Real-time detecting and positioning method for dynamic dangerous area of cargo hoisting

Abstract

The invention relates to the technical field of cargo hoisting, in particular to a real-time detecting and positioning method and system for a cargo hoisting dynamic dangerous area, and the method comprises the following steps: 1, installing a binocular camera on hoisting equipment, and obtaining depth map information of a hoisted object by using a binocular camera technology; 2, whether hoisting operation is carried out or not is judged by setting a corresponding depth threshold value; the binocular vision technology is used for completing the two functions of lifting judgment of the hoisted goods and accurate detection and positioning of the hoisting dynamic dangerous area. The lifting operation state is effectively judged, dynamic dangerous area position information generated by lifting operation is obtained in real time, the problems that whether cargo lifting work is carried out or not cannot be judged, objects stored on the ground and objects to be lifted cannot be distinguished are solved, and the problems that real-time detection of the dynamic dangerous area is difficult, and positioning information is inaccurate are solved; and a new method is provided for dynamic dangerous area detection and positioning.

Description

Technical Field

[0001] The invention relates to the technical field of cargo hoisting, and in particular to a real-time detection and positioning method for dynamic dangerous areas of cargo hoisting. Background Art

[0002] In the field of industrial production, especially when it comes to large-scale cargo lifting operations, bridge cranes, as the most critical lifting equipment, are often used to carry heavy objects. However, bridge crane lifting operations often involve a more complex working environment with many potential risk factors. The falling danger zone generated by the bridge crane lifting cargo is a place that is very likely to cause safety accidents. Therefore, timely and accurate identification of dangerous areas is the basis for effectively preventing potential dangers in lifting operations. In the field of dangerous area intrusion detection, how to determine the dangerous area and obtain the location information of the dangerous area is a prerequisite for intrusion detection. The bridge crane lifting danger area is generated by the lifting and transportation of the lifting cargo. The setting of the danger area is based on the contour of the lifting cargo. Because the lifting cargo moves continuously with the bridge crane and the shape and size of the lifting cargo are inconsistent, compared with the static danger area, the bridge crane lifting danger area type is a more random dynamic danger area, which makes it difficult to determine the location information of the bridge crane lifting danger area. The prerequisite for controlling and managing dangerous areas is to determine the location information of dangerous areas. However, dynamic dangerous areas are difficult to identify and determine because the danger sources are constantly moving. This problem has a great impact on the subsequent invasion control of dangerous areas.

[0003] At present, traditional dynamic danger zone intrusion detection is mainly achieved through manual inspections or employees wearing relevant sensors. By installing positioning sensors on the hoisted cargo, the position information of the hoisted cargo is obtained in real time, thereby determining the dynamic danger zone of the hoisted cargo. With the continuous development of computer vision technology, the invention "An early warning method and system for collision accidents between excavators and people based on computer vision" uses target detection technology to identify the targets of excavators and pedestrians, and calculates the position information of the dynamic danger zone of the excavator through the identified excavator target frame information. Although target detection technology can identify dynamic danger sources in real time, when it is applied to hoisting operations in complex environments, it has two major disadvantages: it cannot determine whether a hoisting operation is currently being performed, and it cannot distinguish between goods stored on the ground and goods being hoisted and transported.

[0004] The present invention aims to solve the problem that dynamic dangerous areas during cargo hoisting are difficult to identify and position information is difficult to determine, and to provide a new method for positioning and detecting dynamic dangerous areas. Summary of the invention

[0005] In view of the above-mentioned defects and problems, the present invention provides a method for real-time detection and positioning of dynamic dangerous areas in cargo hoisting. The method mainly uses binocular vision technology to complete the two functions of judging the lifting of hoisted cargo and accurately detecting and positioning the dynamic dangerous areas of hoisting, thereby effectively judging the state of hoisting operation and obtaining the location information of dynamic dangerous areas generated by hoisting operation in real time, providing the prerequisite for subsequent dynamic dangerous area management and control. At the same time, it provides a new method for positioning and detecting dynamic dangerous areas.

