A method for target detection and recognition on structured roads for intelligent driving vehicles

Abstract

The present invention discloses a method for target detection and recognition on a structured road of an intelligent driving vehicle. The method mainly includes a data acquisition module, a coordinate conversion module, a point cloud data preprocessing module, a ground point cloud segmentation module, a target object segmentation module, and a point cloud classification and recognition module. The three-dimensional point cloud data is derived from a vehicle-mounted laser radar, and the target object is detected and recognized while the vehicle is driving, which can provide the vehicle with more detailed information on obstacles such as vehicles and pedestrians. The present invention uses a vehicle-mounted laser radar to collect three-dimensional point cloud data, and uses the advantages of high laser radar data accuracy and few constraints to solve the problem that the millimeter wave radar detection accuracy is not enough and the surrounding three-dimensional environment cannot be described in detail. The point cloud segmentation algorithm is designed and optimized in combination with a point cloud classification and recognition algorithm based on deep learning, and the optimization effect is obvious, which can ensure the rapid realization of target detection and recognition tasks on structured roads.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent driving, and specifically relates to a method for realizing road surface target detection and recognition based on three-dimensional point cloud data for an intelligent driving vehicle applied on a structured road. Background Art

[0002] With the rapid development of smart transportation and artificial intelligence technology, intelligent driving vehicles not only improve road traffic safety and alleviate parking problems, but also greatly reduce the huge loss of life and property caused by traffic accidents due to improper human operation.

[0003] With the emergence of new energy vehicle companies such as Tesla, intelligent driving technology is receiving widespread attention from all walks of life and is moving towards higher levels of autonomous driving. It is also considered to be a very promising development direction for future transportation systems.

[0004] The intelligent driving system is composed of three parts: perception, decision-making and control. It is a comprehensive intelligent control system. The perception module consists of environmental perception, navigation positioning and vehicle status perception. Environmental perception is like the vision and hearing of the car, which can help the car quickly and accurately obtain driving environment information. It is the prerequisite and foundation for the vehicle to achieve advanced intelligent behaviors such as obstacle avoidance, positioning, and path planning.

[0005] At this stage, in order to improve the capabilities of the car's perception system and increase the level of intelligence of the car, many sensors will be installed in the car, such as lidar, millimeter-wave radar, camera, positioning and navigation systems and other sensors. Three-dimensional lidar has the advantages of high ranging accuracy, large detection range and no influence from lighting conditions. The laser rotates horizontally at high speed to obtain the surrounding environmental information, thereby generating a series of data points, called point cloud data, which can provide a detailed description of the surrounding three-dimensional environment. It is considered by most companies to be an indispensable part of the development of higher-level autonomous driving. Based on the three-dimensional point cloud data collected by lidar, the rapid detection and recognition of objects is a crucial link in environmental perception. Summary of the invention

[0006] The present invention adopts a vehicle-mounted laser radar to collect three-dimensional point cloud data. It uses the advantages of high accuracy and few constraints of laser radar data to solve the problem that millimeter-wave radar detection accuracy is insufficient and it is unable to provide a detailed description of the surrounding three-dimensional environment. It also designs and optimizes the point cloud segmentation algorithm combined with the point cloud classification algorithm based on deep learning. The optimization effect is obvious, which can ensure the rapid realization of target detection and recognition tasks on structured roads.

[0007] In order to solve the technical problems in various links, the present invention designs the following technical solutions:

[0008] In order to achieve rapid detection and recognition of target objects on structured roads, this method is implemented through the following modules:

[0009] Data acquisition module, coordinate conversion module, point cloud data preprocessing module, ground point cloud segmentation module, target object segmentation module, point cloud classification and recognition module.

[0010] This method builds a data processing and visualization software system based on the ROS platform, using Ubuntu as the operating system and C++ and Python as programming languages. It reads the three-dimensional point cloud data collected by the lidar, implements data processing in each module, and finally displays it on the RVIZ visualization platform.

[0011] The data acquisition module is responsible for collecting 3D point cloud data, mainly through the mechanical laser radar installed on the top of the car. The laser radar can be of any brand with 32 or 64 lines.

