An unmanned vehicle obstacle avoidance method based on three-dimensional point cloud depth detection

Abstract

This invention discloses an obstacle avoidance method for unmanned vehicles based on 3D point cloud depth detection, comprising: 1. acquiring and preprocessing the original environmental point cloud using the unmanned vehicle's own TOF camera to obtain a preprocessed non-ground point cloud in the unmanned vehicle's world coordinate system; 2. extracting local geometric features from the point cloud and selecting key points, estimating object positions through voting clustering, and decoding the 3D bounding box of the point cloud; 3. constructing a virtual extended body of the unmanned vehicle, processing its dimensions, and combining its center coordinates and heading angle to achieve global perception of the potential collision space around the unmanned vehicle; 4. judging potential collision risks and making motion decisions based on the real-time interaction between the virtual extended body and the 3D bounding box of the point cloud. This invention can directly process the point cloud acquired by the depth camera, enabling the unmanned vehicle to quickly detect obstacles ahead, thereby achieving intelligent obstacle avoidance.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and autonomous driving technology, and more specifically, to an obstacle avoidance method for unmanned vehicles based on 3D point cloud depth detection. Background Technology

[0002] Mobile robots are automated machines capable of performing tasks. They can be controlled by humans, run pre-programmed procedures, or act according to principles established using artificial intelligence. They represent a broad category of robots, encompassing numerous modules such as environmental perception, navigation and localization, and motion control, involving multiple disciplines including mechanics, automation, and computer science. Obstacle avoidance is a crucial and essential function for mobile robots. The ability to detect static or dynamic objects obstructing their path through sensors and then effectively avoid them using specific methods is a hot topic in current mobile robot research.

[0003] Currently, most common robot obstacle avoidance systems on the market use LiDAR, which offers advantages such as high precision, achieving millimeter-level accuracy. It also boasts strong anti-interference capabilities, is unaffected by ambient lighting, and exhibits high stability, making it suitable for navigation tasks in complex environments. However, LiDAR navigation requires the pre-establishment of accurate maps, and its updating and maintenance are highly demanding. Thanks to the rapid development in the field of computer vision, many new technologies have been proposed in recent years... A point cloud-based 3D object detection method. Because point cloud data preserves the precise depth and geometric structure information of objects, using point clouds for 3D object detection can achieve higher accuracy.

[0004] Currently, there are three main methods for object detection using point clouds: The first method divides the point cloud into a 3D voxel grid. Points within each voxel are aggregated and fed into a 3D convolutional neural network for processing. This method ensures training and inference speed, but voxelization introduces quantization differences, inevitably leading to information loss during voxel sampling. The second method is based on a bird's-eye view of the point cloud, projecting it onto a 2D plane from a ground perspective and then using a mature 2D object detection network for fast processing. However, this method loses height information and fine geometric structure during projection. The third method operates directly on the original point cloud, using point-level networks to extract features. This method preserves the geometric information of the original points and has high processing accuracy, but the number of points in the environment is usually very large, making point-level operations inefficient and difficult to perform in real-time. Summary of the Invention

[0005] The present invention addresses the shortcomings of the existing technology by proposing an obstacle avoidance method for unmanned vehicles based on 3D point cloud depth detection. The aim is to enable unmanned vehicles to quickly detect obstacles ahead by directly processing the point cloud acquired by the depth camera, thereby effectively realizing intelligent obstacle avoidance.

[0006] To solve the technical problem, the present invention adopts the following technical solution: The present invention provides an obstacle avoidance method for unmanned vehicles based on 3D point cloud depth detection, characterized by the following steps: Step S1: Establish the camera coordinate system The world coordinate system of driverless cars ; Step S2: The TOF camera of the autonomous vehicle can acquire data in... Original environment point cloud in coordinate system After coordinate transformation, the following is obtained Environmental point cloud in coordinate system And then Crop the data to obtain non-ground point clouds. ; Step S3, for Noise filtering is performed to remove outliers in the environment, resulting in a cleaned-up non-terrestrial point cloud. ; Step S4, for After preprocessing, the preprocessed non-terrestrial point cloud is obtained. ; Step S5: Use a voting-clustering-regression network model to... Target localization is performed to obtain 3D target detection results; Step S6: Construct the virtual extended body of the autonomous vehicle at time t. And serve as a collision prediction zone; Step S7: Combine virtual extensions Based on the 3D target detection results, path planning is performed for the unmanned vehicle to achieve obstacle avoidance.