[0006] The solution adopted by the present invention to solve the technical problem is: a real-time detection and positioning method for dynamic dangerous areas of cargo hoisting, comprising the following steps:

[0007] Step 1: Install the binocular camera on the lifting equipment and use the binocular camera technology to obtain the depth map information of the hoisted object;

[0008] Step 2: Determine whether to perform lifting operation by setting the corresponding depth threshold;

[0009] Step 3: If the object is in the lifting and transporting stage, the image processing algorithm is used to remove the redundant object connected domains in the depth map;

[0010] Step 4: After completing image processing operations such as depth map filtering, the contour of the hoisted object in the depth map is extracted;

[0011] Step 5: Use the minimum enclosing rectangle method to draw the minimum enclosing rectangle of the object outline and obtain the position information of the hoisted object in the image;

[0012] Step 6: Combined with the actual application scenario and the workshop safety department's requirements for the division of the lifting danger zone, the lifting danger zone is set based on the minimum circumscribed rectangle of the lifting object outline obtained above;

[0013] Step 7: Map the location information of the dangerous area to the RGB image obtained by the binocular camera to provide a prerequisite for subsequent personnel intrusion detection tasks.

[0014] Furthermore, in step 2, in order to improve the accuracy of judging the lifting object of the bridge crane, the depth values ​​of three pixel areas in the field of view of the binocular camera are selected as reference data, and the depth value involved in the judgment is the average value of the depth values ​​in each area. The above object lifting judgment function is realized by using Python code:

[0015]

[0016] Where D i is the average depth value of the ith region; d ij is the depth value of the jth pixel coordinate in the i-th region; N iis the total number of pixels in the i-th region, where i is the number of depth value determination regions.

[0017] Furthermore, the method of removing the redundant object connected domains in the depth map by using the image processing algorithm in step 3 is as follows:

[0018] S31, using a bilateral filter to optimize the acquired depth map to achieve the effect of noise reduction and smoothing and maintaining the contour. The calculation formula is:

[0019]

[0020] In the formula, i and j are the coordinates of the pixels involved in the convolution; k and l are the coordinates of the neighboring pixels, and their weight coefficients ω are:

[0021]

[0022] S32. After the noise reduction process is completed, the pixel values ​​of the depth image are mapped to the range of [0, 255] by using a linear normalization operation, thereby converting the depth image into a grayscale image. The linear normalization mathematical expression is as follows:

[0023]

[0024] Where, d ij represents the depth value of the i-th row and j-th column in the depth map, D min , D max They represent the minimum and maximum depth values ​​in the depth map, respectively. a and b are the normalized ranges. d ijn is the normalized depth value;

[0025] S33. In order to effectively distinguish the background and the target object, the normalized depth map is subjected to K-means clustering operation. The clustering method optimizes the grouping of data points iteratively so that the depth information in the same cluster has a high similarity, while the depth information between different clusters has a large difference. Its mathematical model is realized by minimizing the objective function:

[0026]

[0027] In the formula, c i is the normalized depth value of the center point of the i-th cluster; S i is the set of data points contained in the i-th cluster; d j is the normalized depth value of the jth data point; Used to measure the closeness of data points within the same cluster;

[0028] Through the above clustering operation, the part of the normalized depth map that exceeds the depth threshold is clustered into one category as the background, and the part within the depth threshold is clustered into one category as the target object;

[0029] S34: To achieve the requirement of retaining only the connected domain of the target object in the depth map, the redundant connected domains in the depth map after normalization and clustering are processed by setting pixel coordinate thresholds.

[0030] Furthermore, in step 4, the Canny operator edge detection algorithm is used to draw the object contour, and the steps are as follows:

[0031] S41: firstly use two-dimensional Gaussian low-pass filtering to smooth the image;

[0032] S42: Then, the Sobel operator is used to calculate the gradient magnitude and direction of each pixel at the smoothed depth image in the horizontal and vertical directions;

[0033] S43: Then, non-maximum suppression is used to find the maximum value of the pixel in the gradient direction to determine whether to temporarily define the central pixel point as an edge point;

[0034] S44: Finally, a double threshold algorithm is used to determine edge points and connect and draw the outline of the object.

[0035] Furthermore, the mathematical expression of the two-dimensional Gaussian low-pass filter in S1 for smoothing and denoising the depth map is:

[0036]

[0037] Where σ is the standard deviation of the Gaussian function, which is used to control the smoothing effect of the image.

