Subway intelligent purging path automatic planning system and method based on Yolo6D profile recognition

Abstract

The present application relates to a kind of subway intelligent blowing path automatic planning system based on Yolo6D profile identification, including AGV blowing car equipped with mechanical arm, camera is installed at the end of the mechanical arm, mechanical arm control center is installed in the mechanical arm, and the rotation and movement of mechanical arm and mechanical arm camera are controlled by mechanical arm control center;Gyroscope and ultrasonic sensor are provided on the camera, and 2D shooting and 3D stereoscopic scanning can be carried out by the camera;It also includes Yolo6D identification module and subway database central control platform, the 2D image photographed by the camera is identified by Yolo6D identification module to obtain the part type and position, and the 3D point cloud database and the stereoscopic appearance of part obtained by 3D stereoscopic scanning are all stored in subway database. It also includes the path planning method using the path planning system. The present application adopts the stereoscopic identification technology of Yolo6D, and the automatic blowing path planning is made through the accurate identification, positioning of Yolo6D to part and executes blowing action.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of urban rail transit vehicle maintenance and repair, and relates to a subway intelligent blowing path automatic planning system and method based on Yolo6D contour recognition. BACKGROUND

[0002] In the maintenance of subway trains, the blowing of the bottom of the subway train has always been a key and difficult point of vehicle maintenance and repair. After the train runs for a period of time, the electrical components at the bottom of the subway train will accumulate a large amount of dust due to environmental factors and other factors, thereby affecting the safe and normal operation of the train. According to the requirements of the Metro Design Specification, the maintenance facilities of the subway vehicle depot should include train blowing facilities. The blowing operation is a step that must be taken before the maintenance of urban rail transit vehicles, and the main purpose is to clean the vehicle components, especially the important components at the bottom of the vehicle, so as to facilitate the maintenance of the vehicle.

[0003] According to the current market research, there is no good intelligent blowing system in the domestic vehicle depot that has been put into operation. Most of the blowing operations for the bottom of the vehicle are mainly manual operations, which produce a lot of dust and seriously pollute the environment. The operating personnel must wear protective clothing and respirators, the operating intensity is high, and the work efficiency is low. Therefore, the blowing operation has always been a pain point for the on-site operating personnel, and an intelligent blowing device is urgently needed to solve the above problems. Recently, some people in China have begun to develop subway blowing robots, which mainly use robots to replace manual work to perform blowing operations at the bottom of the subway. That is, by developing an intelligent subway train blowing robot system, the purpose is to reduce the time-consuming of subway train blowing, improve the efficiency of train blowing, and reduce the dust concentration in the environment after blowing operation to improve the working environment.

[0004] In the previous patent of the inventor, the main blowing judgment needs to define the blowing method of various components in advance, and there is no intelligent automatic judgment function. That is, the classification of the components at the bottom of the subway cannot be known, automatic blowing actions cannot be performed according to the appearance of the components, or no blowing action is defined (directly violent blowing), or the shortest blowing path is defined. For example, the components of the subway are cylindrical or triangular prismatic, and the blowing direction should be different normal angles, and the best blowing path is as shown in Figure 1 However, the above system does not define it. SUMMARY

[0005] The present application provides a subway intelligent blowing path automatic planning system and method based on Yolo6D contour recognition to solve the technical problems in the prior art. The main function is to remove dust from the components at the bottom of the subway vehicle on both sides of the lower part of the subway vehicle, accurately identify and locate the components, and plan an automatic blowing path.

[0006] The application comprises the following technical solutions: a subway intelligent blowing path automatic planning system based on Yolo6D contour recognition, comprising an AGV blowing vehicle provided with a mechanical arm, the mechanical arm being installed on the base of the AGV blowing vehicle, a camera being installed at the end of the mechanical arm, a mechanical arm control center being installed in the mechanical arm, the mechanical arm control center controlling the rotation of the mechanical arm and the camera and the movement of the mechanical arm; a gyroscope and an ultrasonic sensor being arranged on the camera, the camera being capable of 2D shooting and 3D stereoscopic scanning; further comprising a central control platform provided with a Yolo6D recognition module and a subway database, the 2D images shot by the camera being subjected to part recognition by the Yolo6D recognition module to obtain the part model and position, and the 3D point cloud database obtained by 3D stereoscopic scanning and the stereoscopic appearance of the parts being stored in the subway database.