[0007] The obstacle avoidance method for unmanned vehicles based on 3D point cloud depth detection described in this invention is characterized in that step S1 includes: S1.1: The origin of the coordinate system is the center point of the autonomous vehicle. Taking the direction of the driverless car as The positive direction of the axis, perpendicular to The direction of the axis vertically upward is The positive direction of the axis is determined by the right-hand rule. The positive direction of the axis is used to construct the world coordinate system of the autonomous vehicle. ; S1.2: Using the light source of the TOF camera As the origin of the coordinate system, the area directly in front of the TOF camera is set as... The positive direction of the axis, perpendicular to The vertical downward direction of the axis is The positive direction of the axis is determined by the right-hand rule. The positive direction of the axis is used to establish the camera coordinate system. .

[0008] Furthermore, in step S2, a random sample consensus algorithm is used to... Cutting: from Three points were randomly selected multiple times and fitted to different planar models, and the calculations were performed. The distances from the remaining points in the model to each plane are calculated, and those distances are less than a threshold. The points are set as interior points. Then, the planar model with the most interior points is used as the ground model. All interior points on the ground model are then deleted, and the remaining point cloud is used as the point cloud for non-ground objects. .

[0009] Furthermore, step S3 includes: exist The KD-Tree accelerated search algorithm is used to obtain the results. any i-th point The closest Find the nearest neighboring points and calculate... and Average distance of neighboring points ,statistics The average distance between any point and all other points and its standard deviation ,if average distance Exceed Then Mark as outlier Then you can Remove all outliers in the array to obtain the result. in, It is a threshold that is a multiple of the standard deviation.

[0010] Furthermore, step S4 includes: S4.1: Set the given number of target points ,if Number of point clouds > Then from Random downsampling If there are sampling points, Number of point clouds < Then from the time Repeat sampling until reached Each sampling point is used to obtain the sampled non-ground point cloud. ; S4.2: Calculation Geometric centroid of all sampling points coordinates Thus, to Any number in the middle sampling points coordinates Perform a translation to obtain the translated [number]. sampling points coordinates ; S4.3: Calculation All translated sampling points are respectively mapped to the world coordinate system of the autonomous vehicle. origin of coordinates The distances are calculated, and the furthest distance is selected from them. Afterwards, Divide by Thus, the normalized first... sampling points ; S4.4: For the first sampling points After performing random rotation, random scaling, and random translation operations on the three-dimensional coordinates, the preprocessed first-order coordinates are obtained. This process generates a preprocessed non-ground point cloud from a number of non-ground points. .

[0011] Furthermore, step S5 includes: S5.1: Using the hierarchical feature extraction network PointNet++ to process preprocessed non-terrestrial point clouds Multi-scale feature learning and downsampling are performed to generate a set of seed points containing local geometric information and contextual semantic information. ; S5.2: with Any j-th seed point Using the center as the reference point, a radius neighborhood search algorithm is employed to determine... The set of K neighboring points within the neighborhood of . Features of its neighboring points ;in, express The k-th neighboring point within the neighborhood of ; express Features; S5.3: Calculate the j-th seed point using equation (1) query vector The key vector of the k-th neighboring point and the value vector of the k-th neighboring point : (1) In equation (1), express Features , , These are three linear transformation weight matrices to be learned; S5.4: Calculate the j-th seed point using equation (2) With the k-th neighbor point Attention weights between : (2) In equation (2), is the dimension of the vector; T denotes transpose; S5.5: Obtained using equation (3) Neighborhood enhancement features and will and After concatenation, a nonlinear transformation is performed to obtain the enhanced features of the j-th seed point. ; = (3) S5.6: Calculation right The offset is calculated to obtain K offsets, and the neighborhood point with the smallest offset is selected as the nearest neighbor. polling stations and to All polling stations Clustering is performed to obtain The clustering results and cluster centers are analyzed, and several center proposal points are randomly selected from the cluster centers. For each selected center proposal point... Find all the polling stations in its neighborhood. ; Based on the geometric and contextual semantic information of the seed point inherited from the voting point, The geometric semantic features are aggregated into each central proposal point. Comprehensive proposal features Each central proposal point and its corresponding comprehensive proposal features A binary tuple that merges into a single candidate target The 3D object detection results are obtained by predicting the bounding boxes of all candidate objects using a multilayer perceptron. The position and parameters.

