A linear array laser imaging target recognition method based on simulated point cloud data set

Abstract

The application discloses a linear array laser imaging target recognition method based on a simulation point cloud dataset, an interactive model of a virtual scanning system and a simulation scene is established, a virtual scanning system motion track is created, a light ray tracing algorithm is used to generate scanning point coordinate data, a virtual scanning system parameter noise model is further established, and noised point cloud data is synthesized; meanwhile, each point in the simulation point cloud data is accurately labeled through the method of establishing a model name-label index table, and finally a simulation point cloud dataset with rich data and accurate labeling is synthesized. Finally, the simulation point cloud dataset is used as training data of a deep learning target recognition algorithm, offline training and online recognition of the algorithm are completed, and accurate laser imaging point cloud target recognition is realized.

Description

Technical Field

[0001] This invention relates to the field of laser imaging and target detection and recognition, specifically to a linear laser imaging target recognition method based on a simulated point cloud dataset. Background Technology

[0002] Laser scanning imaging technology is an active detection technology that uses lasers as an emission source to accurately and rapidly acquire three-dimensional spatial information of targets. Scanning imaging systems based on unmanned aerial vehicle (UAV) platforms are widely used in autonomous navigation, intelligent transportation, environmental monitoring, military target detection, and precision guidance. Unlike passive sensors such as infrared sensors and visible light cameras, array-type laser scanning imaging systems can obtain three-dimensional information of a target scene through push-broom scanning. This three-dimensional data can provide the actual size and shape of the target, and can acquire the target object's position and attitude, making it suitable for refined target detection and precise identification.

[0003] With the development of deep learning technology, significant progress has been made in point cloud target recognition technology based on deep learning. For data-driven supervised deep learning algorithms, the size and quality of the dataset have a significant impact on the algorithm's performance. Accurate labeling of the acquired dataset is essential for its effectiveness as training data. However, due to the high hardware cost of array-type laser imaging systems and the difficulty in changing system parameters once determined, as well as the uncertainty of the actual sensing environment and detection conditions, and the time-consuming and labor-intensive process of manually labeling experimental data, obtaining real-world point cloud data through laser imaging systems is costly and challenging. Using computer simulation to simulate the scanning and imaging process of the imaging system, obtaining simulated point cloud data, and synthesizing a simulated point cloud dataset can effectively solve the problem of obtaining difficult-to-obtain experimental point cloud data. Currently, commonly used laser imaging simulation methods include those based on photon emission, laser spot simulation, and ray tracing. The first two methods are computationally intensive and slow in generating data, while the latter method has higher data generation efficiency, but the simulated data does not include labels, thus often requiring additional data labeling work. Summary of the Invention

[0004] This invention proposes a linear laser imaging target recognition method based on simulated point cloud datasets. It addresses the problems of difficulty in obtaining training data and high cost of constructing point cloud datasets for current deep learning-based point cloud target recognition algorithms. This method provides rich and accurately labeled point cloud datasets for unmanned aerial vehicle target recognition, reduces the cost of data acquisition and labeling in the development of target recognition algorithms, and enables accurate target recognition of linear laser imaging in complex scenarios.

[0005] The technical solution for achieving the present invention is: a method for target recognition in linear laser imaging based on simulated point cloud datasets, characterized by comprising the following steps:

[0006] Step 1: Build a simulation scenario based on the typical working environment of the unmanned aerial vehicle and its target recognition requirements:

[0007] Create 3D terrain and target models for the simulation scene using 3D modeling software. The 3D terrain and target models of the simulation scene should be consistent with the typical working terrain and real targets in the real environment. Add obstacles to the simulation scene according to the target recognition requirements.

[0008] Step 2: Create a model name-label correspondence index table for the simulation scene according to the required labels. Name the terrain, target models and obstacles in the simulation scene. The naming format is "name + serial number". The names of the same type of target are consistent and correspond to the same label. The serial number indicates the number of targets of that type.

[0009] Step 3: Take the geometric center of the target model as the origin O. t The target model moves in the direction of O. t X t Establish a right-handed three-dimensional rectangular coordinate system O along the axis. t X t Y t Z t The coordinate system O is defined at the initial moment. t X t Y t Z t Let O be the ground coordinate system. w X w Y w Z w .

[0010] Step 4: Perform Delaunay triangulation on the simulation scene to obtain a series of triangular facets. Store the triangular facets in an octree structure. Each node of the octree represents a cube voxel. Each node contains 8 child nodes. The sum of the volume elements of the 8 child nodes is equal to the volume of the parent node. Each triangular facet in the simulation scene corresponds to a leaf node.