[0007] A subway intelligent blowing path automatic planning method based on Yolo6D contour recognition, using the path planning system as described above, comprising the following steps:

[0008] S1, when the AGV blowing vehicle reaches the specified location, the camera is carried by the mechanical arm, and the Yolo6D technology is used to determine the type, shape and pose of the parts above;

[0009] S2, first, the part position error is determined, and the part is classified;

[0010] S3, the shape of the part is determined, and a plurality of surfaces to be blown are determined;

[0011] S4, Z-shaped blowing is adopted, and obstacle judgment is performed before blowing, and 2D shooting is performed on the surface of the part before and after the start of blowing to determine whether the cleaning is up to standard;

[0012] S5, after the blowing of all parts is completed, the central control platform is waited for to inform the AGV blowing vehicle that the end point (i.e. the subway train tail) has been reached, and the mechanical arm is moved back to the waiting position.

[0013] Further, the part position error determination process in S2 comprises the following steps:

[0014] C1, if the 2D image is found to contain a part, the relative position of the part in the 2D image to the subway train and the ground, the relative position of the mechanical arm to the AGV blowing vehicle, and the absolute position of the AGV blowing vehicle are used to calculate whether the part is consistent with the position in the database; if there is an error of more than 1 cm, the central control system is informed, the positioning error of the AGV blowing vehicle is recorded, and a specified blowing mode is started;

[0015] C2, if the 2D image is found to be missing again, it indicates that the AGV blowing vehicle has a serious positioning error, at which time the central control system is notified, and the AGV blowing vehicle is started to move forward or backward by 30 cm, and the position of the part is searched, and S1 is repeatedly executed.

[0016] Further, the part classification process in S2 includes the following steps:

[0017] D1, collect all the 2D photos of the parts of the subway chassis in advance;

[0018] D2, execute Yolo6D model, pre-train part classification action, and achieve more than 90% accuracy;

[0019] D3, perform 3D scanning on the part through the camera of the mechanical arm to obtain a 3D point cloud database;

[0020] D4, calculate the three-dimensional information of the part using Yolo6D based on the 3D point cloud database, and classify the part according to the three-dimensional information of the part; when Yolo6D is calculated, 23 convolution layers and 5 maximum pooling layers are used. First, input a complete 2D color image, process it with the network and segment it into SxS grid. The network outputs a 3D tensor of SxSx(9x2+1+C). Each grid in the three-dimensional output tensor corresponds to an output multi-dimensional vector, which includes the predicted two-dimensional image position of 9 control points, the class probability of the target and the overall confidence (i.e. whether there is an object). When entering the subway site and starting the action, the three-dimensional information of the part (9 control points, that is, the coordinate position of 6 faces, three-dimensional block centroid, etc.) is calculated using Yolo6D, which can calculate the normal position of the blowing and the blowing area. Data, then use the Z-shaped loop path to start the automatic blowing action.

[0021] Further, the blowing process in S3 includes the following steps:

[0022] M1, start the mechanical arm, start scanning the point cloud, and take a 2.5D depth map, execute Yolo6D model to obtain part hexahedral data;

[0023] M2, find 5 planes that can be blown by the mechanical arm according to the 9 control points calculated by the obtained Yolo6D model, and calculate the normal vector of each plane according to the centroid point;

[0024] M3, move the mechanical arm to the front of each plane, maintain the normal position of the blowing nozzle and the plane, and implement automatic blowing on the plane;

[0025] M4, the sweeping mode is that four vertices of each plane are necessary points, and a straight line formula is calculated for each two vertices, and the four straight lines are used as boundaries;

[0026] M5, starting from one vertex, sweeping to another vertex, and then calculating a new boundary point of the other boundary relative to the vertex;

[0027] M6, repeatedly performing M5 until more than four vertex positions or more than boundaries are exceeded, and the scanning mode is a zigzag automatic sweeping.

[0028] Further, the obstacle judgment process in S4 includes the following steps:

[0029] E1, in order to judge whether there is an obstruction of the pipeline in front, the system first uses Hough linear transformation to convert the image in the picture into a binary graph of edge line segments according to the collected 2D image, as shown in (a) is the original graph, (b) is the converted edge line graph, and the binary graph is combined with the depth graph to find the position and distance of the line segment in the picture; Figure 7

[0030] E2, according to the position and distance of the line segment in the picture, the distance between the part and the obstacle is preset to a threshold value, according to the position and distance of the line segment in the picture, whether there is a line segment within the threshold value in front of the part is found, so as to judge whether the unknown obstacle exists or not, if the distance between the line segment and the camera is less than the threshold value, the information that the part cannot be swept is sent to the central control platform, and whether the sweeping is suspended is returned by the central control platform;

[0031] E3, if there is no obstacle, the sweeping process is started, the robot arm is moved to the set point, the intelligent sweeping action is executed through the pre-planned sweeping path, the path correction instruction from the central control platform, and the judgment of the on-site point cloud scanning.