[0012] Furthermore, step S6 includes: S6.1: Obtain the autonomous vehicle in the world coordinate system The pose parameters and inherent geometric dimensions are given below, the pose parameters including: center coordinates. and heading angle The inherent geometric dimensions include: the length of the autonomous vehicle. Width of driverless cars The height of driverless cars ; S6.2: Define a scaling factor based on the inherent geometric dimensions of the autonomous vehicle. Thus, the length of the virtual extended body can be obtained using equation (4). Width of the virtual extension and the height of the virtual extension Thus, a virtual extended body is constructed at time t. ,in, This represents the coordinates of the center point of the virtual extended body at time t. The heading angle of the virtual extended body at time t is: , , (4) S6.3: The virtual extension is updated in real time as the autonomous vehicle moves. In This ensures that it always moves in sync with the driverless car and maintains the same direction of travel as the driverless car.

[0013] Furthermore, step S7 includes: S7.1: Calculate the overlap region of the two in space at time t using equation (5). : (5) S7.2: If Then it represents the virtual extended body at time t. With 3D bounding box set If one or more of them overlap, S7.3 is executed; otherwise, the driverless vehicle is controlled to continue driving along the original path. S7.3: Control the unmanned vehicle to stop and wait within the set waiting time, and determine whether the position and parameters of the 3D bounding box in the overlapping area have changed during the waiting time; if they have changed, execute S7.4; otherwise, replan the path. S7.4: Based on the virtual extended body at time t Reassess the changed environmental information. If the value is greater than 0, and less than 0, it means that the original planned path is safe, and the driverless car continues to drive along the original planned path; otherwise, the path is replanned. S7.5: If the waiting time exceeds the threshold Then determine the 3D bounding box set. With a static bounding box, the autonomous vehicle replans its path.

[0014] The present invention provides an electronic device, including a memory and a processor, characterized in that the memory is used to store a program supporting the processor in performing the method described therein, and the processor is configured to execute the program stored in the memory.

[0015] The present invention discloses a computer-readable storage medium storing a computer program, characterized in that the computer program is executed by a processor to perform the steps of the method described thereon.

[0016] Compared with existing technologies, the beneficial effects of this invention are reflected in: 1. This invention directly processes the raw point cloud data obtained by the camera and then uses it as input. Compared with the traditional method of converting the point cloud into a planar form, this effectively avoids the loss of point cloud geometric information and improves the robustness of object detection.

[0017] 2. This invention uses a random sampling consensus algorithm to process environmental point clouds in the world coordinate system. The processing of point clouds can effectively remove the acquired ground point cloud data and reduce the density of environmental point clouds, which can effectively improve the speed of target detection.

[0018] 3. This invention uses the KD-Tree accelerated search algorithm to find the neighboring points of each point in the point cloud. Then, through calculation and comparison, outliers in the neighboring points are obtained. Outliers are removed to obtain clutter-free point cloud data, which can improve the target detection speed.

[0019] 4. This invention employs a voting-clustering-regression network model to extract features and locate targets in the preprocessed point cloud. After domain feature enhancement and weighted voting, data sets about candidate targets can be fused, allowing for the prediction of 3D bounding boxes for all candidate targets. This improves the accuracy of 3D bounding box detection for target objects.

[0020] 5. The present invention proposes a new obstacle avoidance scheme for unmanned vehicles. By analyzing the overlap between the created virtual extension of the unmanned vehicle and the 3D target detection bounding box, the current road conditions of the unmanned vehicle are judged to determine whether it is passable, thereby realizing the dynamic planning and adjustment of the unmanned vehicle trajectory. Attached Figure Description

[0021] Figure 1 This is a flowchart of the obstacle avoidance method for unmanned vehicles based on 3D point cloud depth detection according to the present invention. Figure 2 This is a schematic diagram of the coordinate transformation of the present invention; Figure 3 This is the overall network structure diagram of the three-dimensional point cloud target detection method of the present invention. Detailed Implementation

