Environment perception system for unmanned intelligent torpedo tank truck

Abstract

The application provides an unmanned intelligent torpedo tank truck environment sensing system, which comprises a communication system, an obstacle identification system and a picture display system; the communication system acquires point cloud information and picture information of the environment around the torpedo tank truck; the obstacle identification system performs environment sensing by analyzing the point cloud information from the laser radar and the picture information from the visual sensor, and identifies obstacles invading the current torpedo tank truck operation limit; and the picture display system displays the point cloud information, the picture information, the operation limit area and the identification result of the obstacles in real time. The application mainly aims at the unmanned supervision state of the torpedo tank truck during the full-automatic operation in the factory area, can effectively detect and identify obstacles threatening the driving safety, and ensures the safety of the iron and water transportation, so as to prevent the torpedo tank truck from colliding with objects invading the operation line accidentally.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving, and more specifically, to an environmental perception system for an unmanned intelligent torpedo tanker. Background Technology

[0002] Torpedo ladle cars are large-scale molten iron transport equipment. Their tanks can store molten iron and perform desulfurization and dephosphorization processes during transport. They also feature a tipping mechanism. Torpedo ladle cars have a large storage capacity, low heat loss, and long heat preservation time, simplifying steelmaking processes and saving transportation costs. However, in steel plants, there is often a problem of low turnover due to insufficient manned torpedo ladle capacity. Unmanned intelligent torpedo ladle cars can achieve autonomous operation and management, increasing transport capacity and thus solving this problem.

[0003] Currently, environmental perception algorithms in the field of autonomous driving mainly rely on hardware such as visual sensors and LiDAR, and utilize algorithms such as deep learning to perform functions such as road lane line recognition and obstacle identification. However, for torpedo tankers, such as SmartTPC, operating in a tracked environment, especially within factory areas, these perception algorithms are no longer universally applicable due to the characteristics of numerous lines, switches, crossings, vegetation, and uneven road surfaces. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the purpose of this invention is to provide an environmental perception system for an unmanned intelligent torpedo tanker.

[0005] An environmental perception system for an unmanned intelligent torpedo tanker truck according to the present invention includes:

[0006] The communication system acquires point cloud information and image information of the environment surrounding the torpedo tanker.

[0007] Obstacle recognition system: It performs environmental perception by analyzing point cloud information from lidar and image information from visual sensors, and identifies obstacles that intrude into the current torpedo tanker's operating clearance.

[0008] The display system displays point cloud information, image information, operating boundary areas, and obstacle recognition results in real time.

[0009] Preferably, in the obstacle recognition system:

[0010] The ground is fitted based on point cloud information and image information, and the railway track point cloud is extracted.

[0011] Operational clearance calculations are performed using the torpedo car's pose, trajectory, and rail point cloud.

[0012] Based on point cloud information and image information, obstacles are identified and classified, the distance and size of obstacles are calculated, and it is determined whether the obstacles are within the operating limits.

[0013] Preferably, in the obstacle recognition system, the uneven ground is fitted using a rasterization method based on the point cloud information to obtain a rasterized ground; based on the rasterized ground, the railway point cloud is extracted, and combined with the current pose and running path of the torpedo tanker, the deviation of the running path is corrected, and the running limit of the unmanned intelligent torpedo tanker is calculated.

[0014] Preferably, determining whether an obstacle is within the operating limits includes:

[0015] Semantic analysis is performed on the original input 3D point cloud information to label whether each point belongs to shrub vegetation;

[0016] Based on the current pose of the torpedo tanker, the mission path is transformed into the relative coordinate system of the torpedo tanker to obtain the relative mission path.

[0017] Extract the waypoints of the relative task path within the required detection range, and fit the equation expression of the relative task path.

[0018] Based on the equation expression of the relative task path and the lateral threshold, the region of interest is defined and the 3D point cloud within the region is extracted;

[0019] An uneven ground surface is represented using a rasterization method, and the elevation within each ground grid is determined.

[0020] Based on the elevation difference between the three-dimensional point cloud and its corresponding ground grid, point clouds that may be railway tracks are extracted.

[0021] The point cloud that may be a railway track is filtered based on the lateral distance difference to the relative task path.

[0022] Based on the filtered rail point cloud and the relative task path, an error function is established and nonlinear optimization is performed to maximize the match between the relative task path and the centerline of the rail point cloud, thereby eliminating the angular error of the relative task path.