[0038] Furthermore, in S42, the Sobel operator is used to replace the traditional 2×2 neighborhood first-order partial derivative template and the image window to perform gradient convolution calculation:

[0039]

[0040] Where D is the depth image window, the middle pixel is e, and G x and G y is the horizontal and vertical gradient of the middle pixel, S x and S y is the Sobel operator template;

[0041] Combining the above formula, the gradient size and direction of each pixel are as follows:

[0042]

[0043] Furthermore, the non-maximum suppression in S43 is to determine whether the gradient amplitude of a certain pixel point is the maximum value in the direction along the pixel gradient direction. If so, it is determined to be a possible edge point. If not, the grayscale value of this point is set to 0, thereby removing redundant edge pixels and accurately locating the edge of the image object. After the non-maximum suppression process, the retained local maximum points are filtered by setting two thresholds to remove edge points less than the minimum threshold and retain points between the two thresholds that are greater than the maximum threshold and connected to the real edge.

[0044] Furthermore, in step six, after communicating with the safety department of the production workshop, the minimum enclosing rectangle of the hoisting object's own contour is used as the basis, and the contour range of 0.5 meters is set as the bridge crane hoisting danger zone. The extracted hoisting object contour is used to use the cv2.minAreaRect function provided by OPENCV to draw the minimum enclosing rectangle of the hoisting object, thereby obtaining the location information of this dynamic danger zone.

[0045] The beneficial effects of the present invention are as follows: binocular vision technology is used to complete the two functions of judging the lifting of hoisted cargo and accurately detecting and locating the dynamic dangerous area of ​​hoisting. First, the binocular camera is calibrated and deployed at the designated position of the bridge crane after calibration to obtain the image and depth information of the hoisted cargo, and judge whether the bridge crane is performing hoisting work according to the set depth threshold; secondly, when it is judged that the bridge crane is performing hoisting operations, the acquired depth map of the hoisted object is processed, the contour information of the hoisted object is identified and the minimum circumscribed rectangle of the hoisted object contour is obtained. According to the requirements for the division of the hoisting dangerous area of ​​the bridge crane in the workshop of the enterprise, combined with the minimum circumscribed rectangle of the hoisting object contour, the location information of the hoisting dangerous area of ​​the bridge crane is accurately and flexibly obtained, thereby effectively judging the hoisting operation status and obtaining the dynamic dangerous area location information generated by the hoisting operation in real time, solving the problem of being unable to judge whether to perform cargo lifting work and distinguishing between ground stored objects and hoisted objects, and solving the problem of difficult real-time detection of dynamic dangerous areas and inaccurate positioning information, providing a new method for dynamic dangerous area detection and positioning. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 It is a work flow chart of the present invention;

[0047] Figure 2 It is a Canny operator detection step diagram of the present invention;

[0048] Figure 3 is the depth map after normalized clustering of the present invention;

[0049] Figure 4 The depth map of the present invention only retains the target connected domain;

[0050] Figure 5 It is a contour image of the hoisted object processed by the Canny operator edge detection algorithm of the present invention;

[0051] Figure 6 This is a rendering of the lifting hazardous area identification effect of the present invention. DETAILED DESCRIPTION

[0052] The present invention takes bridge crane hoisting operation as the application scenario and combines it with the actual needs of engineering projects. The existing research in the field of dangerous area intrusion detection mainly focuses on static dangerous areas. The dynamic dangerous areas of bridge crane hoisting have problems such as difficult location division and difficult information determination. A dangerous area detection method based on binocular vision is proposed. The method mainly uses binocular vision technology to complete two research tasks: lifting judgment of hoisted cargo and accurate detection of dangerous areas of bridge crane hoisting, so as to effectively judge the hoisting operation status and obtain the dynamic dangerous area location information generated by the hoisting operation in real time, providing a prerequisite for subsequent dynamic dangerous area control.

[0053] The present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0054] Embodiment 1:

[0055] See also Figure 1-6 The present invention provides a technical solution for a method for real-time detection and positioning of dynamic hazardous areas during cargo hoisting. The technology used by the present invention to obtain the dynamic hazardous source depth map contour information includes but is not limited to binocular vision technology, laser radar technology, etc. Here, binocular vision technology is taken as an example, including the following steps:

[0056] Step 1: After completing the binocular camera calibration and parameter setting, the present invention installs the binocular camera at the designated position of the lifting equipment (for example, the trolley beam above the bridge crane sling) according to the detection task requirements, so as to obtain the image and depth information of the hoisted cargo, realize the functions of real-time and accurate judgment of the hoisting working state of the bridge crane and flexible segmentation of the outline of the hoisted cargo, etc. The specific installation position is to make the hook of the bridge crane located in the center of the field of view.