[0022] In this embodiment, an obstacle avoidance method for unmanned vehicles based on 3D point cloud depth detection is presented. This method operates directly on the original point cloud data, fully preserving the geometric structure information of the environment. The execution flow is as follows: Figure 1 As shown, specifically, the following steps are performed: Step S1: Establish the camera coordinate system and the world coordinate system of the autonomous vehicle; Using the center point of the autonomous vehicle as the origin of the coordinate system The direction of travel is Positive axis direction, perpendicular The axis is vertically upward. The positive direction of the axis is determined by the right-hand rule. The positive axis is used to construct the world coordinate system of the autonomous vehicle. ; using the light source of a TOF camera As the origin of the coordinate system, let the area directly in front of the camera be... Positive axis direction, perpendicular to Vertically downwards The positive direction of the axis is determined by the right-hand rule. The camera coordinate system is established by determining the axial direction. The relationship between the world coordinate system and the camera coordinate system is as follows: Figure 2 As shown. In specific implementations, the depth camera used can be a structured light camera, a TOF camera, or a binocular camera.

[0023] Step S2: Crop the environmental point cloud in the world coordinate system using the random sample consensus algorithm, filtering out point clouds belonging to the ground environment and retaining non-ground point clouds. To reduce the amount of additional computation; In practice, the autonomous vehicle uses its own TOF camera to obtain the camera coordinate system. Original 3D point cloud After performing coordinate transformation, the world coordinate system of the autonomous vehicle is obtained. Original 3D point cloud Point clouds in the world coordinate system are analyzed using a random sampling consensus algorithm. Cut it, and then from The process involves randomly selecting three points multiple times and fitting them into different planar models, while calculating the model in each iteration. The distances from the remaining points in the model to each plane are calculated, and those distances are less than a threshold. The points are set as interior points. Then, the planar model with the most interior points is used as the ground model. All interior points on the ground model are then deleted, and the remaining point cloud is used as the non-ground object point cloud. .

[0024] Step S3: For Noise filtering is performed to remove outliers from the environment, resulting in a cleaned-up non-terrestrial point cloud. ; In practical implementation, non-ground point clouds Perform noise filtering, Each point in Find its nearest neighbor using the KD-Tree accelerated search algorithm. Calculate the nearest neighboring points. Arrive here Average distance of neighboring points And statistics Average distance from any point in the array to all other points and its standard deviation If a point average distance Exceed If so, mark it as an outlier. Calculate all outliers Then remove it to obtain the point cloud after removing noise. .in, It is a threshold that is a multiple of the standard deviation.

[0025] Search each point in the point cloud using a KD-Tree. Find the nearest neighboring points and calculate... The average distance of each neighboring point is calculated; the average distance and standard deviation of all points are calculated; outliers are removed based on the relationship between the average distance of a single point and the average distance and standard deviation of all points, thus filtering out noise points and obtaining a point cloud after noise removal. Step S4: Process the point cloud after removing noise. Preprocessing is performed to obtain the preprocessed non-ground point cloud. To improve the robustness of the model to point cloud data transformations and enhance its generalization ability in different objects and scenes; Step S4.1: Process the point cloud after removing clutter. Preprocessing is performed by uniformly sampling the target point cloud, setting a given number of target points. If the current number of point clouds > Then, randomly downsample from the current point cloud. If the current number of points is [number] points < Then, repeat sampling from the current point cloud until you have... Each point is used to obtain the sampled non-ground point cloud. .

[0026] Step S4.2: Calculation Geometric centroid of all sampling points ,Will any of the first Coordinates of each sampling point Translate to obtain the translated first... Coordinates of each sampling point ; Step S4.3: Calculation The sampling points after translation to the world coordinate system of the autonomous vehicle The origin farthest distance , each point Divide by The normalized first number is obtained sampling points ; Step S4.4: Set each sampling point The XYZ coordinates are connected with other available feature information to form a longer feature vector; Step S4.5: For each sampling point The data undergoes random rotation, scaling, and translation to enhance the model's generalization ability, resulting in preprocessed non-terrestrial point cloud data. .

[0027] Step S5: Based on the preprocessed point cloud A voting-clustering-regression network model is adopted to further extract features and locate targets, so as to achieve refined 3D target detection and target detection results; Step S5.1: Use the hierarchical feature extraction network PointNet++ to process the preprocessed scene point cloud. Multi-scale feature learning and downsampling are performed to generate a set of seed points containing local geometric information and contextual semantic information. ; Step S5.2: Set the seed points Any j-th seed point Using [a specific point] as the center point, a radius neighborhood search algorithm is used to determine [the location / location]. The set of K points in the neighborhood and the characteristics of these neighboring points ;in, express The k-th neighboring point within the neighborhood of ; express Its characteristics.