[0012] The coordinate conversion module is responsible for converting the 3D point cloud data between the laser radar coordinate system and the vehicle coordinate system. The 3D coordinates of the point cloud in the laser radar coordinate system are obtained by parsing the original UDP data packet. The conversion is performed based on the rotation matrix R and the translation matrix T to obtain the 3D coordinates of the point cloud in the vehicle coordinate system. The specific steps include the following:

[0013] Step 1.1, parse the original UDP data packet to obtain the echo distance Range, vertical angle V_theta, horizontal angle H_Beta, turntable angle ωt and other parameters.

[0014] Step 1.2: Calculate the three-dimensional coordinates (x l ,y l ,z l ):

[0015] Angle=-ωt-ω*Δt-(H_Beta) (4-1)

[0016]

[0017] Step 1.3, the rotation matrix R is calculated from the angle vector A = [ε1 ε2 ε3] actually measured when installing the laser radar:

[0018] R=R(ε1)R(ε2)R(ε3) (4-3)

[0019] in:

[0020]

[0021]

[0022]

[0023] Combined with the translation matrix T by formula (4-3), the vehicle coordinate system o v x v y v z v and the coordinate system o of the lidar l x l y l z l conversion formula:

[0024]

[0025] Thus, the three-dimensional coordinates (x v , y v , z v ) of the point cloud in the vehicle coordinate system are obtained.

[0026] The point cloud data preprocessing module is responsible for preprocessing the three-dimensional point cloud in the vehicle coordinate system. Using the PCL open-source library, pass-through filtering and statistical filtering are selected to effectively filter outlier points and point cloud data that is too far from the driving vehicle in the point cloud data.

[0027] The ground point cloud segmentation module is responsible for segmenting the ground point cloud in the preprocessed three-dimensional point cloud data. To achieve the rapid separation of the ground point cloud and the non-ground point cloud, a multi-iteration ground fitting method based on multi-sector projection is designed. The specific principle steps are as follows:

[0028] Step 2.1: Divide the single-frame point cloud data into multiple equal-angle fan-shaped regions on the X-Y plane at an angle θ. The first sector above the X-axis in the first quadrant is numbered as sector 0, and the numbering increases counterclockwise. Calculate which sector each point belongs to according to its projection coordinates (x, y) on the X-Y plane. θ can be set according to the actual situation. In this method, θ = 3°, and a total of 120 sectors are divided.

[0029] Step 2.2: Sort the points in each sector according to the Z-axis size starting from Z0. Given the height h of the lidar relative to the ground, obtain the value of Z0 - (-h). If it is greater than the threshold M, it is determined that the point is off the ground, and the entire sector is excluded. Otherwise, continue to judge D = Z1 - Z0. If D < N and Z1 - (-h) < M, continue to consider it; otherwise, no longer consider the subsequent points. In this way, the fitting point P i (f i = 0) is obtained. The thresholds M and N can be set according to the actual situation. In this method, the threshold is set to 0.2m.

[0030] Step 2.3: Use the least squares method according to the fitting point P i (fi =0) to fit the plane equation G(x,y,z)=EstimatePlane[P i (f i =0)].

[0031] Step 2.4: Traverse the point cloud P i (f i =1), calculate P i (f i =1) and the vertical projection distance between the plane G(x,y,z) i ; Set the distance threshold ε dist , determine distance i ≤ε dist Is it true? If true, point P i The logo of i Set to 0 and return to the ground point cloud; otherwise, remain unchanged. Distance threshold ε dist It can be set according to the actual situation. In this method, the threshold ε is set dist is 0.2m.

[0032] Step 2.5: Set the number of iterations N iter Repeat steps 2.3 to 2.4 to implement plane fitting iteration. i Points that are 0 are all ground point clouds. After deleting them, the goal of segmenting the ground point cloud in the point cloud space can be achieved. Iteration number N iter You can set it according to the actual situation. In this method, the number of iterations is set to 5.

[0033] The target object segmentation module is responsible for segmenting the non-ground point cloud into different objects. It uses an optimized point cloud segmentation algorithm based on Euclidean distance. The specific principles and steps are as follows:

[0034] Step 3.1, load the point cloud data P and store it sequentially, establish the corresponding kd tree, and establish the pointer index list L of the element to be processed and the clustering result list C.