[0023] Calculate the runtime limits based on the corrected relative task path;

[0024] Perform height filtering based on rasterized ground on the 3D point cloud within the region of interest;

[0025] Point clouds that have been filtered and marked as shrub vegetation are removed.

[0026] The remaining point cloud is segmented and clustered;

[0027] Based on the distance and size of the clustered point cloud, the corresponding region of the point cloud within the image captured by the visual sensor is inferred;

[0028] Pixels within the visual sensor image are selected, and deep learning is used to identify the types of pixels one by one, and the types are matched with the clustered point cloud.

[0029] If an obstacle point cloud is identified through the corresponding identification, a decision to avoid it is made based on the distance, size, and type of the obstacle point cloud.

[0030] Preferably, the method of representing uneven ground using a gridded approach includes:

[0031] Select the detection area based on the relative path;

[0032] Use height filtering to remove point clouds with higher threshold values;

[0033] Set the grid size and create a two-dimensional grid matrix in a horizontal plane;

[0034] Iterate through each grid cell sequentially, and use the lowest height of all point clouds within the grid cell as the height of this grid cell; if there is no point cloud in the grid cell, mark it as no point cloud;

[0035] Iterate through all grid cells without point clouds. If not all adjacent grid cells are without point clouds, set the weights using a two-dimensional Gaussian distribution and use the weighted average height as the height of the grid cell. Continue until all grid cells have a height value.

[0036] Traverse all grids and erode higher grids using adjacent lower grids. That is, if the height of an adjacent grid is less than the height of the current grid, the height of the current grid is reset to the lowest height among the adjacent grids.

[0037] Preferably, the correction path includes:

[0038] The extracted rail point cloud is distinguished and divided into left rail points and right rail points;

[0039] Randomly extract some points from the left side of the railway track and some points from the right side of the railway track;

[0040] The correction angle variable of the railway point cloud is introduced and initial values ​​are set, and a rotation matrix is ​​constructed;

[0041] A rotation matrix is ​​applied to a portion of the extracted left-side rail points and a portion of the right-side rail points to obtain a rotated point cloud.

[0042] Construct an error function, which is the sum of the differences between the distances from each point cloud after rotation to the relative path equation and half the standard track gauge;

[0043] The error function is nonlinearly optimized to obtain the optimal correction angle for the rail point cloud;

[0044] Evaluate the correction angle of this optimal rail point cloud. If it is better than the initial value of the correction angle variable, retain the correction angle of the optimal rail point cloud; otherwise, discard it. Continue until the maximum number of iterations is reached and the final correction angle is obtained.

[0045] Preferably, the obstacle recognition system fuses and complements information from lidar and visual sensors:

[0046] Using a deep learning model of 3D point cloud, semantic analysis is performed on the input point cloud to label the point clouds of trees and vegetation without taking braking measures on the trees and vegetation.

[0047] Based on the rasterized ground, using an elevation threshold, point clouds above the torpedo tanker chassis are extracted and clustered; based on the size and location of the clustered point clouds, the corresponding range of the clustered point clouds in the image captured by the visual sensor is calculated, and a correspondence is established.

[0048] Using a deep learning model for images, the two-dimensional images corresponding to the clustered point clouds are sequentially classified to determine whether they are pedestrians and vehicles that pose a threat to traffic, or trees and vegetation that do not pose a threat to traffic.

[0049] Preferably, the communication system communicates with the onboard PLC control system of the torpedo tanker to obtain the torpedo tanker's travel route information, including obtaining the coordinate point set of the current travel route of the torpedo tanker and transmitting the location, size and type information of the identified obstacles.

[0050] Preferably, the communication system communicates with the positioning module to obtain the current pose information, current position and heading information of the torpedo tanker, as well as to send the identified obstacle information.

[0051] Preferably, when an obstacle is detected that has intruded into the current operating clearance of the torpedo tanker, the torpedo tanker is controlled to brake.

[0052] Compared with the prior art, the present invention has the following beneficial effects:

[0053] 1. This invention is mainly aimed at the situation where torpedo tankers are in a state of unattended operation during fully automated operation within the factory area. In order to prevent torpedo tankers from colliding with objects that accidentally intrude into the operating route and causing safety accidents, an environmental perception system for torpedo tankers is designed.