[0011] Step 5: Add a virtual scanning system to the simulation scene. Set the parameters of the virtual scanning system according to the parameters of the lidar on the UAV, including the field of view angle θ. fov Angular resolution θ s and detection range [t] min ,t max To simulate the dynamic rendezvous process between an unmanned aerial vehicle (UAV) and a real target, a timeline is created. The motion speed v, scanning frequency f, and motion direction of the virtual scanning system are set. The position where the virtual scanning system emits scanning rays each time is calculated, and a keyframe is created at that position. The time start point is set to t0 = 0, and the time when the virtual scanning system emits scanning rays in the i-th time is t_i.i =v i / f. To achieve simulated point cloud generation for arbitrary flight trajectories, the trajectory equation of the virtual scanning system is established, and the ground coordinate system O is calculated at each emitted scanning ray moment during the detection phase of the virtual scanning system. w X w Y w Z w Spatial coordinates (x) i ,y i ,z i Simultaneously calculate the direction of the virtual scanning system's motion velocity (v) at that location. xi ,v yi ,v zi Import spatial coordinates (x) i ,y i ,z i The motion trajectory is constructed by combining the motion velocity and direction at the corresponding position.

[0012] Step 6: For each keyframe created in Step 5, use the spatial coordinates (x, y) of the virtual scanning system corresponding to the keyframe itself. i ,y i ,z i (O) is the origin. s O s X s Along the direction of the motion velocity, O s Y s The axis is perpendicular to the trajectory plane, establishing a virtual scanning system coordinate system O. s X s Y s Z s .

[0013] Step 7: At each keyframe, establish an interaction model with the simulation scene based on the parameters of the virtual scanning system. For a field of view of θ... fov Angular resolution is θ s The virtual scanning system has a corresponding number of emitted scanning rays, n = θ. fov / θ s The simulation point generation is modeled as a problem of finding the intersection of the scanning ray and the triangular elements of the simulation scene, thus obtaining the coordinates of the intersection point between the scanning ray and the simulation scene.

[0014] Step 8: Improve the realism of the point cloud by adding Gaussian white noise to the simulated point cloud data, that is, add distance error to the ideal distance value to generate simulated point cloud data containing Gaussian white noise.

[0015] Step 9: Automatically label the intersection points of the scanning ray and the triangular facet according to the model name-label correspondence index table of the simulation scene. If the scanning ray has a return intersection point, search for the model name corresponding to the triangular facet at that point, assign the label corresponding to the model name at that point, and complete the accurate labeling of each intersection point.

[0016] Step 10: In the virtual scanning system coordinate system O s X s Y s Z s The following is generated with O s Multiple scanning rays emitted from the emission point into the field of view are used to obtain a contour line of the corresponding keyframe simulation scene. Coordinate transformation is then used to transform the coordinates of the contour line points to the ground coordinate system O. w X w Y w Z w Down.

[0017] Step 11: Segment the outlines corresponding to all keyframes to obtain the point cloud data of the simulation scene.

[0018] Step 12: Repeat steps 5 to 11 to create different motion trajectories for virtual scanning systems with different scanning field of view and different angular resolutions. Simulate the motion at different flight speeds, yaw angles, pitch angles, and roll angles to generate N frames of simulated point cloud data under different rendezvous conditions, thus forming a simulated point cloud dataset.

[0019] Step 13: Divide the simulated point cloud dataset into training, validation, and test sets in a 6:2:2 ratio, and input the dataset into the RandLA-Net model to complete the pre-training of the RandLA-Net model.

[0020] Step 14: Input the point cloud data to be identified into the trained RandLA-Net model, and output the category prediction result of each point in the point cloud data to be identified, thereby realizing the offline training and online recognition of the target recognition algorithm.

[0021] Compared with the prior art, the significant advantages of this invention are:

[0022] (1) This invention establishes an interactive model between a virtual scanning system and a simulation scene, triangularizes the simulation scene, and then models the process of the real scanning system collecting point clouds as the process of intersecting the beam ray with the scene triangular elements, thereby realizing the simulation generation of point cloud data. Furthermore, the octree coding structure is used to encode the triangular elements, which greatly improves the simulation efficiency.