[0035] Step 3.2: Create a new current cluster pointer index list C k , extract a "measurement point" from L as the seed "measurement point", delete it from L and add C k ; Using the seed "measurement point" as the center, use the kd tree to perform a fast neighborhood search to find all neighboring "measurement points", and delete the neighboring "measurement points" from L and add them to C k In the process of recursive growth, the new neighboring "measurement point" is continuously deleted from L and added to C. k , until no new neighboring "measurement points" are added to Ck So far, the current cluster C k All the "measurement points" belonging to it have been found, and C k Join to C.

[0036] Step 3.3: Repeat step 3.2 until L is an empty list. At this time, C is a single target object cluster set, and the non-ground point cloud segmentation is completed.

[0037] The optimization contents of this method for the point cloud segmentation algorithm are as follows:

[0038] For any measuring point P i Set a dynamic search radius to avoid the problem of the field search radius being too large or too small. The dynamic search radius formula is as follows:

[0039]

[0040] The value of coefficient k must satisfy Where θ' is the horizontal angle resolution of the laser radar. The coefficient k can be set according to the actual situation when the conditions are met. This method selects the coefficient k = 0.02.

[0041] When clustering, for a single cluster C in the cluster set C k Screening is performed to eliminate extremely large or extremely small clusters, and only the target object clusters of primary concern are retained. The cluster constraint range can be set according to actual conditions. This method selects the range [100, 15000].

[0042] The point cloud classification and recognition module is responsible for identifying the different segmented objects. It uses a point cloud classification algorithm based on the PointNet++ neural network, including network training and online target recognition. The specific contents are as follows:

[0043] In the network training part, a sample library is built based on the KITTI public dataset, which is widely used in the field of 3D point cloud target classification research. When training the PointNet++ classification network, a multi-scale MSG model with good robustness is selected, the activation function is Relu, the learning method is stochastic gradient descent (SGD), the initial learning rate is 0.001, the batch size is 24, and the epoch is 200.

[0044] In the online target recognition part, since the objects that a vehicle can encounter on a structured road are mainly vehicles and pedestrians, other obstacles do not need to be classified in detail. Therefore, for the objects segmented from a frame of point cloud data, three categories are classified into vehicles, pedestrians and other obstacles, and the recognition process is optimized by setting preliminary screening. The specific steps are as follows:

[0045] Step 4.1: Preliminary screening of targets: select candidate targets based on whether the length, width and height of the target object meet the geometric feature rules of pedestrians and vehicles. c ≤2m,w c ≤2.1m,l c If the object is smaller than 6m, it is considered as a candidate for identification, and those that do not meet the requirement are classified as other obstacles.

[0046] Among them, h c Indicates the height of the target object, w c Indicates the width of the target object, l c Indicates the length of the target object.

[0047] Step 4.2: For the candidate recognition targets, put them into the trained PointNet++ neural network to extract features, and use the fully connected layer to get the final result. After completing the classification of the targets, the output results are combined into a target list, and each target object is finally divided into pedestrians, vehicles and other obstacles.

[0048] Beneficial effects of the present invention: The target detection and recognition method for intelligent driving vehicles on structured roads provided by the present invention detects and recognizes target objects while the vehicle is driving, and can provide the vehicle with more detailed information about obstacles such as vehicles and pedestrians. The present invention uses a vehicle-mounted laser radar to collect three-dimensional point cloud data, and uses the advantages of high laser radar data accuracy and few constraints to solve the problem that the millimeter wave radar detection accuracy is insufficient and the surrounding three-dimensional environment cannot be described in detail. The point cloud segmentation algorithm is designed and optimized in combination with a point cloud classification and recognition algorithm based on deep learning, and the optimization effect is obvious, which can ensure the rapid realization of target detection and recognition tasks on structured roads. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is the overall module diagram

[0050] Figure 2 This is a schematic diagram of the laser radar installation location

[0051] Figure 3 It is the coordinate transformation flow chart

[0052] Figure 4 It is the ground point cloud segmentation flow chart

[0053] Figure 5 This is the target object segmentation flow chart

[0054] Figure 6 This is the point cloud classification and recognition module diagram DETAILED DESCRIPTION

[0055] The following is a detailed description of the embodiments of the present invention in conjunction with the accompanying drawings in the specification. The present invention uses a vehicle-mounted laser radar to collect three-dimensional point cloud data, and utilizes the advantages of high accuracy and few constraints of laser radar data to solve the problem that the millimeter-wave radar detection accuracy is insufficient and the surrounding three-dimensional environment cannot be described in detail. The point cloud segmentation algorithm is designed and optimized in combination with a point cloud classification and recognition algorithm based on deep learning. The optimization effect is obvious, which can ensure the realization of target detection and recognition tasks on structured roads.