[0054] 2. This invention utilizes an environmental perception system for torpedo tank cars, enabling them to perceive the environment during automatic operation in complex rail-based conditions and effectively detect and identify obstacles that threaten driving safety.

[0055] 3. This invention enables the detection and identification of obstacles within the operating clearance, allowing the torpedo car to brake in time before a collision, thus ensuring the safety of molten iron transportation. Attached Figure Description

[0056] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0057] Figure 1 This is a structural diagram of the torpedo tanker control system. Detailed Implementation

[0058] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0059] This invention discloses an environmental perception system primarily based on lidar and visual sensors, providing obstacle perception and identification along the path for a torpedo tank car (unmanned intelligent torpedo tank car) transporting molten iron in a rail-based environment. The environmental perception system includes: a communication system, an obstacle identification system, a display system, and a data recording system.

[0060] The communication system is primarily used to communicate with sensors and other system modules of the torpedo tanker, including communication with lidar and vision sensors. Communication with lidar obtains point cloud information of the surrounding environment, while communication with vision sensors obtains image information. Communication with the torpedo tanker's onboard PLC control system obtains the torpedo tanker's driving route information, including the coordinate point set of the current trajectory of the unmanned intelligent torpedo tanker, and transmits information such as the location, size, and type of identified obstacles for the onboard PLC control system to make obstacle avoidance decisions. Communication with the positioning module obtains the torpedo tanker's current pose, current position, and heading information, as well as transmits information about identified obstacles.

[0061] The obstacle recognition system primarily analyzes and processes point cloud data from LiDAR and image information from visual sensors to perceive the environment and identify obstacles that intrude into the current clearance zone of the torpedo tanker, ensuring the safe operation of the torpedo tanker. The obstacle recognition system includes a road surface fitting module, a rail detection module, a path fitting module, a clearance calculation module, and an obstacle recognition and classification module. Specifically, the obstacle recognition system receives and processes point cloud information from LiDAR, fits the ground surface, and extracts rail point cloud data. By combining the pose, operating path, and rail point cloud information of the unmanned intelligent torpedo tanker, accurate clearance calculations are performed, reducing the false alarm probability of obstacles intruding into the clearance zone. By combining the point cloud data from LiDAR and image information captured by visual sensors, obstacles are identified and classified, their distance and size are calculated, and it is determined whether they are within the clearance zone and affect the operational safety of the unmanned intelligent torpedo tanker. In the obstacle recognition system, the uneven ground is fitted using a rasterization method based on point cloud information. Based on the rasterized ground, the railway point cloud is extracted, and combined with the current pose and running path of the unmanned intelligent torpedo tanker, the deviation of the running path is corrected, and the running clearance of the unmanned intelligent torpedo tanker is accurately calculated. By combining the point cloud of lidar and the image information captured by the visual sensor, obstacles are identified and classified, their distance and size are calculated, and it is determined whether they are within the clearance and affect the operational safety of the unmanned intelligent torpedo tanker.

[0062] More specifically, the obstacle recognition system is used to identify whether there are obstacles affecting driving safety within the operating limits of the unmanned intelligent torpedo tanker, ensuring driving safety. Its implementation involves: semantic analysis of the original input 3D point cloud information to label whether each point belongs to shrub vegetation; transforming the task path to the relative coordinate system of the unmanned intelligent torpedo tanker based on its current pose to obtain the relative task path; extracting waypoints within the detection range of the relative task path and fitting them using a cubic polynomial to obtain the equation expression of the relative task path; defining the region of interest based on the equation expression of the relative task path and a lateral threshold, and extracting the 3D point cloud within that region; representing uneven ground using a rasterization method to determine the elevation within each ground grid; extracting point clouds that may be railway tracks based on the elevation difference between the point cloud and its corresponding ground grid; and filtering point clouds that may be railway tracks based on the lateral distance difference to the relative task path. Based on the filtered rail point cloud and the relative task path, an error function is established and nonlinear optimization is performed using the Gauss-Newton method to maximize the alignment between the relative task path and the centerline of the rail point cloud, eliminating angular errors in the relative task path. The operating clearance is calculated based on the corrected relative task path. Height filtering based on rasterized ground is applied to the 3D point cloud within the region of interest. Points marked as vegetation in the filtered point cloud are removed. The remaining point cloud is segmented and clustered. Based on the distance and size of the clustered point clouds, their corresponding regions within the image captured by the visual sensor are estimated. Pixels within the visual sensor image are selected, and deep learning is used to sequentially identify their types, mapping the types to the clustered point clouds. If point clouds identified as pedestrians, vehicles, or other threats to driving safety exist, it is calculated whether they are within the operating clearance. If obstacles affecting driving safety exist within the clearance, their distance, size, and type are output for the onboard PLC control module to make decisions.

