generates an optimized luminance channel. Noise in the optimized luminance channel is effectively suppressed while edge information is preserved, resulting in a clearer and easier-to-analyze image.

[0168] Step S214: Combine the optimized luminance channel with the chroma channel to generate a compensated video image sequence.

[0169] After obtaining the optimized luminance and chroma channels, they need to be composited to generate a compensated video image sequence. The compositing process integrates the information from the optimized luminance and chroma channels to recover a complete color image.

[0170] Specifically, the brightness information of the luminance channel can be combined with the color information of the chrominance channel. In the YUV color space, the values ​​of the Y, U, and V channels can be converted back to the RGB color space using a set conversion formula to obtain the compensated color image.

[0171] The above synthesis process is performed on each frame of the video image sequence to generate the compensated video image sequence. The compensated video image sequence shows improvements in sharpness, illumination uniformity, and color accuracy.

[0172] Step S215: Re-execute the multi-level feature extraction process based on the compensated video image sequence to generate an updated fused feature map.

[0173] Based on the compensated video image sequence, multi-level feature extraction processing is re-executed. The specific steps of multi-level feature extraction processing are similar to those performed on the original video image sequence, including edge enhancement and morphological closing operations on single-frame images, multi-category region segmentation based on a preset safety equipment semantic segmentation model, anchor point coordinate detection processing on the seat belt wearing area, bi-branch contour parsing processing on the self-locking device connection area, suspension point trajectory prediction processing on the speed difference device suspension area, channel-specific normalization processing of relevant parameters, and multi-level feature channel fusion processing.

[0174] By re-performing multi-level feature extraction, more accurate and richer feature information can be extracted from the compensated video image sequence. This feature information includes environmental features and features of the working subject. After integrating them, an updated fused feature map is generated. The updated fused feature map can more accurately reflect the actual situation of the current high-altitude operation scenario.

[0175] Step S216: Perform a difference comparison between the updated fused feature map and the historical fused feature map to generate a feature vector of changes in the work scenario.

[0176] Historical fusion feature maps are fusion feature maps generated at different points in time, recording the past feature information of high-altitude operation scenarios. The updated fusion feature map is compared with the historical fusion feature map to identify changes in the operation scenario at different points in time.

[0177] Difference comparison processing can be achieved by calculating the differences between corresponding features in the updated fused feature map and the historical fused feature map. For example, for environmental features such as illumination gradient parameters and work subject features such as safety belt anchor point position parameters, the differences between them in the updated map and the historical map can be calculated respectively.

[0178] By combining the differences among all corresponding features, a feature vector of work scenario changes is generated. This feature vector contains information about changes in the work scenario at different points in time, such as changes in lighting conditions, changes in the position and posture of workers, and changes in the use of safety equipment. This information is crucial for judging the dynamic changes in the work scenario and the development trend of risks.

[0179] Step S217: Update the dynamic risk evolution map according to the feature vector of the change in the work scenario, and regenerate the optimized risk warning signal.

[0180] The dynamic risk evolution map is updated based on the feature vector of changes in the work scenario. The dynamic risk evolution map is a map that reflects the changes in risk over time and space. Changes in the work scenario may lead to changes in the state and evolution trend of the risk.

[0181] The process of updating the dynamic risk evolution map can adjust the risk status, risk diffusion direction, and risk trend in the dynamic risk evolution map based on the information in the feature vector of changes in the work scenario. For example, if the feature vector of changes in the work scenario shows that the position of the workers has changed significantly, it may affect the direction and scope of risk diffusion. In this case, it is necessary to adjust the risk diffusion direction prediction vector and risk trend prediction results in the dynamic risk evolution map accordingly.

[0182] After updating the dynamic risk evolution map, optimized risk warning signals are regenerated. The process of regenerating risk warning signals is similar to before: based on the updated dynamic risk evolution map, a risk decision model is invoked for pattern recognition processing to generate new risk level classification results and a set of risk location coordinates. Then, optimized risk warning signals are generated based on these results. The optimized risk warning signals can more accurately reflect the risk situation of the current high-altitude work scenario.