[0023] (2) This invention constructs the motion trajectory of the scanning system by setting keyframes. According to the trajectory creation method of this invention, a motion trajectory highly similar to that of the actual scanning system during operation can be established, thereby obtaining point cloud simulation data that is more consistent with reality. In addition, by establishing a relationship model between scanning system parameters and measurement noise, noisy point cloud data can be generated, further improving the realism of the simulation data.

[0024] (3) This invention establishes a name-label index table for various targets in a simulation scene, enabling the return of the three-dimensional coordinates of a simulation point while simultaneously returning the corresponding label, thus completing the automatic annotation of the point cloud dataset. This overcomes the limitation of difficulty in conducting experiments through a real scanning system, and eliminates the need for laborious and time-consuming manual annotation. Accurate annotated point cloud datasets can be obtained through simulation, and the target recognition algorithm can be trained using the simulated point cloud dataset, achieving offline training and online recognition of the target recognition algorithm. Attached Figure Description

[0025] Figure 1 This is a flowchart of the linear laser imaging target recognition method based on a simulated point cloud dataset according to the present invention.

[0026] Figure 2 This is the simulation scene diagram established in step 1 of the present invention.

[0027] Figure 3 This is the mesh generation result of the simulation scene in step 3 of the present invention.

[0028] Figure 4 This is the motion trajectory of the scanning system constructed in the simulation scene in step 4 of the present invention.

[0029] Figure 5 This is the interaction model between the scanning light and the simulation scene of the virtual scanning system in step 6 of the present invention.

[0030] Figure 6 This is a flowchart of the interaction model calculation between the scanning light and the simulation scene of the virtual scanning system in step 6 of the present invention.

[0031] Figure 7 These are the two scan lines obtained in step 10 of the embodiment.

[0032] Figure 8 This is a frame of point cloud data synthesized in step 11 of the embodiment. Detailed Implementation

[0033] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0034] Furthermore, the technical solutions of the various embodiments of the present invention can be combined with each other, but only if they are feasible to those skilled in the art. If the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0035] The following section will further introduce the specific implementation method, as well as the technical difficulties and inventive points of this invention, using this design example as an example.

[0036] Combination Figures 1-6 A target recognition method for linear laser imaging based on simulated point cloud datasets includes the following steps:

[0037] Step 1: Build a simulation scene based on the typical working environment of the unmanned aerial vehicle and its target recognition requirements: Create a 3D terrain and 3D target model of the simulation scene using 3D modeling software. The 3D terrain and 3D target model of the simulation scene should be consistent with the typical working terrain and real targets in the real environment. Add obstacles to the simulation scene according to the target recognition requirements.

[0038] Step 2: Create a model name-label correspondence index table for the simulation scene according to the required labels. Name the terrain, target models and obstacles in the simulation scene. The naming format is "name + serial number". The names of the same type of target are consistent and correspond to the same label. The serial number indicates the number of targets of that type.

[0039] Step 3: Take the geometric center of the target model as the origin O. t The target model moves in the direction of O. t X t Establish a right-handed three-dimensional rectangular coordinate system O along the axis. t X t Y t Z t The coordinate system O is defined at the initial moment. t X t Y t Z t Let O be the ground coordinate system. w X w Y w Z w .

[0040] Step 4: Perform Delaunay triangulation on the simulation scene to obtain a series of triangular facets. Store the triangular facets in an octree structure. Each node of the octree represents a cube voxel. Each node contains 8 child nodes. The sum of the volume elements of the 8 child nodes is equal to the volume of the parent node. Each triangular facet in the simulation scene corresponds to a leaf node.

[0041] Step 5: Add a virtual scanning system to the simulation scene. Set the parameters of the virtual scanning system according to the parameters of the lidar on the UAV, including the field of view angle θ. fov Angular resolution θ s and detection range [t] min ,t max To simulate the dynamic rendezvous process between an unmanned aerial vehicle (UAV) and a real target, a timeline is created. The motion speed v, scanning frequency f, and motion direction of the virtual scanning system are set. The position where the virtual scanning system emits scanning rays each time is calculated, and a keyframe is created at that position. The time start point is set to t0 = 0, and the time when the virtual scanning system emits scanning rays in the i-th time is t_i. i =v i / f; To achieve simulated point cloud generation for arbitrary flight trajectories, the trajectory equation of the virtual scanning system can be established, and the ground coordinate system O corresponding to each emitted scanning ray during the detection phase of the virtual scanning system can be calculated. w X w Y w Z w Spatial coordinates (x) i ,y i ,z i Simultaneously calculate the direction of the virtual scanning system's motion velocity (v) at that location. xi ,v yi ,v zi Import spatial coordinates (x) i ,y i ,z i The motion trajectory is constructed by combining the motion velocity and direction at the corresponding position.