[0032] Further, the judgment of whether the cleaning is up to standard in S4 includes the following steps:

[0033] F1, before starting the sweeping, the 2D photograph of the surface of the swept part is executed again, and after the sweeping, the robot arm returns to above the base, and the 2D photograph is executed again, and the photograph comparison is executed; the photograph is converted into a negative format, and the speckles in the two negative films are compared, the number and degree of the speckles are quantified, and the value is taken as the probability after cleaning, which is returned to the central control platform, and whether the part needs to be cleaned again is judged by the central control platform;

[0034] F2, if the part needs to be swept again, the central control platform returns the same part model and the same position to the robot arm control center;

[0035] F3, if the part does not need to be swept again, the central control platform returns the next part model and different position to the robot arm control center.​

[0036] The present application has the advantages and positive effects:

[0037] 1、The system can classify the parts at the bottom of the subway, can establish the blowing path of the parts according to the stereogram returned by Yolo6D, find out the normal direction of the blowing nozzle according to the position of the hexahedron, directly establish the automatic blowing path for the area of each part, and execute the blowing action, that is, can execute the automatic blowing action according to the appearance of the parts.

[0038] 2、2D photographing is performed on the surface of the parts before and after starting blowing, and the impurities in the 2D picture negative before blowing and the 2D picture negative after executing blowing are compared, the number of impurities is quantified to judge whether the cleaning is clean, and the cleaning quality is ensured. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1 is the preferred blowing path schematic diagram of the circular cross-section part.

[0040] Figure 2 is the preferred blowing path schematic diagram of the triangular cross-section part.

[0041] Figure 3 is the blowing process schematic diagram of the present application.

[0042] Figure 4 is the Yolo6D architecture diagram.

[0043] Figure 5 is the Yolo6D identification result schematic diagram.

[0044] Figure 6 is the Z-shaped planning diagram of automatic blowing.

[0045] Figure 7 is the Hough linear conversion schematic diagram. DETAILED DESCRIPTION

[0046] In order to further disclose the invention content, characteristics and effects of the present application, the following examples are specifically illustrated as follows in combination with the drawings.

[0047] Refer to the drawings Figures 3-7An automated planning system for intelligent sweeping paths in subways based on Yolo6D contour recognition includes an AGV sweeping vehicle equipped with a robotic arm. The robotic arm is mounted on a base, and a camera is installed at the end of the robotic arm. A robotic arm control center is installed inside the robotic arm, which controls the rotation and movement of the robotic arm and the camera. The camera is equipped with a gyroscope and an ultrasonic sensor, and can perform 2D imaging and 3D stereo scanning. The system also includes a Yolo6D recognition module and a subway database central control platform. The 2D images captured by the camera are used by the Yolo6D recognition module to identify the parts, obtaining the part models and locations. The 3D point cloud database obtained by the 3D stereo scanning and the stereo appearance of the parts are all stored in the subway database.

[0048] First, during the purging process, we collected a large number of photos of all the components on the subway's underside. We then trained and categorized these photos, using a YOLO6D model for target inspection. Simultaneously, we recorded the component location information for each frame, along with the converted point cloud data, in the database. Because the robotic arm's camera is equipped with a gyroscope, the arm's movement can be calculated, allowing the camera to capture the absolute position of each frame. Using this information, we can calculate the global position of the components in each frame.

[0049] Yolo6D operations utilize a total of 23 convolutional layers and 5 max-pooling layers, such as... Figure 4 As shown, a complete 2D color image is first input, processed by a network, and segmented into an S×S mesh. The network outputs an S×S×(9×2+1+C) 3D tensor. Each mesh in the 3D output tensor corresponds to a multi-dimensional vector, which includes the predicted 2D image position of 9 control points, the target class probability, and the overall confidence (i.e., whether the object exists). When entering the subway site and starting the action, YOLO6D is used to calculate the 3D information of the components (the position and dimensions of the 9 control points, i.e., 6 faces), which can be used to deduce the normal position of the purging, the purging area, and other data. Then, an automatic purging action is started using a Z-shaped pattern.