[0063] The image display system is used to display in real time the point cloud and image information captured and fed back by lidar and visual sensors, the operating limit area of ​​the unmanned intelligent torpedo tanker, and the recognition results of obstacles, and to label the categories of obstacles, so that users can intuitively observe the recognition results of the obstacle recognition system.

[0064] The data recording system is primarily used to record data such as the positioning coordinates, mission path, pose information, and sensor information of the torpedo tanker during operation, facilitating offline testing and troubleshooting. Specifically, the data recording system records the mission path, pose information, and sensor data of the unmanned intelligent torpedo tanker during operation, allowing developers to utilize this data for offline testing and problem diagnosis.

[0065] The present invention will now be described in more detail.

[0066] This invention addresses the unique rail-based operating environment of unmanned intelligent torpedo tankers by establishing an environmental perception system based on lidar and visual sensors. This system can effectively detect and identify obstacles within the driving limits of the unmanned intelligent torpedo tanker that threaten driving safety, and issue corresponding obstacle information and warnings.

[0067] The following section provides a more detailed explanation of ground rasterization.

[0068] This invention addresses the problem of uneven ground in the rail-based operation environment of unmanned intelligent torpedo tank trucks, making it difficult to approximate ground height using a single continuous mathematical model. It employs a rasterization method to process 3D point clouds to represent ground elevation. The processing flow of this method is as follows:

[0069] a. Select a general detection area based on the relative path to reduce unnecessary computation;

[0070] b. Use appropriate height filtering to remove taller point clouds;

[0071] c. Set a reasonable grid size and create a two-dimensional grid matrix on a horizontal plane;

[0072] d. Iterate through each grid cell, setting the grid cell's height to the lowest height of all point clouds within its boundaries. If a grid cell contains no point clouds, mark it as having no point clouds.

[0073] e. Iterate through all grids without point clouds. If not all grids adjacent to a grid are without point clouds, set the weights using a two-dimensional Gaussian distribution and use the weighted average height as the height of the grid.

[0074] f. Repeat step e until all grid cells have a height value;

[0075] g. Traverse all grids and erode the higher grids using adjacent lower-order grids. That is, if the height of an adjacent grid is less than the height of the current grid, the height of the current grid is reset to the lowest height among the adjacent grids.

[0076] The correction of relative paths will be explained in more detail below.

[0077] The heading angle fed back by the positioning module of the unmanned intelligent torpedo tanker has a certain error, resulting in a large lateral error when calculating the clearance at a relatively long distance, thus causing false alarms and missed alarms for obstacles. This invention addresses this problem by using a combination of a random sampling consensus algorithm and a nonlinear optimization algorithm to eliminate interference from the point clouds of other tracks at switch locations and the point clouds of low vegetation that are misidentified as rails, accurately calculating the angle that the relative path needs to be corrected; and by using a reverse solution to solve for the correction angle, the calculation process is simplified. The processing flow of this method is as follows:

[0078] a. The extracted rail point cloud is distinguished into left rail points and right rail points;

[0079] b. Using a random sampling consensus algorithm, randomly extract some points on the left side rail and some points on the right side rail respectively;

[0080] c. Introduce the correction angle variable of the railway point cloud and set initial values, and construct the rotation matrix.

[0081] d. Apply a rotation matrix to a portion of the extracted left-side rail points and a portion of the right-side rail points to obtain the rotated point cloud.

[0082] e. Construct an error function, which is the sum of the differences between the distances from each rotated point cloud to the relative path equation and half the standard track gauge.

[0083] f. Apply the Gauss-Newton method to perform nonlinear optimization of the error function to obtain the optimal correction angle for the railway point cloud.

[0084] g. Evaluate the correction angle of this optimal rail point cloud. If it is better than the correction angle obtained in the previous random sampling consistent loop, retain it; otherwise, discard it.