[0028] Step S5.3: Set each seed point Features as query vector Neighborhood point features as a key vector Sum value vector Each is achieved through three learnable linear transformation weight matrices. , , The above three sets of vectors are obtained as shown in equation (1): (1) Step S5.4: Calculate the attention weights between the seed point and its neighbors. The weights between the seed point and its neighbors are calculated using Equation (2) as shown in the scaling dot product attention mechanism. (2) In equation (2), This represents the self-feature of the j-th seed point. After linear transformation The resulting vector Features of the Kth neighboring point After linear transformation The resulting vector This represents the dot product of the query and the key. It is a query vector and key vector The dimensionality, that is, the point cloud features obtained after linear transformation. / The dimension of a vector.

[0029] Step S5.5: Perform attention-weighted summation on the value vectors of the neighboring points, which can be obtained from equation (3). Neighborhood Enhancement Features and enhance neighborhood features Features of the original seed point Enhanced seed point features are obtained by fusing features through feature concatenation and nonlinear transformation. ; = (3) Step S5.6: Enhance the seed point features The data is input into a weighted voting process, which calculates the offset of each point from the target center and selects the neighborhood point with the smallest offset as the target center. polling stations ,right All polling stations Clustering is performed to obtain The clustering results and their cluster centers were analyzed, and several proposed center points were selected. .

[0030] For each center proposal Find all the polling stations in its neighborhood. Then the voting point is obtained from the seed point, that is, the voting point inherits geometric information and contextual semantic information.

[0031] Will The geometric semantic features are aggregated into each central proposal point. Comprehensive proposal features The central proposal point and its corresponding comprehensive proposal features Merge into a candidate target tuple Predict 3D bounding boxes for all candidate targets using a multilayer perceptron. and the target category A complete network model is as follows: Figure 3 As shown.

[0032] A voting-clustering-regression network model is adopted. A hierarchical feature extraction network is used to learn and downsample the preprocessed scene point cloud to obtain a seed point set. The seed points are selected as center points, and a radius neighborhood search is used to determine the neighboring points and their features. The attention weights are then calculated. and neighborhood enhancement features ,Will and The enhanced seed point features are obtained by fusion. ,Will The input weighted voting yields voting points, which are then processed to obtain center proposal points and comprehensive proposal features. These two are fused into candidate target tuples, which can predict the 3D bounding boxes of all candidate targets. and the target category .

[0033] Step S6: Construct the virtual extended body of the autonomous vehicle at time t As a collision prediction zone; Step S6.1: Obtain the autonomous vehicle in the world coordinate system The real-time pose parameters and intrinsic geometry are shown below. The pose parameters include the center coordinates. and heading angle The inherent geometric dimensions are .in, It is an unmanned vehicle commander, It is the width of the driverless car, It's an autonomous vehicle.

[0034] Step S6.2: Based on the geometry of the autonomous vehicle Define a scaling factor Construct the virtual extended body at time t .in, This represents the coordinates of the center point of the virtual extended body at time t. This represents the heading angle of the virtual extended body at time t. Indicates the length of the virtual extension. Indicates the width of the virtual extension. The height of the virtual extension is represented by formula (4): , , (4) Step S6.3: As the autonomous vehicle moves, the virtual extended body Its pose parameters are updated in real time. It always moves synchronously with the center of the vehicle and maintains the same direction of travel as the driverless car.

[0035] Construct a virtual extension of the unmanned vehicle at time t. Set a scaling factor based on the actual size of the unmanned vehicle, calculate the size of the virtual extension, and determine the heading angle and coordinates.

[0036] Step S7: Combine the virtual extended body with the 3D target detection results to perform path planning; Step S7.1: Combine virtual extensions With 3D bounding box Determine whether the two overlap in space: .

[0037] Step S7.2: If Then it represents the virtual extended body at time t. and If an overlap occurs, proceed to step S7.3; otherwise, control the autonomous vehicle to continue traveling along the original path.