[0183] Furthermore, the method may also include: Step S310: Associate and store the risk level classification results, the risk location coordinate set, and the corresponding video image segments to establish a risk event tracing database.

[0184] To facilitate the tracing and analysis of risk events, it is necessary to link and store the risk level classification results, risk location coordinates, and corresponding video image clips to establish a risk event tracing database.

[0185] Association-based storage refers to binding risk level classification results, risk location coordinate sets, and corresponding video image clips, so that they have a clear correspondence in the database. For example, a unique identifier can be assigned to each risk event, and the risk level classification results, risk location coordinate sets, and corresponding video image clips can all be associated with that identifier.

[0186] Risk event tracing databases can be implemented using either relational or non-relational databases. Relational databases, such as MySQL and Oracle, are suitable for storing structured data and facilitate data querying and management. Non-relational databases, such as MongoDB and Redis, are suitable for storing unstructured data, such as video and image clips.

[0187] Step S320: Perform periodic feature mining processing on the historical data in the risk event tracing database to generate a frequent risk pattern feature set, which includes multiple sample risk feature maps.

[0188] To identify potential risk patterns from historical data in the risk event tracing database, see the following steps: Step S321: Extract all risk event data within a preset time window of the historical data in the risk event tracing database, and generate a spatiotemporal distribution matrix.

[0189] The risk event tracing database contains historical data on risk events at different times and locations. To perform feature mining, it is necessary to extract all risk event data within a preset time window.

[0190] The choice of a preset time window needs to be determined based on the specific application scenario and analysis objectives. For example, if you want to analyze short-term risk patterns, you can choose a shorter time window, such as one day or one week; if you want to analyze long-term risk patterns, you can choose a longer time window, such as one month or one year.

[0191] After extracting all risk event data within a preset time window, this data needs to be organized into a spatiotemporal distribution matrix. The spatiotemporal distribution matrix is ​​a two-dimensional matrix, with one dimension representing time and the other representing spatial location. Each element in the matrix represents the number of risk events or other relevant information that occurred at that time and spatial location.

[0192] For example, assuming the time dimension can be divided into multiple time intervals and the spatial dimension can be divided into multiple spatial grids, then each element in the spatiotemporal distribution matrix can represent the number of risk events occurring in a certain spatial grid within a certain time interval. By generating the spatiotemporal distribution matrix, the distribution of risk events in time and space can be visually displayed.

[0193] Step S322: Perform cluster analysis on the spatiotemporal distribution matrix to identify dense and sparse regions of risk events.

[0194] Cluster analysis is an important step in data mining of spatiotemporal distribution matrices. Its purpose is to group similar risk event data points into the same category, thereby identifying dense and sparse areas of risk events.

[0195] Various clustering algorithms can be used for cluster analysis, such as K-means clustering and hierarchical clustering. Taking K-means clustering as an example, the basic idea of ​​this algorithm is to first randomly select K cluster centers, then assign each data point to the category of the nearest cluster center, then update the position of the cluster centers, and repeat this process until the cluster centers no longer change or the preset number of iterations is reached.

[0196] When performing cluster analysis on a spatiotemporal distribution matrix, each element in the matrix is ​​treated as a data point, and clusters are formed based on their values ​​in the temporal and spatial dimensions. After clustering, each category corresponds to a distribution area of ​​risk events. If a category contains a large number of data points, it indicates that the area is a dense region of risk events; if a category contains a small number of data points, it indicates that the area is a sparse region of risk events.

[0197] By identifying dense and sparse areas of risk events, we can understand the aggregation of risk events in time and space, providing key areas of focus for subsequent causal relationship analysis and risk pattern mining.

[0198] Step S323: Perform causal relationship analysis on the risk events in the dense area to generate a risk transmission path map.

[0199] For example, step S3231: extract the timestamp parameters and three-dimensional spatial coordinate parameters of each risk event in the dense area to generate a spatiotemporal attribute vector group.