[0085] h. Repeat the random sampling loop until the maximum number of loops is reached and the final correction angle is obtained;

[0086] i. If the correction angle of the railway point cloud is reversed, then it is the angle that the relative path equation needs to be corrected.

[0087] j. Apply the angle to be corrected to the relative path points and refit the relative path equation.

[0088] The following section provides a more detailed explanation of the perception fusion of LiDAR and visual sensors.

[0089] This invention utilizes both lidar and visual sensors to compensate for the shortcomings of each, establishing a perception system adapted to the operating environment of the unmanned intelligent torpedo tanker. By classifying obstacles, it separates objects that pose a safety hazard to the unmanned intelligent torpedo tanker from those that do not, improving the accuracy of obstacle perception and reducing false alarm and false positive rates. The processing flow of this method is as follows:

[0090] a. Using a deep learning model of 3D point cloud, semantic analysis is performed on the input point cloud to annotate the point clouds of trees and vegetation. Trees and weeds often intrude into the operating limits of the unmanned intelligent torpedo tanker's track and on both sides of the track, but this does not pose a safety hazard; therefore, the unmanned intelligent torpedo tanker should not take special measures such as braking in response.

[0091] b. Based on the rasterized ground, use an elevation threshold to extract and cluster the point cloud above the chassis of the unmanned intelligent torpedo tanker truck;

[0092] c. Based on the size and location of the clustered point cloud, the corresponding range of the point cloud in the image captured by the visual sensor is calculated using the basic pinhole model, and the corresponding relationship is established.

[0093] d. Using a deep learning model for images, perform category identification on the two-dimensional images corresponding to the clustered point clouds in sequence to determine whether they are objects such as pedestrians and vehicles that pose a threat to driving, or objects such as tree branches and weeds that do not pose a threat to driving.

[0094] e. Send obstacle information.

[0095] Those skilled in the art will understand that, in addition to implementing the system, apparatus, and their modules provided by this invention in purely computer-readable program code, the same program can be implemented in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers by logically programming the method steps. Therefore, the system, apparatus, and their modules provided by this invention can be considered a hardware component, and the modules included therein for implementing various programs can also be considered structures within the hardware component; alternatively, modules for implementing various functions can be considered both software programs implementing the method and structures within the hardware component.

[0096] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. An environmental perception system for an unmanned intelligent torpedo tanker truck, characterized in that, include: Communication system: Acquires point cloud information and image information of the environment surrounding the torpedo tanker; Obstacle recognition system: It performs environmental perception by analyzing point cloud information from lidar and image information from visual sensors, and identifies obstacles that intrude into the current torpedo tanker's operating clearance. Display system: Displays point cloud information, image information, operating limit area, and obstacle recognition results in real time; In the obstacle recognition system: The ground is fitted based on point cloud information and image information, and the railway track point cloud is extracted. Operational clearance calculations are performed using the torpedo car's pose, trajectory, and rail point cloud. Based on point cloud information and image information, obstacles are identified and classified, the distance and size of obstacles are calculated, and it is determined whether the obstacles are within the operating limits. In the obstacle recognition system, the uneven ground is fitted using a rasterization method based on point cloud information to obtain a rasterized ground. Based on the rasterized ground, the railway point cloud is extracted, and combined with the current pose and running path of the torpedo tanker, the deviation of the running path is corrected, and the running limit of the unmanned intelligent torpedo tanker is calculated. The corrective operation path includes: The extracted rail point cloud is distinguished and divided into left rail points and right rail points; Randomly extract some points from the left side of the railway track and some points from the right side of the railway track; The correction angle variable of the railway point cloud is introduced and initial values ​​are set, and a rotation matrix is ​​constructed; A rotation matrix is ​​applied to a portion of the extracted left-side rail points and a portion of the right-side rail points to obtain a rotated point cloud. Construct an error function, which is the sum of the differences between the distances from each point cloud after rotation to the relative path equation and half the standard track gauge; The error function is nonlinearly optimized to obtain the optimal correction angle for the rail point cloud; Evaluate the correction angle of this optimal rail point cloud. If it is better than the initial value of the correction angle variable, retain the correction angle of the optimal rail point cloud; otherwise, discard it. Continue until the maximum number of iterations is reached and the final correction angle is obtained. The angle of correction of the rail point cloud is the inverse of the angle that the relative path equation needs to be corrected. Apply the angle to be corrected to the relative path points and refit the relative path equation.