[0042] Step 6: For each keyframe created in Step 5, use the spatial coordinates (x, y) of the virtual scanning system corresponding to the keyframe itself. i ,y i ,z i (O) is the origin. s O s X s Along the direction of the motion velocity, O s Y s The axis is perpendicular to the trajectory plane, establishing a virtual scanning system coordinate system O. s X s Y s Z s.

[0043] Step 7: At each keyframe, establish an interaction model with the simulation scene based on the parameters of the virtual scanning system. For a field of view of θ... fov Angular resolution is θ s The virtual scanning system has a corresponding number of emitted scanning rays, n = θ. fov / θ s The simulation point generation is modeled as a problem of finding the intersection of the scanning ray and the triangular elements of the simulation scene, thus obtaining the coordinates of the intersection point between the scanning ray and the simulation scene.

[0044] The interaction model between the virtual scanning system and the simulation scene is as follows:

[0045] Starting with the shallowest octree voxels, a list of voxels intersecting with the scan ray is obtained. Then, for each voxel in this list, its sub-voxels intersecting with the scan ray are solved, resulting in a deeper list. The final result is the list of voxels intersecting with the scan ray. The coordinates of the intersection points between the scan ray and the triangular facets in the intersecting voxel list are calculated. For a scan ray that intersects with multiple triangular facets, within the measurement range of the virtual scanning system [t]... min ,t max The point of intersection closest to the virtual scanning system is taken as the sampling point of the scanning ray in the target scene.

[0046] The specific method for determining whether the scanning ray intersects with the voxel is as follows:

[0047] Assume the virtual scanning system is in the ground coordinate system O w X w Y w Z w The position coordinates below are O(x) o ,y o ,z o The direction of the scanning ray is (α, β), where α is the direction of the scanning ray at X. s O s Y s Projection on the plane and O s X s The angle between the axis and β is the angle between the scanning ray and O. s Z s The angle between the axes, assuming P is a point on the scanning ray, and the distance between O and P is t (t>0), (X t ,Y t Z t Let t represent the coordinates of a point on the scanning ray that is a distance t from point O. Then the equation of the scanning ray is:

[0048]

[0049] To determine whether a scan ray intersects a voxel, it is only necessary to determine whether there exists any point on the scan ray within the voxel's volume space. Assume the coordinates of the lower left and upper right corners of a voxel are (x, y, y) and (x, y, y). min ,y min ,z min ) and (x max ,y max ,z max The conditions for the scanning ray to intersect with the voxel must be met:

[0050]

[0051] The method for calculating the coordinates of the intersection point between the scanning ray and the triangular element is as follows:

[0052] The starting point of the scanning ray is O(x) o ,y o ,z o If the direction of the scanning ray is (α, β), then the unit direction vector of the scanning ray can be expressed as: The vertex coordinates of the triangular facets are V1(x1,y1,z1), V2(x2,y2,z2), and V3(x3,y3,z3). Further solve the following equations:

[0053]

[0054] Where u and v represent positional relationship constants, and t also represents the distance between the intersection of the scanning ray and the triangular element. If the scanning ray and the triangular element have an intersection, the solution of the equation simultaneously satisfies the conditions t≥0, u≥0, v≥0 and u+v≤1. The intersection coordinates can be solved by combining the intersection distance t with equation (1). For scanning rays within the detection range, the coordinates of the intersection of the scanning ray and the triangular element are returned.

[0055] Step 8: Noise is unavoidable during the lidar imaging process. To improve the similarity between the simulated point cloud data and the actual scanned imaging data, this invention adds Gaussian white noise to the simulated point cloud data to improve the realism of the point cloud, as detailed below:

[0056] Establish a noise model related to the lidar parameters, and express the probability density function P(R) of lidar distance measurement as follows:

[0057]

[0058] Where R t The distance between the lidar and the target is represented by R, where R represents the measured distance, and δR is the ranging accuracy.

[0059]

[0060] In the formula, τ represents the peak half-width of the laser pulse, c is the speed of light, and SNR represents the signal-to-noise ratio of the receiving system. By setting parameters such as the pulse half-width and signal-to-noise ratio of the scanning system, the ranging accuracy δR of the scanning system is obtained, and a Gaussian distributed distance error value with a mean of 0 and a standard deviation of δR is generated. The distance error is added to the ideal distance value to generate simulated point cloud data containing Gaussian white noise.