[0057] After the binocular camera is installed and deployed, the binocular camera SDK toolkit is used to test the camera's working effect, and its color RGB and depth map information can be obtained. According to the image information obtained, the bridge crane hook is located in the center of the field of view. During the hoisting operation, the overall outline of the hoisted cargo can be fully captured, laying the foundation for subsequent tasks. By taking advantage of the fact that the depth map can use connected domains with different pixel values ​​to express the shape of objects with different depth values, the target object and background in a complex environment can be effectively distinguished, ensuring the integrity of the shape and outline of the target object.

[0058] Step 2: Determine whether to perform lifting operations by setting the corresponding depth threshold.

[0059] Before detecting personnel intrusion in the dangerous area of ​​bridge crane hoisting, it is necessary to accurately determine whether cargo hoisting is in progress. If the cargo is in hoisting operation, a dangerous area based on the outline of the hoisted cargo is generated. In order to accurately determine the cargo hoisting status, the depth map obtained by the binocular camera is used, and combined with the position of the hoisted object in the binocular camera's field of view, three pixel coordinate areas are selected in the depth map. The depth value of the pixel coordinates in the area is obtained to determine whether the object is lifted. The pixel coordinate area is located within the cargo outline during the entire cargo hoisting process.

[0060] Combined with the actual working scene, according to the judgment, when the depth value of the hoisted object is less than 8.55 meters, it means that the bridge crane is in the state of hoisting the object. Analysis of the size structure of the hoisted object shows that the structure of the hoisted object is complex, and different depth values ​​exist on the same surface. In order to improve the accuracy of the bridge crane's lifting object judgment, the depth values ​​of three pixel areas in the binocular camera's field of view are selected as reference data, and the depth value involved in the judgment is the average of the depth values ​​in each area. When the average depth value of two or more areas is less than 8.55 meters, it means that the bridge crane is in the state of hoisting the object. The above object lifting judgment function is realized by using python code:

[0061]

[0062] Where D i is the average depth value of the ith region; d ij is the depth value of the jth pixel coordinate in the i-th region; N i is the total number of pixels in the i-th region, where i is the number of depth value determination regions.

[0063] The present invention combines the actual application scenario and the depth information obtained by the binocular camera to determine that the lifting operation range of the hoisting cargo is generally between 6 meters and 9 meters. Therefore, the depth image is pre-processed by setting a depth threshold (which can be customized according to different deployment operation scenarios), and the pixels with depth values ​​less than 6 meters and greater than 9 meters in the depth image are set to be consistent with the background pixels, thereby discarding the depth information within the relevant range. By setting the depth threshold to process the depth map, only the depth information between 6 meters and 9 meters is used in the subsequent task of extracting the contour of the target object, which not only effectively reduces the difficulty of the task, but also saves the computing cost to a certain extent.

[0064] Step 3: If it is determined that the object is in the lifting and transporting stage, the image processing algorithm is used to remove the redundant object connected domains in the depth map. The operation steps are as follows:

[0065] S31. After obtaining a depth map within a specified depth range by setting a depth threshold, the obtained depth map is optimized using a bilateral filter that considers both the spatial proximity weight and the pixel value similarity weight, thereby reducing the impact of noise on the contour of the object in the depth map, so as to achieve the effect of noise reduction and smoothing, and maintain the contour. The bilateral filter is improved based on the Gaussian filtering principle, calculates the spatial proximity weight of each point in the image to the center point, and then multiplies it with the pixel value similarity weight, and uses the result as the weight of the bilateral filter to participate in the filtering operation of the depth map. The calculation formula is:

[0066]

[0067] In the formula, i and j are the coordinates of the pixels involved in the convolution; k and l are the coordinates of the neighboring pixels, and their weight coefficients ω are:

[0068]

[0069] S32. After the noise reduction process is completed, the pixel values ​​of the depth image are mapped to the range of [0, 255] by using a linear normalization operation, thereby converting the depth image into a grayscale image. The linear normalization mathematical expression is as follows:

[0070]

[0071] In the formula, it is assumed that the depth map D = {d ij}, d ij represents the depth value of the i-th row and j-th column in the depth map, D min , D max They represent the minimum and maximum depth values ​​in the depth map, respectively. a and b are the normalized ranges. d ijn is the normalized depth value;