[0200] Within a dense region, each risk event has its own time and spatial location. A timestamp parameter records the specific time of the risk event's occurrence, while three-dimensional spatial coordinates determine its specific location. Combining the timestamp and coordinates of each risk event forms a spatiotemporal attribute vector. This process is applied to all risk events within the dense region, ultimately yielding a set of spatiotemporal attribute vectors that accurately describe the temporal and spatial distribution of risk events within the dense region.

[0201] Step S3232: Perform synchronization interpolation processing on the spatiotemporal attribute vector group to generate a time-aligned event sequence set, and perform discretization mapping processing on the three-dimensional spatial coordinate parameters according to a preset spatial grid resolution to generate a spatial grid encoded sequence.

[0202] Synchronization interpolation is used to make the timestamp parameters of different risk events comparable. Since the occurrence times of different risk events may vary, interpolation adjusts the timestamp parameters to make the event sequence more continuous in time, thus generating a time-aligned set of event sequences. Simultaneously, discretization mapping is performed on the three-dimensional spatial coordinate parameters according to a preset spatial grid resolution. This involves mapping continuous three-dimensional spatial coordinates to discrete spatial grids, each with a unique code. The final result is a spatial grid-coded sequence that clearly represents the spatial distribution of risk events.

[0203] Step S3233: Input the time-aligned event sequence and the spatially grid-coded event sequence into the event association analysis model to calculate the temporal delay correlation and spatial coupling degree between adjacent events.

[0204] Event correlation analysis models are built upon machine learning or deep learning algorithms. By learning from extensive historical data, they establish correlation models between events. When time-aligned event sequences and spatially grid-coded sequences are input into the model, it analyzes the temporal order and time intervals of adjacent events to calculate temporal correlation. Simultaneously, it considers the spatial relationships of adjacent events within the spatial grid to calculate spatial coupling. Temporal correlation reflects the temporal correlation between two events, while spatial coupling reflects their spatial correlation.

[0205] Step S3234: Perform dual-weight normalization processing on the time delay correlation degree and spatial coupling degree to generate event causal strength parameters.

[0206] Since the values ​​of temporal correlation and spatial coupling may differ, a dual-weighted normalization process is required to comprehensively consider them. Temporal correlation and spatial coupling are normalized separately to ensure their values ​​fall within a uniform range. Then, the normalized temporal correlation and spatial coupling are weighted and combined according to preset weighting coefficients to generate an event causality strength parameter. This parameter comprehensively considers the temporal and spatial correlation of events, providing a more complete reflection of the causal relationships between events.

[0207] Step S3235: Based on a preset causal strength threshold, the event causal strength parameters are filtered to extract causal event pairs that meet the conditions.

[0208] The preset causal strength threshold is a critical value set based on actual conditions and experience. Only event pairs whose causal strength parameters exceed this threshold are considered to have a strong causal relationship. The causal strength parameters are then filtered to extract causal event pairs that meet the criteria; these causal event pairs are key components of the risk transmission path map.

[0209] Step S3236: Generate a node connection relationship topology graph based on the spatiotemporal attribute vector of the causal event pair, and assign the causal strength parameter as the edge weight value to the corresponding connection edge to generate a weighted risk transmission path graph.

[0210] Based on the spatiotemporal attribute vectors of causal event pairs, the temporal and spatial positional relationships of each event can be determined, thereby generating a node connection topology graph. In this topology graph, each node represents a risk event, and the connecting edges represent the causal relationships between events. A causal strength parameter is assigned as the weight value to the corresponding connecting edges, generating a weighted risk transmission path graph. This graph not only reflects the causal relationships between events but also demonstrates the strength of those causal relationships.

[0211] Step S324: Extract key node features and edge connection features from the risk transmission path graph to generate a sample risk feature graph in the frequent risk pattern feature set.

[0212] After obtaining the weighted risk transmission path map, it is necessary to extract key node features and edge connection features from it to generate a sample risk feature map in the frequent risk pattern feature set. For details, please refer to the relevant description of step S141 above.