[0061] Step 9: In order to apply the simulated point cloud data to the target recognition algorithm training, the intersection points of the scanning ray and the triangular facet are automatically labeled according to the model name-label correspondence index table of the simulation scene. If the scanning ray has a return intersection point, the model name corresponding to the triangular facet of that point is searched, and the label corresponding to the model name of that point is assigned, thus completing the accurate labeling of each intersection point.

[0062] Step 10: In the virtual scanning system coordinate system O s X s Y s Z s The following is generated with O s Multiple scanning rays emitted from the emission point into the field of view are used to obtain a contour line of the corresponding keyframe simulation scene. Coordinate transformation is then used to transform the coordinates of the contour line points to the ground coordinate system O. w X w Y w Z w Down.

[0063] Step 11: Segment the outlines corresponding to all keyframes to obtain the point cloud data of the simulation scene.

[0064] Step 12: Repeat steps 5 to 11 to create different motion trajectories for virtual scanning systems with different scanning field of view and different angular resolutions. Simulate the motion at different flight speeds, yaw angles, pitch angles, roll angles, and other attitudes to generate N frames of simulated point cloud data under different intersection conditions, forming a simulated point cloud dataset.

[0065] Step 13: Divide the simulated point cloud dataset into training, validation, and test sets in a 6:2:2 ratio, and input the dataset into the RandLA-Net model to complete the pre-training of the RandLA-Net model.

[0066] Step 14: Input the point cloud data to be identified into the trained RandLA-Net model, and output the category prediction result of each point in the point cloud data to be identified, thereby realizing the offline training and online recognition of the target recognition algorithm.

[0067] Example 1

[0068] like Figure 1 As shown in the figure, this embodiment discloses a target recognition method for linear laser imaging based on a simulated point cloud dataset, which includes the following steps:

[0069] Step 1: Build a simulation scene based on the typical working environment and the target recognition requirements of the scanning system. In this embodiment, a plane is used as the terrain model, and an M1A1 tank is used as the target to be identified. Trees and rocks are added as obstacles in the simulation scene. The 3D modeling software Blender is used to create a simulation scene. Figure 2 The simulation scenario shown.

[0070] Step 2: Establish a model name-tag correspondence index table for the simulation scene according to the required labels. Name the terrain, target models, and obstacles in the simulation scene. The naming format is "name + serial number". Targets of the same type have the same name and correspond to the same label. The serial number indicates the number of targets of that type. In this embodiment, the models in the simulation scene are divided into 4 categories: ground (model name "ground 001", label 0), tank (model name "tank 001", label 1), trees (model names "tree 001", "tree 002", "tree 003", "tree 004", label 2), and rocks (model names "rock 001", "rock 002", label 3). The model name-tag index table for the simulation scene is established as follows.

[0071] ground 0 tank 1 Trees 2 stone 3

[0072] Step 3: Take the geometric center of the target model as the origin O. t The target model moves in the direction of O. t X t Establish a right-handed three-dimensional rectangular coordinate system O along the axis. t X t Y t Z t The coordinate system O is defined at the initial moment. t X t Y t Z t Let O be the ground coordinate system. w X w Y w Z w .

[0073] Step 4: Perform Delaunay triangulation on the simulation scene to obtain a series of triangular facets, such as... Figure 3 As shown, triangular facets are stored in an octree structure. Each node of the octree represents a volume element of a cube. Each node contains 8 child nodes. The sum of the volume elements of the 8 child nodes is equal to the volume of the parent node. In the simulation scenario, each triangular facet corresponds to a leaf node.

[0074] Step 5: Add a virtual scanning system to the simulation scene and set the field of view θ of the virtual scanning system.fov =96°, angular resolution θ s =1°, flight altitude 30m, detection range [1m, 100m], the angle between the laser emission plane and the virtual scanning system's direction of motion is 60°, the roll angle, pitch angle, and yaw angle of the scanning system are all set to 0°, the flight speed is set to 100m / s, the scanning frequency is f = 500Hz, and the coordinate system is O w X w Y w Z w The spatial coordinates of the starting point of the mid-scan are (-28, 0, 30). The detection time for one frame of point cloud is 200ms, so the time interval between two adjacent scan lines is 2ms. One frame of point cloud data contains 100 scan lines, therefore 100 keyframes are set to form the motion trajectory, as follows. Figure 4 .