[0056] This embodiment provides a method for detecting and identifying targets on a structured road for an intelligent driving vehicle. Figure 1 As shown, the method includes a data acquisition module 10 , a coordinate conversion module 20 , a point cloud data preprocessing module 30 , a ground point cloud segmentation module 40 , a target object segmentation module 50 and a point cloud classification and recognition module 60 .

[0057] The method provided in this embodiment builds a data processing and visualization software system based on the ROS platform, using Ubuntu as the operating system and C++ and Python as the programming languages. The three-dimensional point cloud data collected by the laser radar is read, data processing is implemented in each module, and finally displayed on the RVIZ visualization platform.

[0058] In this embodiment, the data acquisition module 10 uses a mechanical laser radar to collect three-dimensional point cloud data. Figure 2 As shown, the laser radar 11 is installed on the top of the car. The brand of the laser radar is Velodyne, and the number of lines is 64.

[0059] The coordinate conversion module 20 in this embodiment is responsible for converting the three-dimensional point cloud data between the laser radar coordinate system and the vehicle coordinate system. Figure 3 As shown, it includes steps 21, 22, and 23, which are specifically as follows:

[0060] 21. Parse the original UDP data packet to obtain parameters such as echo distance Range, vertical angle V_theta, horizontal angle H_Beta, turntable angle ωt, etc.

[0061] 22. Calculate the three-dimensional coordinates (x l ,y l ,z l ):

[0062] Angle=-ωt-ω*Δt-(H_Beta) (1-1)

[0063]

[0064] 23. The rotation matrix R is calculated from the angle vector A = [ε1 ε2 ε3] actually measured when the laser radar is installed:

[0065] R=R(ε1)R(ε2)R(ε3) (1-3)

[0066] in:

[0067]

[0068]

[0069]

[0070] Combining formula (1-3) with the translation matrix T, we get the vehicle coordinate system o v x v y v z v With the laser radar coordinate system o l x l y l z l The conversion formula is:

[0071]

[0072] The three-dimensional coordinates (x v ,y v ,z v ).

[0073] The point cloud data preprocessing module 30 of this embodiment is responsible for implementing point cloud data filtering, using the PCL open source library, selecting through filtering and statistical filtering, and effectively filtering outliers in the point cloud data and point cloud data that is too far away from the moving vehicle.

[0074] The ground point cloud segmentation module 40 of this embodiment is responsible for realizing the rapid separation of ground point cloud and non-ground point cloud. In this embodiment, a multi-iteration ground fitting method based on multi-sector projection is designed, such as Figure 4 As shown, it includes steps 41, 42, 43, 44, and 45, which are specifically as follows:

[0075] 41. Divide the single-frame point cloud data into multiple sectors with equal angles at an angle of θ on the XY plane. The first sector on the X-axis of the first quadrant is numbered 0, and the numbering increases counterclockwise. Calculate which sector the point belongs to based on the projection coordinates (x, y) of each point on the XY plane. θ can be set according to the actual situation. In this embodiment, θ=3° is set, and a total of 120 sectors are divided.

[0076] 42. Sort the points in each sector according to the Z-axis size starting from Z0. Obtain the value of Z0 - (-h) from the height h of the lidar relative to the ground. If it is greater than the threshold M, determine that the point is off the ground and exclude the entire sector. Otherwise, continue to judge D = Z1 - Z0. If D < N and Z1 - (-h) < M, continue to consider it; otherwise, no longer consider the subsequent points. In this way, the fitting point P is obtained. i (f i = 0). The thresholds M and N can be set according to the actual situation. In this embodiment, the threshold is set to 0.2m.

[0077] 43. Use the least squares method to fit according to the fitting point P i (f i = 0) to obtain the plane equation G(x, y, z) = EstimatePlane[P i (f i = 0)].