[0050] Example 2: See Appendix Figures 3-7 An automatic path planning method for intelligent subway cleaning based on Yolo6D contour recognition is presented. Using the path planning system described above, the method includes the following steps: After the subway stops, the undercarriage cover is disassembled and stored, and personnel evacuate after completion. After confirming the safety of the work area, three AGV cleaning vehicles are powered on, started, and moved to the starting point. The robotic arm camera is moved to the front of the train, activated, and scans the front of the train to confirm its position and begin executing the predefined cleaning task.

[0051] S1, when the AGV blowing car arrives at the designated place, the camera is carried by the mechanical arm, and the Yolo6D technology is used to judge the type, shape and pose of the upper parts.

[0052] S2, first judge the position error of the parts, and classify the parts;

[0053] The part position error judgment process in S2 includes the following steps:

[0054] C1, if the parts exist in the 2D image, the relative position of the parts in the 2D image to the subway train and the ground, the relative position of the mechanical arm to the AGV blowing car, and the absolute position of the AGV blowing car are used to calculate whether the parts are consistent with the position in the database; If there is an error of more than 1cm, inform the central control system, record the positioning error of the AGV blowing car, and start to execute the specified blowing mode (fast, basic, deep blowing, etc. Three kinds);

[0055] C2, if the parts do not exist in the 2D image, it means that the AGV blowing car has serious positioning error, at this time, the central control system is informed, and the AGV blowing car is started to move forward or backward by 30cm, the part position is searched, and S3 is repeatedly executed.

[0056] The part classification process in S2 includes the following steps:

[0057] D1, collect all the 2D photos of the parts of the subway chassis in advance;

[0058] D2, execute Yolo6D model, pre-train part classification action, and achieve more than 90% accuracy;

[0059] D3, perform 3D scanning on the parts through the mechanical arm camera to obtain 3D point cloud database;

[0060] D4, calculate the three-dimensional information of the parts by Yolo6D for 3D point cloud database, and classify the parts according to the three-dimensional information of the parts; such as Figures 4-5As shown, when Yolo6D operation is performed, 23 convolution layers and 5 maximum pooling layers are used in total. First, a complete 2D color image is input, processed by the network, and segmented into SxS grids. The network outputs a 3D tensor of SxSx(9x2+1+C). Each grid in the three-dimensional output tensor corresponds to an output multi-dimensional vector, which includes the predicted two-dimensional image positions of the nine control points, the class probability of the target, and the overall confidence (i.e., whether there is an object). When entering the subway site and starting to perform actions, the three-dimensional information of the parts (9 control points, i.e., the coordinate positions of the 6 faces, the center of the three-dimensional block, etc.) is calculated by Yolo6D, the normal position of the blowing and the area of the blowing can be calculated, and then the Z-shaped loop method is used to start the automatic blowing action.

[0061] S3, the shape of the part is determined, and several faces to be blown are determined.

[0062] S4, Z-shaped blowing is adopted, and obstacle judgment is performed before blowing. 2D photographing is performed on the surface of the part before and after starting blowing to determine whether the cleaning meets the standard;

[0063] The obstacle judgment process in S4 includes the following steps:

[0064] E1, in order to judge whether there is an obstruction of pipeline in front, the system first uses Hough linear transformation to convert the image in the picture into a binary image of edge line segment according to the collected 2D image, and combines the binary image with the depth image to find the position and distance of the line segment in the picture;

[0065] E2, according to the characteristics of the distance of the object, it is found out whether there is a close distance line segment in front of the part before moving, so as to judge whether there is an unknown obstacle. If the distance is less than the threshold value (the distance of the line segment obstacle from the camera), the information that the part cannot be blown is sent to the central control platform, and whether the blowing is suspended is returned by the central control platform;

[0066] E3, if there is no obstacle, the blowing process is started, the mechanical arm is moved to the designated point, the intelligent blowing action is performed through the pre-planned blowing path, the path correction instruction from the central control platform, and the judgment of the on-site point cloud scanning.

[0067] The blowing process in S4 includes the following steps:

[0068] M1, start the mechanical arm, start scanning the point cloud, and take a 2.5D depth map, execute the Yolo6D model to obtain the hexahedral data of the part;

[0069] M2, according to the nine control points calculated by the obtained Yolo6D model, find out the five planes that can be blown by the mechanical arm, and calculate the normal vector of each plane according to the center point.