[0075] Step 6: For each keyframe created in Step 5, use the spatial coordinates (x, y) of the virtual scanning system corresponding to the keyframe itself. i ,y i ,z i (O) is the origin. s O s X s Along the direction of the motion velocity, O s Y s The axis is perpendicular to the trajectory plane, establishing a virtual scanning system coordinate system O. s X s Y s Z s .

[0076] Step 7: At each keyframe, establish a simulation scene interaction model based on the virtual scanning system parameters. In this embodiment, based on the set θ... fov and θ s The virtual scanning system has n = 96 lines. The generation of simulation points is modeled as a problem of finding the intersection of scanning rays and triangular elements of the simulation scene. In the coordinate system of the scanning system, 96 scanning rays are generated with the position of the virtual scanning system as the emission point and emitted into the field of view.

[0077] Taking a keyframe as an example, such as Figure 5 As shown, in the virtual scanning system coordinate system O s X s Y s Z s In the diagram, the starting coordinates of the scanning ray are O(0,0,0), and its direction is (α,β). The unit direction vector of the scanning ray can be... Assume P is a point on the scanning ray, and the distance between O and P is t (t>0), (X t ,Y t Zt Let t represent the coordinates of a point on the scanning ray that is a distance t from point O. Then the equation of the scanning ray is:

[0078]

[0079] The interaction model calculation process between the virtual scanning system and the simulation scene is as follows: Figure 6 As shown, starting from the shallowest octree voxels, a list of voxels intersecting with the scanning ray is obtained. Then, in each voxel in the voxel list, the sub-voxels intersecting with the scanning ray are solved, and a new list at a deeper level is obtained. The final result is the list of voxels intersecting with the scanning ray. The coordinates of the three vertices V1, V2, and V3 of the triangular facet in the intersecting voxel list are obtained, and equation (3) is solved. If the solution of the equation satisfies the conditions t≥0, u≥0, v≥0 and u+v≤1, then the value of t is saved. Since a scanning ray may intersect with multiple triangular facets at the same time, a distance value list {t1, t2, ..., t3} can be obtained. n Considering occlusion during the actual scanning process, we take the list {t1,t2,…,t}. n The minimum distance value t in} m As the distance value measured by the virtual scanning system.

[0080] Step 8: To improve the similarity between the simulated point cloud data and the real imaging data, Gaussian white noise is added to the simulated point cloud data to synthesize noisy point cloud data. The pulse width τ of the virtual scanning system is set to 2ns, and the signal-to-noise ratio SNR is set to 10. A distance error value Δt that conforms to the distribution of equation (4) is generated. The error value is added to the distance measurement value to obtain the distance t containing the measurement error. f =t m +Δt, if t f Within the measurement range of the virtual scanning system, the coordinates of the intersection point in the virtual scanning system coordinate system can be calculated as P(t). f cosαcosβ,t f sinαcosβ,t f sinβ).

[0081] Step 9: Automatically label the intersection points of the scan ray and the triangular facets based on the simulation scene model name-label correspondence index table established in Step 2. If the scan ray returns to an intersection point, search for the model name corresponding to the triangular facet at that point and assign the label corresponding to that model name. This completes the accurate labeling of each intersection point, as shown in Table 1.

[0082] Table 1. Three-dimensional coordinates and corresponding labels of some simulation points.

[0083]

[0084] Step 10: Calculate the intersection points of the 96 scanning rays in the keyframe with the simulation scene to obtain a contour line of the simulation scene corresponding to the keyframe. Transform the coordinates of the contour line points to the ground coordinate system O. w X w Y w Z w Below, such as Figure 7 The simulation results are for two contour lines.

[0085] Step 11: Stitch together the contour lines corresponding to all keyframes to obtain the point cloud data of the simulation scene, such as... Figure 8 As shown.

[0086] Step 12: Based on the above steps, create different motion trajectories for virtual scanning systems with different scanning field of view and different angular resolutions. Simulate different flight speeds, yaw angles, pitch angles, roll angles, and other attitudes to generate N frames of simulated point cloud data under different intersection conditions, forming a simulated point cloud dataset.

[0087] Step 13: Divide the simulation point cloud dataset containing N frames of simulation point cloud data into a training set, a validation set, and a test set in a 6:2:2 ratio. Input the dataset into the RandLA-Net model to complete the pre-training of the RandLA-Net model.