[0078] 44. Traverse the point cloud P i (f i = 1), calculate the perpendicular projection distance distance between P i (f i = 1) and the plane G(x, y, z); set the distance threshold ε i , and judge whether distance dist ≤ ε i holds. If it holds, set the flag f dist of the point P i to 0 and classify it into the ground point cloud; otherwise, keep it unchanged. The distance threshold ε i can be set according to the actual situation. In this embodiment, the threshold ε dist is set to 0.2m. dist is 0.2m.

[0079] 45. Set the number of iterations N iter , and repeat steps 43 to 44 to implement plane fitting iteration. All points with the flag f i being 0 are ground point clouds. After deletion, the goal of separating the ground point cloud in the point cloud space can be achieved. The number of iterations N iter can be set according to the actual situation. In this embodiment, the number of iterations is set to 5.

[0080] In this embodiment, the target object segmentation module 50 is responsible for implementing the segmentation of non-ground point clouds into different objects. In this embodiment, an optimized point cloud segmentation algorithm based on Euclidean distance is used, as shown in Figure 5 , which includes steps 51, 52, 53, and is specifically as follows:

[0081] 51. Load the point cloud data P and store it sequentially, establish the corresponding kd tree, and establish the pointer index list L of the element to be processed and the clustering result list C.

[0082] 52. Create a new current cluster pointer index list C k , extract a "measurement point" from L as the seed "measurement point", delete it from L and add C k ; Using the seed "measurement point" as the center, use the kd tree to perform a fast neighborhood search to find all neighboring "measurement points", and delete the neighboring "measurement points" from L and add them to C k In the process of recursive growth, the new neighboring "measurement point" is continuously deleted from L and added to C. k , until no new neighboring "measurement points" are added to C k So far, the current cluster C k All the "measurement points" belonging to it have been found, and C k Join to C.

[0083] 53. Repeat step 52 until L is an empty list, at which point C is a single target object cluster set, and the non-ground point cloud segmentation is completed.

[0084] In this embodiment, the optimization contents for the point cloud segmentation algorithm are as follows:

[0085] For any measuring point P i Set a dynamic search radius to avoid the problem of the field search radius being too large or too small. The dynamic search radius formula is as follows:

[0086]

[0087] The value of coefficient k must satisfy Where θ′ is the horizontal angle resolution of the laser radar. The coefficient k can be set according to the actual situation when the conditions are met. In this embodiment, the coefficient k=0.02 is selected.

[0088] When clustering, for a single cluster C in the cluster set C k Screening is performed to eliminate extremely large or extremely small clusters, and only the target object clusters of primary concern are retained. The cluster constraint range can be set according to actual conditions. In this embodiment, the selected range is [100, 15000].

[0089] The point cloud classification and recognition module 60 of this embodiment is responsible for identifying the different segmented objects. In this embodiment, a point cloud classification algorithm based on the PointNet++ neural network is used, such as Figure 6 The diagram includes two parts: network training 61 and online target recognition 62, which are as follows:

[0090] In the network training 61 part of this embodiment, a sample library is constructed based on the KITTI public dataset widely used in the field of three-dimensional point cloud target classification research. When training the PointNet++ classification network, a multi-scale MSG model with good robustness is selected, the activation function is Relu, the learning method is stochastic gradient descent (SGD), the initialization learning rate is 0.001, the Batch Size is 24, and the Epoch is 200.

[0091] In the online target recognition 62 part of this embodiment, since the objects that a vehicle can encounter on a structured road are mainly vehicles and pedestrians, other obstacles do not need to be classified in detail. Therefore, for the objects segmented from a frame of point cloud data, three categories are classified into vehicles, pedestrians and other obstacles, and the recognition process is optimized by setting preliminary screening, such as Figure 6 As shown, it includes steps 621 and 622, which are specifically as follows:

[0092] 621. Preliminary screening of targets: select candidate targets for identification based on whether the length, width and height of the target object meet the geometric feature rules of pedestrians and vehicles. If the clustered target object meets 0.3m≤h c ≤2m,w c ≤2.1m,l c If the object is smaller than 6m, it is considered as a candidate for identification, and those that do not meet the requirement are classified as other obstacles.