[0213] Step S330: Optimize the decision parameters of the risk decision model based on the set of frequent risk pattern features to generate an updated risk decision model.

[0214] Step S331: Input the sample risk feature map from the frequent risk pattern feature set into the feature convolution module of the risk decision model to extract the initial convolution feature map.

[0215] The frequent risk pattern feature set contains multiple sample risk feature maps, which reflect the characteristic information of frequently occurring risk patterns in high-altitude operation scenarios. These sample risk feature maps are input into the feature convolution module of the risk decision model. This feature convolution module performs local pattern extraction processing on the input maps through convolution operations to generate an initial convolutional feature map. The initial convolutional feature map contains feature information at different scales and levels from the sample risk feature maps.

[0216] Step S332: Perform gradient backpropagation on the initial convolutional feature map and calculate the weight gradient correction amount of each convolutional kernel.

[0217] Backpropagation is an important algorithm in deep learning used to update model parameters. During backpropagation, the gradient correction amount for each convolutional kernel's weights is calculated based on the error between the model's output and the expected result. The weight gradient correction amount represents the direction and magnitude of the adjustment needed for the weights of each convolutional kernel to reduce error. By continuously adjusting the weights of the convolutional kernels, the model's output can be made closer to the expected result, thereby improving the model's accuracy.

[0218] Step S333: Update the kernel parameters of the feature convolution module according to the weight gradient correction amount to generate an optimized multi-scale feature extractor.

[0219] Based on the calculated weight gradient correction, the convolution kernel parameters of the feature convolution module are updated. The update process involves adding the weight gradient correction to the original convolution kernel parameters to obtain the new parameters. The updated kernel parameters better adapt to the sample risk feature maps in the frequent risk pattern feature set, thereby improving the feature extraction capability of the feature convolution module. After parameter updates, an optimized multi-scale feature extractor is generated, which can more accurately extract features at different scales and levels from the sample risk feature maps.

[0220] Step S334: Input the sample risk feature map into the optimized multi-scale feature extractor to generate an updated multi-scale risk feature set.

[0221] The sample risk feature map is input into the optimized multi-scale feature extractor, which performs feature extraction processing on the input map to generate an updated multi-scale risk feature set. The updated multi-scale risk feature set contains more accurate and richer risk feature information.

[0222] Step S335: The attention allocation module of the risk decision model performs channel attention weight redistribution processing on the updated multi-scale risk feature set to generate dynamic adjustment parameters for focusing on risk features.

[0223] The attention allocation module of the risk decision-making model is responsible for weighting the multi-scale risk feature set, enabling the model to focus more on important risk features. By performing channel attention weight reallocation on the updated multi-scale risk feature set, the attention allocation module recalculates the importance weight of each channel based on the feature information of the set. Based on these importance weights, dynamic adjustment parameters are generated to focus on risk features; these parameters are used to adjust the model's attention level towards different channel features.

[0224] Step S336: Update the weight allocation matrix of the attention allocation module according to the dynamic adjustment parameters to generate an optimized risk decision model.

[0225] The weight allocation matrix of the attention allocation module is updated based on the dynamically adjusted parameters generated for the focused risk features. The weight allocation matrix records the importance weight of each channel, and the updated matrix enables the model to more accurately focus on important risk features. By updating the weight allocation matrix, an optimized risk decision-making model is generated, which exhibits higher accuracy and effectiveness in feature extraction and risk analysis.

[0226] Step S340: Deploy the updated risk decision-making model to the real-time early warning system, replacing the original model version.

[0227] After generating the updated risk decision-making model, it needs to be deployed to the real-time early warning system to replace the original model version.

[0228] Figure 2 The illustration shows exemplary hardware and software components of a camera image recognition-based high-altitude operation risk warning system 100, which can implement the ideas of this application, according to some embodiments of this application. For example, a processor 120 can be used in the camera image recognition-based high-altitude operation risk warning system 100 and to perform the functions in this application.