[0088] Step 14: Input the point cloud data to be identified into the trained RandLA-Net model, and output the category prediction result of each point in the point cloud data to be identified, thereby realizing the offline training and online recognition of the target recognition algorithm.

Claims

1. A method for target recognition in linear laser imaging based on simulated point cloud datasets, characterized in that, Includes the following steps: Step 1: Build a simulation scenario based on the typical working environment of the unmanned aerial vehicle and its target recognition requirements: Create 3D terrain and target models for the simulation scene using 3D modeling software. The 3D terrain and target models of the simulation scene should be consistent with the typical working terrain and real targets in the real environment. Add obstacles to the simulation scene according to the target recognition requirements. Step 2: Create a model name-tag correspondence index table for the simulation scene according to the required labels. Name the terrain, target models and obstacles in the simulation scene. The naming format is "name + number". The names of the same type of target are consistent and correspond to the same label. The number indicates the number of targets of that type. Step 3: Take the geometric center of the target model as the origin O. t The target model moves in the direction of O. t X t Establish a right-handed three-dimensional rectangular coordinate system O along the axis. t X t Y t Z t The coordinate system O is defined at the initial moment. t X t Y t Z t Let O be the ground coordinate system. w X w Y w Z w ; Step 4: Perform Delaunay triangulation on the simulation scene to obtain a series of triangular facets. Store the triangular facets in an octree structure. Each node of the octree represents a cube voxel. Each node contains 8 child nodes. The sum of the volume elements of the 8 child nodes is equal to the volume of the parent node. Each triangular facet in the simulation scene corresponds to a leaf node. Step 5: Add a virtual scanning system to the simulation scene. Set the parameters of the virtual scanning system according to the parameters of the lidar on the UAV, including the field of view angle θ. fov Angular resolution θ s and detection range [t] min ,t max To simulate the dynamic rendezvous process between an unmanned aerial vehicle (UAV) and a real target, a timeline is created. The motion speed v, scanning frequency f, and motion direction of the virtual scanning system are set. The position where the virtual scanning system emits scanning rays each time is calculated, and a keyframe is created at that position. The time start point is set to t0 = 0, and the time when the virtual scanning system emits scanning rays in the i-th time is t_i. i =v i / f; To achieve simulated point cloud generation for arbitrary flight trajectories, the trajectory equation of the virtual scanning system is established, and the ground coordinate system O is calculated at each emitted scanning ray moment during the detection phase of the virtual scanning system. w X w Y w Z w Spatial coordinates (x) i ,y i ,z i Simultaneously calculate the direction of the virtual scanning system's motion velocity (v) at that location. xi ,v yi ,v zi Import spatial coordinates (x) i ,y i ,z i The motion trajectory is constructed by combining the direction of motion velocity at the corresponding position. Step 6: For each keyframe created in Step 5, use the spatial coordinates (x, y) of the virtual scanning system corresponding to the keyframe itself. i ,y i ,z i (O) is the origin. s O s X s Along the direction of the motion velocity, O s Y s The axis is perpendicular to the trajectory plane, establishing a virtual scanning system coordinate system O. s X s Y s Z s ; Step 7: At each keyframe, establish an interaction model with the simulation scene based on the parameters of the virtual scanning system. For a field of view of θ... fov Angular resolution is θ s The virtual scanning system has a corresponding number of emitted scanning rays, n = θ. fov / θ s The simulation point generation is modeled as a problem of finding the intersection of the scanning ray and the triangular facets of the simulation scene, thus obtaining the coordinates of the intersection point between the scanning ray and the simulation scene; Step 8: Improve the realism of the point cloud by adding Gaussian white noise to the simulated point cloud data, that is, add distance error to the ideal distance value to generate simulated point cloud data containing Gaussian white noise; Step 9: Automatically label the intersection points of the scanning ray and the triangular facet according to the model name-label correspondence index table of the simulation scene. If the scanning ray has a return intersection point, search for the model name corresponding to the triangular facet at that point, assign the label corresponding to the model name at that point, and complete the accurate labeling of each intersection point. Step 10: In the virtual scanning system coordinate system O s X s Y s Z s The following is generated with O s Multiple scanning rays emitted from the emission point into the field of view are used to obtain a contour line of the corresponding keyframe simulation scene. Coordinate transformation is then used to transform the coordinates of the contour line points to the ground coordinate system O. w X w Y w Z w Down; Step 11: Stitch together the outlines corresponding to all keyframes to obtain the point cloud data of the simulation scene; Step 12: Repeat steps 5 to 11 to create different motion trajectories for virtual scanning systems with different scanning field of view and different angular resolutions. Simulate the motion at different flight speeds, yaw angles, pitch angles, and roll angles to generate N frames of simulated point cloud data under different rendezvous conditions, thus forming a simulated point cloud dataset. Step 13: Divide the simulated point cloud dataset into training set, validation set, and test set according to a 6:2:2 ratio, and input the dataset into the RandLA-Net model to complete the pre-training of the RandLA-Net model; Step 14: Input the point cloud data to be identified into the trained RandLA-Net model, and output the category prediction result of each point in the point cloud data to be identified, thereby realizing the offline training and online recognition of the target recognition algorithm.