[0072] S33. In order to effectively distinguish the background and the target object, the normalized depth map is subjected to K-means clustering operation, and is divided into two categories according to the pixel range, thereby highlighting the structure of each object in the image. This clustering method iteratively optimizes the grouping of data points so that the depth information within the same cluster has a high similarity, while the depth information between different clusters is quite different. Its mathematical model is implemented by minimizing the objective function, as shown in Formula 5. The function defines the sum of the squares of the distances between each pixel and the center of its cluster, and assigns it to the cluster closest to it. Data points are assigned to each cluster by calculation, and the cluster center is updated until the cluster center no longer changes significantly, thereby achieving the classification of objects in the depth map. Through the above clustering operation, the part of the normalized depth map that exceeds the depth threshold is clustered into one category as the background, and the part within the depth threshold is clustered into one category as the target object;

[0073]

[0074] In the formula, c i is the normalized depth value of the center point of the i-th cluster; S i is the set of data points contained in the i-th cluster; d j is the normalized depth value of the jth data point; It is used to measure the compactness of data points within the same cluster. According to the clustering requirements, the value of i is 1 or 2;

[0075] Through the above clustering operation, the part of the normalized depth map that exceeds the depth threshold is clustered into one category as the background, and the part within the depth threshold is clustered into one category as the target object, such as Figure 3 As shown;

[0076] S34, from Figure 3 It can be seen that although the depth map after normalized clustering has completely separated the target object from the background, there are still redundant object connected domains that do not participate in subsequent tasks. In order to achieve the requirement of retaining only the connected domain of the target object in the depth map, the redundant connected domains in the depth map are processed by setting the pixel coordinate threshold. Specifically, according to the task requirements, the pixel coordinates of the center point of the hoisted cargo are determined. This pixel coordinate is always in the connected domain of the target object during the entire hoisting process. Through relevant algorithm operations, only the connected domain containing the pixel coordinate is retained, thereby removing other redundant connected domains. After all the above depth map processing operations, the depth map that finally only contains the target object is as shown in Figure 4 As shown, the depth map only contains the connected domain of the hoisted object, which reduces the computational cost of subsequent tasks to a certain extent and provides a favorable image basis for the subsequent extraction of the hoisted object contour.

[0077] Step 4: After completing image processing operations such as depth map filtering, the contour of the hoisted object in the depth map is extracted, thereby laying the foundation for subsequent detection of personnel intrusion into the dangerous area of ​​the hoisted object. After comprehensive analysis of various image edge detection algorithms, the present invention adopts the Canny operator edge detection algorithm with high signal-to-noise ratio and high detection accuracy. Its basic detection idea is as follows:

[0078] S41: firstly use two-dimensional Gaussian low-pass filtering to smooth the image;

[0079] Among them, a two-dimensional Gaussian filter is used to reduce the noise of the depth map. The mathematical expression is shown in Formula 6. σ is the standard deviation of the Gaussian function, which is used to control the smoothing effect of the image.

[0080]

[0081] S42: Then, the Sobel operator is used to calculate the gradient magnitude and direction of each pixel at the smoothed depth image in the horizontal and vertical directions;

[0082] The Sobel operator replaces the traditional 2×2 neighborhood first-order partial derivative template and performs gradient convolution calculation with the image window. The mathematical expression is shown in Equation 7, where D is the depth image window, the middle pixel is e, and G x and G y is the horizontal and vertical gradient of the middle pixel, S x and S y is the Sobel operator template,

[0083]

[0084] Combining the above formula, the gradient size and direction of each pixel are shown in formulas 9 and 10:

[0085]

[0086] S43: Then, non-maximum suppression is used to find the maximum value of the pixel in the gradient direction to determine whether to temporarily define the central pixel point as an edge point;

[0087] S44: Finally, a double threshold algorithm is used to determine edge points and connect and draw the outline of the object.

[0088] Non-maximum suppression is to determine whether the gradient amplitude of a certain pixel point is the maximum value in the direction along the pixel gradient direction. If so, it is determined as a possible edge point. If not, the gray value of this point is set to 0, thereby removing redundant edge pixels and accurately locating the edges of image objects. After non-maximum suppression processing, the retained local maximum points are filtered by setting two thresholds, removing edge points less than the minimum threshold, and retaining points between the two thresholds that are greater than the maximum threshold and connected to the real edge. In summary, the processing effect is shown in the figure below. Figure 5 shown.