[0093] Among them, h c Indicates the height of the target object, w c Indicates the width of the target object, l c Indicates the length of the target object.

[0094] 622. For the candidate recognition targets, put them into the trained PointNet++ neural network to extract features, and use the fully connected layer to get the final result. After completing the classification of the targets, the output results are combined into a target list, and each target object is finally divided into pedestrians, vehicles and other obstacles.

[0095] In summary, the target detection and recognition method for intelligent driving vehicles on structured roads provided by the present invention can detect and recognize target objects while the vehicle is driving, and can provide the vehicle with more detailed information about obstacles such as vehicles and pedestrians. The present invention uses a vehicle-mounted laser radar to collect three-dimensional point cloud data, and uses the advantages of high accuracy and few constraints of laser radar data to solve the problem that the millimeter wave radar detection accuracy is insufficient and the surrounding three-dimensional environment cannot be described in detail. The point cloud segmentation algorithm is designed and optimized in combination with a point cloud classification and recognition algorithm based on deep learning, and the optimization effect is obvious, which can ensure the rapid realization of target detection and recognition tasks on structured roads.

[0096] The above embodiments are merely illustrative of the principles and effects of the present invention, and are not intended to limit the present invention. Anyone familiar with the art may modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by a person of ordinary skill in the art without departing from the spirit and technical concept disclosed by the present invention shall still be covered by the claims of the present invention.