2. The target recognition method for linear laser imaging based on simulated point cloud datasets according to claim 1, characterized in that: In step 7, the interaction model between the virtual scanning system and the simulation scene is as follows: Starting with the shallowest octree voxels, a list of voxels intersecting with the scanning ray is obtained. Then, for each voxel in this list, its child voxels intersecting with the scanning ray are solved, resulting in a deeper list. The final result is the list of voxels intersecting with the scanning ray. The coordinates of the intersection points between the scanning ray and the triangular facets in the intersecting voxel list are calculated. For a scanning ray that intersects with multiple triangular facets, the coordinates are calculated within the measurement range of the virtual scanning system [t]. min ,t max The point of intersection closest to the virtual scanning system is taken as the sampling point of the scanning ray in the target scene.

3. The target recognition method for linear laser imaging based on simulated point cloud datasets according to claim 2, characterized in that, The determination of whether the scanning ray intersects with the voxel is as follows: Assume the virtual scanning system is in the ground coordinate system O w X w Y w Z w The position coordinates below are O(x) o ,y o ,z o The direction of the scanning ray is (α, β), where α is the direction of the scanning ray at X. s O s Y s Projection on the plane and O s X s The angle between the axis and β is the angle between the scanning ray and O. s Z s The angle between the axes, assuming P is a point on the scanning ray, and the distance between O and P is t, t>0, (X t ,Y t Z t Let t represent the coordinates of a point on the scanning ray that is a distance t from point O. Then the equation of the scanning ray is: To determine whether a scan ray intersects a voxel, it is only necessary to determine whether there exists any point on the scan ray within the voxel's volume space. Assume the coordinates of the lower left and upper right corners of a voxel are (x, y, y) and (x, y, y). min ,y min ,z min ) and (x max ,y max ,z max The conditions for the scanning ray to intersect with the voxel must be met:

4. The target recognition method for linear laser imaging based on simulated point cloud datasets according to claim 3, characterized in that, The method for calculating the coordinates of the intersection point between the scanning ray and the triangular element is as follows: The starting point of the scanning ray is O(x) o ,y o ,z o If the direction of the scanning ray is (α, β), then the unit direction vector of the scanning ray is expressed as: The vertex coordinates of the triangular facets are V1(x1,y1,z1), V2(x2,y2,z2), and V3(x3,y3,z3). Further solve the following equations: Where u and v both represent positional relationship constants. If the scanning ray intersects with the triangular element, the solution of the equation simultaneously satisfies the conditions t≥0, u≥0, v≥0 and u+v≤1. Based on the intersection distance t and equation (1), the coordinates of the intersection point are solved. For scanning rays within the detection range, the coordinates of the intersection point between the scanning ray and the triangular element are returned.

5. The target recognition method for linear laser imaging based on simulated point cloud datasets according to claim 1, characterized in that: Step 8: Improve the realism of the point cloud by adding Gaussian white noise to the simulated point cloud data, as detailed below: Establish a noise model related to the lidar parameters, and express the probability density function P(R) of lidar distance measurement as follows: Where R t The distance between the lidar and the target is represented by R, where R represents the measured distance, and δR is the ranging accuracy. In the formula, τ represents the peak half-width of the laser pulse, c is the speed of light, and SNR represents the signal-to-noise ratio of the receiving system. By setting the pulse half-width and signal-to-noise ratio of the scanning system, the ranging accuracy δR of the scanning system is obtained, and a Gaussian distributed distance error value with a mean of 0 and a standard deviation of δR is generated. The distance error is added to the ideal distance value to generate simulated point cloud data containing Gaussian white noise.

Images