2. The unmanned intelligent torpedo tanker environmental perception system according to claim 1, characterized in that, The determination of whether an obstacle is within the operating limits includes: Semantic analysis is performed on the original input 3D point cloud information to label whether each point belongs to shrub vegetation; Based on the current pose of the torpedo tanker, the mission path is transformed into the relative coordinate system of the torpedo tanker to obtain the relative mission path. Extract the waypoints of the relative task path within the required detection range, and fit the equation expression of the relative task path. Based on the equation expression of the relative task path and the lateral threshold, the region of interest is defined and the 3D point cloud within the region is extracted; An uneven ground surface is represented using a rasterization method, and the elevation within each ground grid is determined. Based on the elevation difference between the three-dimensional point cloud and its corresponding ground grid, point clouds that may be railway tracks are extracted. The point cloud that may be a railway track is filtered based on the lateral distance difference to the relative task path. Based on the filtered rail point cloud and the relative task path, an error function is established and nonlinear optimization is performed to maximize the match between the relative task path and the centerline of the rail point cloud, thereby eliminating the angular error of the relative task path. Calculate the runtime limits based on the corrected relative task path; Perform height filtering based on rasterized ground on the 3D point cloud within the region of interest; Point clouds that have been filtered and marked as shrub vegetation are removed. The remaining point cloud is segmented and clustered; Based on the distance and size of the clustered point cloud, the corresponding region of the point cloud within the image captured by the visual sensor is inferred; Pixels within the visual sensor image are selected, and deep learning is used to identify the types of pixels one by one, and the types are matched with the clustered point cloud. If an obstacle point cloud is identified through the corresponding identification, a decision to avoid it is made based on the distance, size, and type of the obstacle point cloud.

3. The unmanned intelligent torpedo tanker environmental perception system according to claim 2, characterized in that, The use of a rasterization method to represent uneven ground includes: Select the detection area based on the relative path; Use height filtering to remove point clouds with higher threshold values; Set the grid size and create a two-dimensional grid matrix in a horizontal plane; Iterate through each grid cell sequentially, and use the lowest height of all point clouds within the grid cell as the height of this grid cell; if there is no point cloud in the grid cell, mark it as no point cloud; Iterate through all grid cells without point clouds. If not all adjacent grid cells are without point clouds, set the weights using a two-dimensional Gaussian distribution and use the weighted average height as the height of the grid cell. Continue until all grid cells have a height value. Traverse all grids and erode higher grids using adjacent lower grids. That is, if the height of an adjacent grid is less than the height of the current grid, the height of the current grid is reset to the lowest height among the adjacent grids.

4. The unmanned intelligent torpedo tanker environmental perception system according to claim 1, characterized in that, The obstacle recognition system fuses and complements information from LiDAR and visual sensors: Using a deep learning model of 3D point cloud, semantic analysis is performed on the input point cloud to label the point clouds of trees and vegetation without taking braking measures on the trees and vegetation. Based on the rasterized ground, an elevation threshold is used to extract and cluster point clouds above the torpedo tanker chassis; Based on the size and location of the clustered point cloud, the corresponding range of the clustered point cloud in the image captured by the visual sensor is calculated, and the corresponding relationship is established. Using a deep learning model for images, the two-dimensional images corresponding to the clustered point clouds are sequentially classified to determine whether they are pedestrians and vehicles that pose a threat to traffic, or trees and vegetation that do not pose a threat to traffic.

5. The unmanned intelligent torpedo tanker environmental perception system according to claim 1, characterized in that, The communication system communicates with the onboard PLC control system of the torpedo tanker to obtain the torpedo tanker's driving route information, including obtaining the coordinate point set of the current track route of the torpedo tanker and transmitting the location, size and type information of the identified obstacles.

6. The unmanned intelligent torpedo tanker environmental perception system according to claim 1, characterized in that, The communication system communicates with the positioning module to obtain the current pose information, current position and heading information of the torpedo tanker, as well as to send the identified obstacle information.

7. The unmanned intelligent torpedo tanker environmental perception system according to claim 1, characterized in that, When an obstacle is detected that has intruded into the current operating clearance of the torpedo tanker, the torpedo tanker is controlled to brake.

Images