Claims

1. A method for detecting and identifying targets on a structured road for an intelligent driving vehicle, characterized in that: The target detection and recognition method for structured roads of intelligent driving vehicles is implemented through the following modules: Data acquisition module, coordinate conversion module, point cloud data preprocessing module, ground point cloud segmentation module, target object segmentation module, point cloud classification and recognition module; The data collected by the data acquisition module is three-dimensional point cloud data, which comes from a mechanical laser radar installed on the top of the car, and the number of laser lines is 32 or 64; The coordinate conversion module converts between the laser radar coordinate system and the vehicle coordinate system based on the rotation matrix R and the translation matrix T; The point cloud data preprocessing module uses the PCL open source library to select through filtering and statistical filtering to filter outliers in the point cloud data and point cloud data that is too far away from the moving vehicle; the ground point cloud segmentation module is used to achieve rapid separation of ground point clouds and non-ground point clouds; The target object segmentation module uses the target detection and recognition method: builds a data processing and visualization software system based on the ROS platform, uses Ubuntu as the operating system, and uses C++ and Python as programming languages; reads the three-dimensional point cloud data collected by the laser radar, implements data processing in each module, and finally displays it on the RVIZ visualization platform; The point cloud classification and recognition module uses a point cloud classification algorithm based on the PointNet++ neural network to perform point cloud classification and recognition; The ground point cloud segmentation module designs a multi-iteration ground fitting method based on multi-sector projection to achieve rapid separation of ground point clouds and non-ground point clouds, and is performed in the following steps: Step 2.1, divide the single-frame point cloud data into multiple sectors with equal angles at an angle θ on the XY plane, with the first sector on the X-axis of the first quadrant as sector number 0, and the numbering increases counterclockwise. Calculate which sector the point belongs to based on the projection coordinates (x, y) of each point on the XY plane; θ can be set according to the actual situation; Step 2.2: Sort the points in each sector according to the Z-axis size starting from Z0. From the height h of the lidar relative to the ground, obtain the value of Z0 - (-h). If it is greater than the threshold M, it is determined that the point is off the ground and the entire sector is excluded; otherwise, continue to judge D = Z1 - Z0. If D < N and Z1 - (-h) < M, continue to consider; otherwise, no longer consider the subsequent points, and thus obtain the fitting point P i (f i = 0); The thresholds M and N can be set according to the actual situation Step 2.3: Use the least squares method to fit the point P i (f i =0) to fit the plane equation G(x,y,z)=EstimatePlane[P i (f i =0)]; Step 2.4: Traverse the point cloud P i (f i =1), calculate P i (f i =1) and the vertical projection distance between the plane G(x,y,z) i ; Set the distance threshold ε dist , determine distance i ≤ε dist Is it true? If true, point P i The logo of i Set to 0 and include it in the ground point cloud; otherwise, keep it unchanged; distance threshold ε dist It can be set according to the actual situation; Step 2.5: Set the number of iterations N iter Repeat steps 2.3 to 2.4 to implement plane fitting iteration; all the labels f i Points that are 0 are all ground point clouds. After deleting them, the goal of segmenting the ground point clouds in the point cloud space can be achieved; the number of iterations N iter It can be set according to the actual situation; The target object segmentation module uses an optimized point cloud segmentation algorithm based on Euclidean distance and proceeds according to the following steps: Step 3.1, load the point cloud data P and store it sequentially, establish the corresponding kd tree, establish the pointer index list L of the element to be processed and the clustering result list C; Step 3.2: Create a new current cluster pointer index list C k , extract a "measurement point" from L as the seed "measurement point", delete it from L and add it to C k ; Using the seed "measurement point" as the center, use the kd tree to perform a fast neighborhood search to find all neighboring "measurement points", and delete the neighboring "measurement points" from L and add them to C k In the recursive growth process, the new "measurement point" is used as the center for neighborhood search, and the new neighboring "measurement point" is continuously deleted from L and added to C. k , until no new neighboring "measurement points" are added to C k So far, the current cluster C k All the "measurement points" belonging to it have been found, and C k Join to C; Step 3.3, repeat step 3.2 until L is an empty list, at which point C is a single target object cluster set, and the non-ground point cloud segmentation is completed; The optimized point cloud segmentation algorithm based on Euclidean distance includes the following optimization contents: For any measuring point P i Set the adaptive search radius to avoid the problem of the neighborhood search radius being too large or too small. The adaptive search radius formula is as follows: The value of coefficient k must satisfy Where θ′ is the horizontal angle resolution of the laser radar; the coefficient k can be set according to the actual situation when the conditions are met; When clustering, for a single cluster C in the cluster set C k Screening is performed to eliminate extremely large or extremely small clusters, and only the target object clusters of primary concern are retained; the cluster constraint range can be set according to actual conditions; The point cloud classification and recognition module uses a point cloud classification algorithm based on the PointNet++ neural network, which includes two parts: network training and online target recognition: In the network training part, a sample library was built based on the 3D point cloud target classification dataset. When training the PointNet++ classification network, a multi-scale MSG model was selected, the activation function was Relu, the learning method was stochastic gradient descent, the initial learning rate was 0.001, the batch size was 24, and the epoch was 200. In the online target recognition part, for the objects segmented from a frame of point cloud data, three categories are classified into vehicles, pedestrians and other obstacles, and the recognition process is optimized by setting preliminary screening. The specific steps are as follows: Step 4.1: Preliminary screening of targets: selecting candidate targets for identification based on whether the length, width and height of the target object meet the geometric feature rules of pedestrians and vehicles; If the clustering target object satisfies 0.3m≤h c ≤2m,w c ≤2.1m,l c If the distance is ≤6m, it is considered as a candidate for identification, and those that do not meet the requirements are classified as other obstacles; Among them, h c Indicates the height of the target object, w c Indicates the width of the target object, l c Indicates the length of the target object; Step 4.2: For the candidate recognition target, put it into the trained PointNet++ neural network to extract features, and use the fully connected layer to get the final result; After completing the classification of the targets, the output results are organized into a target list, and each target object is finally divided into pedestrians, vehicles, and other obstacles.

2. The method for detecting and identifying targets on structured roads of an intelligent driving vehicle according to claim 1, characterized in that: The coordinate conversion module is performed according to the following steps: Step 1.1, parse the original UDP data packet to obtain the echo distance Range, vertical angle V_theta, horizontal angle H_Beta, turntable angle ωt parameters; Step 1.2: Calculate the three-dimensional coordinates (x l ,y l ,z l ): Angle=-ωt-ω*Δt-(H_Beta) (2-1) Step 1.3, the rotation matrix R is calculated from the angle vector A = [ε1 ε2 ε3] actually measured when installing the laser radar: R=R(ε1)R(ε2)R(ε3) (2-3) in: Combining formula (2-3) with the translation matrix T, we get the vehicle coordinate system o v x v y v z v With the laser radar coordinate system o l x l y l z l The conversion formula is: The three-dimensional coordinates (x v ,y v ,z v ).

Images