[0089] Step 5: After completing the contour extraction of the hoisted object, use the minimum enclosing rectangle method to obtain the position information of the hoisted object, including the upper left, lower right and center point coordinates, rotation angle and other information of the minimum enclosing rectangle of the object contour, so as to obtain the position information of the object contour in the image, so as to accurately set the danger zone.

[0090] Step 6: Based on the actual application scenario, after communicating with the safety department of the production workshop, the minimum bounding rectangle of the hoisting object's own contour is used as the basis, and the contour range of 0.5 meters is set as the bridge crane hoisting danger zone. Using the extracted hoisting object contour, use the cv2.minAreaRect function provided by OPENCV to draw the minimum bounding rectangle of the hoisting object, so as to obtain the location information of this dynamic danger zone;

[0091] Step 7:

[0092] The location information of the dangerous area is mapped to the RGB image obtained by the binocular camera, providing a prerequisite for subsequent personnel intrusion detection tasks.

[0093] Compared with traditional dangerous area detection methods, the hardware deployment and labor costs of the present invention are saved by nearly 80%, which breaks the limitations of traditional dangerous area detection methods, solves the problem of difficult detection of dynamic dangerous areas, provides a new solution for dynamic dangerous area location identification, and lays a good foundation for subsequent dangerous area personnel intrusion detection.

[0094] Embodiment 2:

[0095] On the basis of the first embodiment, the present invention provides a real-time detection and positioning system for dynamic dangerous areas during cargo hoisting, including a cargo hoisting and lifting judgment module, a depth map processing module and a dynamic dangerous area identification module.

[0096] Cargo lifting judgment module: The dynamic danger zone is generated by cargo lifting and transportation operations. The danger source is the constantly moving lifting cargo. Therefore, before dynamic danger zone detection, it is necessary to determine whether lifting operations are being performed. The cargo lifting judgment module can obtain the depth value of the lifting cargo in real time, determine whether the lifting cargo is being lifted and transported, and provide a prerequisite for dynamic danger zone detection;

[0097] Depth map processing module: The depth map processing module is the basic module of the present invention. The danger source generated by the dynamic danger zone is the hoisted cargo itself. There are many redundant noise contours in the contour information of the connected domain of the hoisted cargo obtained by using the depth map. The depth map processing module can process the acquired depth map to remove the relevant noise contours so that the depth map only contains the connected domain of the hoisted cargo itself, which provides a reference basis for the subsequent dynamic danger zone identification and determination of the dangerous zone location information;

[0098] Dynamic danger zone identification module: The dynamic danger zone identification module is a functional terminal output module of the present invention. The module is based on the connected domain of the hoisted cargo in the processed depth map, and according to the actual division regulations of the dynamic danger zone, it identifies and sets the dynamic danger zone to realize real-time detection of the dynamic danger zone location information.

[0099] The above description is only a preferred embodiment of the present invention and does not limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for real-time detection and positioning of dynamic dangerous areas in cargo hoisting, characterized in that: The following steps are involved: Step 1: Install the binocular camera on the lifting equipment and use the binocular camera technology to obtain the depth map information of the hoisted object; Step 2: Determine whether to perform lifting operation by setting the corresponding depth threshold; Step 3: If the object is in the lifting and transporting stage, the image processing algorithm is used to remove the redundant object connected domains in the depth map; Step 4: After completing image processing operations such as depth map filtering, the contour of the hoisted object in the depth map is extracted; Step 5: Use the minimum enclosing rectangle method to draw the minimum enclosing rectangle of the object outline and obtain the position information of the hoisted object in the image; Step 6: Combined with the actual application scenario and the workshop safety department's requirements for the division of the lifting danger zone, the lifting danger zone is set based on the minimum circumscribed rectangle of the lifting object outline obtained above; Step 7: Map the location information of the dangerous area to the RGB image obtained by the binocular camera to provide a prerequisite for subsequent personnel intrusion detection tasks.

2. A method for real-time detection and positioning of dynamic dangerous areas of cargo hoisting according to claim 1, characterized in that: In step 2, in order to improve the accuracy of judging the lifting object of the bridge crane, the depth values ​​of three pixel areas in the field of view of the binocular camera are selected as reference data, and the depth value involved in the judgment is the average value of the depth value in each area. The above object lifting judgment function is realized by using Python code: Where D i is the average depth value of the ith region; d ij is the depth value of the jth pixel coordinate in the i-th region; N i is the total number of pixels in the i-th region, where i is the number of depth value determination regions.