[0229] The high-altitude operation risk warning system 100 based on camera image recognition can be a general-purpose server or a special-purpose server; both can be used to implement the high-altitude operation risk warning method based on camera image recognition of this application. Although only one server is shown in this application, for convenience, the functions described in this application can be implemented in a distributed manner on multiple similar platforms to balance the processing load.

[0230] For example, a high-altitude operation risk warning system 100 based on camera image recognition may include a network port 110 connected to a network, one or more processors 120 for executing program instructions, a communication bus 130, and various forms of storage media 140, such as a disk, ROM, or RAM, or any combination thereof. Exemplarily, the high-altitude operation risk warning system 100 based on camera image recognition may also include program instructions stored in ROM, RAM, or other types of non-transitory storage media, or any combination thereof. The methods of this application can be implemented according to these program instructions. The high-altitude operation risk warning system 100 based on camera image recognition also includes an I / O interface 150 between the computer and other input / output devices.

[0231] For ease of explanation, only one processor is described in the high-altitude operation risk warning system 100 based on camera image recognition. However, it should be noted that the high-altitude operation risk warning system 100 based on camera image recognition in this application may also include multiple processors. Therefore, the steps performed by one processor as described in this application may also be performed jointly or individually by multiple processors. For example, if the processor of the high-altitude operation risk warning system 100 based on camera image recognition performs steps A and B, it should be understood that steps A and B may also be performed jointly by two different processors or individually by one processor. For example, the first processor performs step A, the second processor performs step B, or the first processor and the second processor jointly perform steps A and B.

[0232] Furthermore, this embodiment of the invention also provides a readable storage medium, which has computer-executable instructions pre-set in it. When the processor executes the computer-executable instructions, the above-mentioned high-altitude operation risk warning method based on camera image recognition is implemented.

[0233] It should be noted that, in order to simplify the description of the present invention and thus help to understand one or more embodiments of the invention, multiple features may sometimes be grouped into one embodiment, drawing or description thereof in the foregoing description of the embodiments of the present invention.

Claims

1. A method for early warning of risks in high-altitude operations based on camera image recognition, characterized in that, The method includes: A video image sequence of a high-altitude operation scene is acquired, and multi-level feature extraction processing is performed on the video image sequence to generate a fused feature map containing environmental feature set and operation subject feature set; The fused feature map is segmented spatially to obtain multiple local region feature sub-maps. Regional feature comparison analysis is performed on each local region feature sub-map to generate a spatial risk distribution map containing risk area identification information. Based on preset temporal feature fusion rules, temporal feature fusion processing is performed on the spatial risk distribution map of consecutive frames in the video image sequence to generate a dynamic risk evolution map. The risk decision-making model is invoked to perform pattern recognition processing on the dynamic risk evolution map, generating a risk level classification result and a set of risk location coordinates; A risk warning signal is generated based on the risk level classification result and the risk location coordinate set, and the risk warning signal is sent to the monitoring terminal.

2. The high-altitude operation risk early warning method based on camera image recognition according to claim 1, characterized in that, The step of performing multi-level feature extraction processing on the video image sequence to generate a fused feature map containing environmental feature sets and subject feature sets includes: Edge enhancement and morphological closing operations are performed on single frames in the video image sequence to generate an enhanced target region segmentation map; Based on the preset safety equipment semantic segmentation model, the target area segmentation map is divided into multiple categories of regions to separate the seat belt wearing area, the self-locking device connection area and the speed difference device suspension area. Anchor point coordinate detection processing is performed on the seat belt wearing area to extract the spatial position parameters of the main anchor point and auxiliary anchor point of the seat belt; A bi-branch contour analysis process is performed on the self-locking device connection area to separate the self-locking device body contour and the connecting rope contour, and the minimum bounding box center distance parameter of the self-locking device body contour and the connecting rope contour is calculated. The suspension point trajectory prediction process is performed on the suspension area of ​​the speed difference device to extract the instantaneous position coordinates of the suspension device in the current frame and its offset angle parameter with respect to the preset baseline. The spatial position parameters of the main anchor point and auxiliary anchor point of the safety belt, the center distance parameter between the self-locking device body and the connecting rope, and the offset angle parameter of the suspension device are respectively subjected to channel normalization processing to generate a dynamic feature matrix of the safety equipment. Multi-level feature channel fusion processing is performed on the dynamic feature matrix of the safety equipment to generate the main feature set of the operation subject, which includes the distribution of seat belt anchor points, the connection status of self-locking devices and the offset of speed difference devices. This set is then cross-sourcely spliced ​​with the ambient light gradient parameters extracted by background segmentation to generate the fused feature map.