[0070] M3, move the mobile robot arm in front of each face, maintain the position of the blowing nozzle and the normal line of the plane, and automatically blow the plane;

[0071] M4, the blowing method is that four vertices of each plane are the necessary points, and a straight line formula is calculated for each two vertices, and the four straight lines are used as boundaries;

[0072] M5, starting from one vertex, blowing to another vertex, and then calculating a new boundary point of the vertex relative to another boundary;

[0073] M6, repeatedly execute M5 until the position of more than four vertices or the boundary is exceeded, and the scanning method is zigzag automatic blowing.

[0074] The obstacle judgment process in S4 includes the following steps:

[0075] E1, in order to judge whether there is an obstruction of the pipeline in front, the system first uses Hough linear transformation to convert the image in the picture into a binary image of edge line segments according to the collected 2D image, as shown in (a) is the original image, (b) is the converted edge line image, and the binary image is combined with the depth image to find the position and distance of the line segment in the picture; Figure 6

[0076] E2, according to the position and distance of the line segment in the picture, the distance between the part and the obstacle is preset to a threshold value, according to the position and distance of the line segment in the picture, whether there is a line segment within the threshold value in front of the part is found, so as to judge whether the unknown obstacle exists or not, if the distance between the line segment and the camera is less than the threshold value, the information that the part cannot be blown is sent to the central control platform, and whether the blowing is suspended is returned by the central control platform;

[0077] E3, if there is no obstacle, the blowing process is started, the robot arm is moved to the designated point, the intelligent blowing action is executed through the pre-planned blowing path, the path correction instruction from the central control platform, and the judgment of the on-site point cloud scanning.

[0078] The judgment of whether the cleaning is up to standard in S4 includes the following steps:

[0079] F1, before starting blowing, the 2D photograph of the surface of the part to be blown is executed again, and after blowing, the robot arm returns to above the base, and the 2D photograph is executed again, and the photograph comparison is executed; the photograph is converted into a negative format, and the speckles in the two negative films are compared, the number and degree of the speckles are quantified, and the value is taken as the probability after cleaning, which is returned to the central control platform, and whether the part needs to be cleaned again is determined by the platform;

[0080] ​F2, if the need for re-blowing, the control platform is returned to the same model, the same position of the mechanical arm control center parts;

[0081] F3, if the need for re-blowing, the control platform is returned to the next model, different position of the mechanical arm control center parts.

[0082] S5, after the execution of all parts of the blowing, waiting for the control notice AGV blowing car has reached the end (i.e. subway tail), while moving the mechanical arm back to the waiting position.

[0083] Although the above describes the preferred embodiments of the present application, the present application is not limited to the above-described specific embodiments, the above-described specific embodiments are only illustrative, not limiting, those of ordinary skill in the art can make many forms under the inspiration of the present application without departing from the purpose of the present application and the scope of the claims. These are within the scope of the present application.

Claims

1. A subway intelligent purging path automatic planning method based on Yolo6D contour recognition, using a subway intelligent purging path automatic planning system based on Yolo6D contour recognition, comprising an AGV purging vehicle equipped with a mechanical arm, characterized in that: The mechanical arm is installed on a base, a camera is installed at the end of the mechanical arm, a mechanical arm control center is installed in the mechanical arm, and the mechanical arm control center controls the rotation and movement of the mechanical arm and the mechanical arm camera; a gyroscope and an ultrasonic sensor are arranged on the camera, the camera can take 2D pictures and 3D stereoscopic scanning; further comprising a Yolo6D recognition module and a subway database central control platform, the 2D images taken by the camera are subjected to part recognition by the Yolo6D recognition module to obtain the part model and position, and the 3D point cloud database and the stereoscopic appearance of the part obtained by 3D stereoscopic scanning are stored in the subway database, and the characteristic is that the following steps are included: S1 When the AGV blowing vehicle arrives at the specified location, the mechanical arm drives the mechanical arm camera, and uses the Yolo6D technology to judge the type, shape and pose of the part above; S2 First, judge the part position error, and classify the part; S3 Determine the shape of the part and the number of surfaces to be blown; S4 Z-shaped blowing is adopted, and obstacle judgment is performed before blowing, and 2D pictures are taken on the surface of the part before and after blowing to judge whether the cleaning meets the standard; S5 After executing the blowing of all parts, wait for the central control platform to inform the AGV blowing vehicle that it has arrived at the end point, and move the mechanical arm back to the waiting position.