3. A method for real-time detection and positioning of dynamic dangerous areas of cargo hoisting according to claim 1, characterized in that: The method of removing the redundant object connected domain in the depth map by using the image processing algorithm in step 3 is as follows: S31, using a bilateral filter to optimize the acquired depth map to achieve the effect of noise reduction and smoothing and maintaining the contour. The calculation formula is: Where i and j are the coordinates of the pixels involved in the convolution; k and l are the coordinates of the neighboring pixels, and their weight coefficients ω are: S32. After the noise reduction process is completed, the pixel values ​​of the depth image are mapped to the range of [0, 255] by using a linear normalization operation, thereby converting the depth image into a grayscale image. The linear normalization mathematical expression is as follows: Where, d ij represents the depth value of the i-th row and j-th column in the depth map, D min , D max Respectively represent the minimum and maximum depth values ​​in the depth map, a and b are the normalized ranges, and d ijn is the normalized depth value; S33. In order to effectively distinguish the background and the target object, the normalized depth map is subjected to K-means clustering operation. The clustering method optimizes the grouping of data points iteratively so that the depth information in the same cluster has a high similarity, while the depth information between different clusters has a large difference. Its mathematical model is realized by minimizing the objective function: In the formula, c i is the normalized depth value of the center point of the i-th cluster; S i is the set of data points contained in the i-th cluster; d j is the normalized depth value of the jth data point; Used to measure the closeness of data points within the same cluster; Through the above clustering operation, the part of the normalized depth map that exceeds the depth threshold is clustered into one category as the background, and the part within the depth threshold is clustered into one category as the target object; S34: To achieve the requirement of retaining only the connected domain of the target object in the depth map, the redundant connected domains in the depth map after normalization and clustering are processed by setting pixel coordinate thresholds.

4. A method for real-time detection and positioning of dynamic dangerous areas of cargo hoisting according to claim 1, characterized in that: In step 4, the Canny operator edge detection algorithm is used to draw the object contour, and the steps are as follows: S41: firstly use two-dimensional Gaussian low-pass filtering to smooth the image; S42: Then, the Sobel operator is used to calculate the gradient magnitude and direction of each pixel at the smoothed depth image in the horizontal and vertical directions; S43: Then, non-maximum suppression is used to find the maximum value of the pixel in the gradient direction to determine whether to temporarily define the central pixel point as an edge point; S44: Finally, a double threshold algorithm is used to determine edge points and connect and draw the outline of the object.

5. A method for real-time detection and positioning of dynamic dangerous areas of cargo hoisting according to claim 4, characterized in that: The mathematical expression for the two-dimensional Gaussian low-pass filtering in S41 to smooth and reduce noise on the depth map is: Where σ is the standard deviation of the Gaussian function, which is used to control the smoothing effect of the image.

6. A method for real-time detection and positioning of dynamic dangerous areas of cargo hoisting according to claim 4, characterized in that: In S42, the Sobel operator is used to replace the traditional 2×2 neighborhood first-order partial derivative template and the image window to perform gradient convolution calculation: Where D is the depth image window, the middle pixel is e, and G x and G y are the horizontal and vertical gradients of the middle pixel, S x and S y is the Sobel operator template; Combining the above formula, the gradient size and direction of each pixel are as follows:

7. A method for real-time detection and positioning of dynamic dangerous areas of cargo hoisting according to claim 4, characterized in that: The non-maximum suppression in S43 is to determine whether the gradient amplitude of a certain pixel point is the maximum value in the direction along the pixel gradient direction. If so, it is determined to be a possible edge point. If not, the grayscale value of this point is set to 0, thereby removing redundant edge pixel points and accurately locating the edge of the image object. After the non-maximum suppression process, the retained local maximum points are filtered by setting two thresholds to remove edge points less than the minimum threshold and retain points between the two thresholds that are greater than the maximum threshold and connected to the real edge.

8. The method for real-time detection and positioning of dynamic dangerous areas of cargo hoisting according to claim 1 is characterized in that: In step six, after communicating with the safety department of the production workshop, the minimum enclosing rectangle of the hoisting object's own contour is used as the basis, and the contour range of 0.5 meters is set as the bridge crane hoisting danger zone. The extracted hoisting object contour is used to use the cv2.minAreaRect function provided by OPENCV to draw the minimum enclosing rectangle of the hoisting object, so as to obtain the location information of this dynamic danger zone.

Images