3. The high-altitude operation risk early warning method based on camera image recognition according to claim 1, characterized in that, The step of performing regional feature comparison analysis on each of the local region feature sub-maps to generate a spatial risk distribution map containing risk area identification information includes: Extract the location coordinates of the safety equipment wearing area of ​​the workers from the feature sub-image of the local area, and generate safety belt anchor point parameters, climbing self-locking device connection point parameters, and speed difference device suspension point parameters; The anchor point parameters of the seat belt are spatially projected and matched with the preset human torso reference coordinate system to calculate the contour overlap between the fixed area of ​​the seat belt and the reference area. The parameters of the climbing self-locking device connection point are subjected to dual-branch feature analysis processing to separate the position coordinates of the rope end and the human connection end, and the offset ratio of the distance between the two ends to the standard safety distance is calculated. The suspension point parameters of the speed difference device are input into the three-dimensional spatial trajectory prediction model to generate the dynamic movement trajectory of the suspension device, and the vertical projection deviation between the trajectory end point and the preset high-hanging and low-use safety zone is calculated. The contour overlap, offset ratio and vertical projection deviation are standardized to generate a standardized set of safety equipment matching parameters. The standardized safety equipment matching parameter set is input into the multi-level risk decision tree model to generate risk area identification information including seat belt failure identification, self-locking device misalignment identification, and speed difference violation identification. The risk area identification information is then mapped to the spatial grid coordinate system of the local area feature sub-map to generate the spatial risk distribution map.

4. The high-altitude operation risk early warning method based on camera image recognition according to claim 1, characterized in that, The process of performing temporal feature fusion processing on the spatial risk distribution map of consecutive frames in the video image sequence based on preset temporal feature fusion rules to generate a dynamic risk evolution map includes: Extract time-series features from the spatial risk distribution map of consecutive frames to generate risk state change trajectories; The trajectory of the risk state change is subjected to motion vector analysis to generate a risk diffusion direction prediction vector; Based on the risk evolution patterns in the historical risk event database, the risk diffusion direction prediction vector is subjected to pattern matching processing to generate risk trend prediction results. The risk state change trajectory is mapped to the time dimension channel of the spatial grid coordinate system. The risk diffusion direction prediction vector is converted into a direction encoding matrix and superimposed on the spatial grid coordinate system. A time decay weight coefficient is generated based on the risk trend prediction result. A dynamic risk evolution map is generated by splicing the spatiotemporal channels.

5. The high-altitude operation risk early warning method based on camera image recognition according to claim 1, characterized in that, The process of calling the risk decision model to perform pattern recognition processing on the dynamic risk evolution map generates a set of risk level classification results and risk location coordinates, including: Extract key node features and edge connection features from the risk transmission path graph of the dynamic risk evolution graph to generate a key risk feature graph; The key risk feature map is input into a pre-trained risk decision model, which includes a feature convolution module, an attention allocation module, and a classification output module. The key risk feature map is processed by the feature convolution module to extract local patterns, generating a multi-scale risk feature set. The attention allocation module performs weight allocation processing on the multi-scale risk feature set to generate a focused risk feature set. The classification output module performs probability mapping processing on the focused risk feature set to generate the risk level classification result and the risk location coordinate set.