2. The subway intelligent purging path automatic planning method based on Yolo6D contour recognition according to claim 1, characterized in that: The part position error judgment process in S2 includes the following steps: C1, if the 2D image is found to exist again, the relative position of the part in the 2D image, the relative position of the mechanical arm, and the absolute position of the AGV blowing vehicle are used to calculate whether the part position is consistent with the database position; if there is an error of more than 1cm, inform the central control system, record the positioning error of the AGV blowing vehicle, and start executing the specified blowing mode; C2, if the 2D image is found to be missing again, it means that the AGV blowing vehicle has a serious positioning error, so inform the central control system and start moving forward and backward 30cm to search for the part position and repeat S1.

3. The subway intelligent purging path automatic planning method based on Yolo6D contour recognition according to claim 1, characterized in that: The part classification process in S2 includes the following steps: D1, collect all subway chassis part 2D photos in advance; D2, execute Yolo6D model, pre-train part classification action, and achieve more than 90% accurate recognition; D3, execute 3D scanning on the part through the mechanical arm camera to obtain the 3D point cloud database; D4, calculate the 3D information of the part using Yolo6D for the 3D point cloud database to establish a global blowing path; when Yolo6D is operated, 23 convolution layers and 5 maximum pooling layers are used, a complete 2D color image is first input, the network is processed and segmented into S×S grid; the network outputs a 3D tensor of S×S×(9×2+1+C), each grid in the three-dimensional output tensor corresponds to an output multi-dimensional vector, which includes the predicted two-dimensional image position of 9 control points, the class probability of the target and the overall confidence.

4. The subway intelligent purging path automatic planning method based on Yolo6D contour recognition according to claim 1, characterized in that: The blowing process in S3 includes the following steps: M1, start the robot arm, start scanning the point cloud, and take the point cloud depth map, execute the Yolo6D model to obtain the six-hedron data of the part; M2, find out the five faces that can be swept by the robot arm according to the nine control points calculated by the Yolo6D model, and calculate the normal vector of each face according to the centroid point; M3, move the robot arm to the front of each face, maintain the position of the blowing nozzle and the normal of the plane, and implement automatic blowing on the plane; M4, the blowing method is that the four vertices of each plane are the points that must be passed through, and a straight line formula is calculated for each two vertices, and the four straight lines are used as boundaries; M5, start from one vertex, sweep to another vertex, and then calculate the new boundary point of the other boundary relative to the vertex; M6, repeatedly execute M5 until the position of the four vertices is exceeded or the boundary is exceeded, and the scanning method is Z-shaped automatic blowing.

5. The subway intelligent purging path automatic planning method based on Yolo6D contour recognition according to claim 1, characterized in that: The obstacle judgment process in S4 includes the following steps: E1, in order to judge whether there is a pipeline line in front of the barrier, the system first uses Hough linear transformation to convert the image in the picture into a binary image of edge line segments according to the collected 2D image, and combines the binary image with the depth map to find the position and distance of the line segment in the picture; E2, according to the position and distance of the line segment in the picture, the distance between the part and the obstacle is preset to a threshold value, and according to the position and distance of the line segment in the picture, it is found out whether there is a line segment within the threshold value in front of the part, so as to judge the existence of unknown obstacles, if the distance between the line segment and the camera is less than the threshold value, the information that the part cannot be swept is sent to the central control platform, and whether the sweeping is suspended is returned by the central control platform; E3, if there is no obstacle, start the blowing process, move the robot arm to the designated point, execute intelligent blowing action through the pre-planned blowing path, plus the path correction instruction from the central control platform, and the judgment of the on-site point cloud scanning.

6. The subway intelligent purging path automatic planning method based on Yolo6D contour recognition according to claim 1, characterized in that: The judgment of whether the cleaning is up to standard in S4 includes the following steps: F1, before starting blowing, perform 2D photography of the part surface again, and after blowing, move the robot arm to the top of the base, perform 2D photography again, and perform photo comparison; convert the photos to negative format, compare the speckles in the two negatives, quantify the number and degree of speckles, and return the value to the central control platform as the probability of cleaning, and determine whether it needs to be cleaned again by the central control platform; F2, if it needs to be blown again, the central control platform returns the same part model and position to the robot arm control center; F3, if it does not need to be blown again, the central control platform returns the next part model and position to the robot arm control center. ​

Images