6. The high-altitude operation risk early warning method based on camera image recognition according to claim 1, characterized in that, After generating a risk warning signal based on the risk level classification result and the risk location coordinate set, and sending the risk warning signal to the monitoring terminal, the method further includes: The system monitors the update status of the video image sequence of the high-altitude operation scene in real time. When the change in the position of the operator in the video image sequence exceeds the first preset threshold, the system identifies the outline features of the operator in the current video image sequence. When the clarity index of the outline features of the operator is lower than the second preset threshold, the system calculates the expected set of position coordinates for future moments based on the operator's motion trajectory prediction algorithm. Based on the expected set of position coordinates, a focal length adjustment parameter is generated, and the corresponding image acquisition device is controlled to perform a multi-stage focusing operation. After completing the multi-stage focusing operation, the light intensity distribution characteristics in the re-acquired image are detected. When an overexposed or underexposed area is detected in the re-acquired image, dynamic light equalization processing is initiated to obtain the equalized image. The image after equalization is subjected to color space conversion to generate a standardized luminance channel and chrominance channel. Then, a bilateral filtering algorithm is used to suppress noise in the standardized luminance channel to generate an optimized luminance channel. The optimized luminance channel and the chroma channel are combined to generate a compensated video image sequence; Based on the compensated video image sequence, the multi-level feature extraction process is re-executed to generate an updated fused feature map. The updated fused feature map is compared with the historical fused feature map to generate a feature vector of changes in the work scenario. The dynamic risk evolution map is updated based on the feature vector of the change in the work scenario, and the optimized risk warning signal is regenerated.

7. The high-altitude operation risk early warning method based on camera image recognition according to claim 1, characterized in that, After generating a risk warning signal based on the risk level classification result and the risk location coordinate set, the method further includes: The risk level classification results, the risk location coordinate set, and the corresponding video image segments are associated and stored to establish a risk event tracing database. Periodic feature mining processing is performed on the historical data in the risk event tracing database to generate a frequent risk pattern feature set, which includes multiple sample risk feature maps. The decision parameters of the risk decision model are optimized based on the set of features of frequent risk patterns to generate an updated risk decision model; The updated risk decision-making model is deployed to the real-time early warning system, replacing the original model version.

8. The high-altitude operation risk early warning method based on camera image recognition according to claim 7, characterized in that, The process of periodically mining historical data in the risk event tracing database to generate a set of frequent risk pattern features includes: Extract all risk event data within a preset time window from the historical data in the risk event tracing database, and generate a spatiotemporal distribution matrix; Cluster analysis is performed on the spatiotemporal distribution matrix to identify dense and sparse regions of risk events; Causal relationship analysis is performed on risk events within the dense area to generate a risk transmission path map. Key node features and edge connection features are extracted from the risk transmission path graph to generate a sample risk feature graph in the frequent risk pattern feature set.

9. The high-altitude operation risk early warning method based on camera image recognition according to claim 7, characterized in that, The step of optimizing the decision parameters of the risk decision model based on the frequent risk pattern feature set to generate an updated risk decision model includes: Input the sample risk feature map from the frequent risk pattern feature set into the feature convolution module of the risk decision model to extract the initial convolution feature map; Gradient backpropagation is performed on the initial convolutional feature map to calculate the weight gradient correction of each convolutional kernel; The kernel parameters of the feature convolution module are updated according to the weight gradient correction amount to generate an optimized multi-scale feature extractor. The sample risk feature map is input into the optimized multi-scale feature extractor to generate an updated multi-scale risk feature set; The attention allocation module of the risk decision model performs channel attention weight redistribution processing on the updated multi-scale risk feature set to generate dynamic adjustment parameters that focus on risk features. The weight allocation matrix of the attention allocation module is updated according to the dynamically adjusted parameters to generate an optimized risk decision model.

10. A high-altitude operation risk early warning system based on camera image recognition, characterized in that, The device includes a processor and a memory, the memory being connected to the processor. The memory is used to store programs, instructions, or code, and the processor is used to execute the programs, instructions, or code in the memory to implement the high-altitude operation risk warning method based on camera image recognition as described in any one of claims 1-9.

Images