[0038] Step S7.3: Control the unmanned vehicle to stop and wait within the set waiting time, and determine the time within the waiting time. Check if the location and parameters have changed; if they have changed, proceed to step S7.4; otherwise, replan the path.

[0039] Step S7.4: Based on the virtual extended volume at time t Reassess the changed environmental information. Whether the value is greater than 0 indicates whether the originally planned path is safe; if... If the original planned path is safe, the driverless vehicle will continue to travel along the original planned path; otherwise, the path needs to be replanned.

[0040] Step S7.5: If the waiting time exceeds the threshold Then determine the 3D bounding box. For static bounding boxes, the autonomous vehicle needs to replan its path.

[0041] By combining the virtual extension and target detection results, it can be determined whether the two overlap, thus determining whether the autonomous vehicle's driving path is feasible.

[0042] In this embodiment, an electronic device includes a memory and a processor. The memory stores a program that supports the processor in executing the above-described method, and the processor is configured to execute the program stored in the memory.

[0043] In this embodiment, a computer-readable storage medium stores a computer program, which is executed by a processor to perform the steps of the above method.

Claims

1. An obstacle avoidance method for unmanned vehicles based on 3D point cloud depth detection, characterized in that, Includes the following steps: Step S1: Establish the camera coordinate system The world coordinate system of driverless cars ; Step S2: The TOF camera of the autonomous vehicle can acquire data in... Original environment point cloud in coordinate system After coordinate transformation, the following is obtained Environmental point cloud in coordinate system And then Crop the data to obtain non-ground point clouds. ; Step S3, for Noise filtering is performed to remove outliers in the environment, resulting in a cleaned-up non-terrestrial point cloud. ; Step S4, for After preprocessing, the preprocessed non-terrestrial point cloud is obtained. ; Step S5: Use a voting-clustering-regression network model to... Target localization is performed to obtain 3D target detection results; Step S6: Construct the virtual extended body of the autonomous vehicle at time t. And serve as a collision prediction zone; Step S7: Combine virtual extensions Based on the 3D target detection results, path planning is performed for the unmanned vehicle to achieve obstacle avoidance.

2. The obstacle avoidance method for unmanned vehicles based on 3D point cloud depth detection according to claim 1, characterized in that, Step S1 includes: S1.1: The origin of the coordinate system is the center point of the autonomous vehicle. Taking the direction of the driverless car as The positive direction of the axis, perpendicular to The direction of the axis vertically upward is The positive direction of the axis is determined by the right-hand rule. The positive direction of the axis is used to construct the world coordinate system of the autonomous vehicle. ; S1.2: Using the light source of the TOF camera As the origin of the coordinate system, the area directly in front of the TOF camera is set as... The positive direction of the axis, perpendicular to The vertical downward direction of the axis is The positive direction of the axis is determined by the right-hand rule. The positive direction of the axis is used to establish the camera coordinate system. .

3. The obstacle avoidance method for unmanned vehicles based on 3D point cloud depth detection according to claim 1, characterized in that, In step S2, the random sample consensus algorithm is used to... Cutting: from Three points were randomly selected multiple times and fitted to different planar models, and the calculations were performed. The distances from the remaining points in the model to each plane are calculated, and those distances are less than a threshold. The points are set as interior points. Then, the planar model with the most interior points is used as the ground model. All interior points on the ground model are then deleted, and the remaining point cloud is used as the point cloud for non-ground objects. .

4. The obstacle avoidance method for unmanned vehicles based on 3D point cloud depth detection according to claim 1, characterized in that, Step S3 includes: exist The KD-Tree accelerated search algorithm is used to obtain the results. any i-th point The closest Find the nearest neighboring points and calculate... and Average distance of neighboring points ,statistics The average distance between any point and all other points and its standard deviation ,if average distance Exceed Then Mark as outlier Then you can Remove all outliers in the array to obtain the result. in, It is a threshold that is a multiple of the standard deviation.

5. The obstacle avoidance method for unmanned vehicles based on 3D point cloud depth detection according to claim 1, characterized in that, Step S4 includes: S4.1: Set the given number of target points ,if Number of point clouds > Then from Random downsampling If there are sampling points, Number of point clouds < Then from the time Repeat sampling until reached Each sampling point is used to obtain the sampled non-ground point cloud. ; S4.2: Calculation Geometric centroid of all sampling points coordinates Thus, to Any number in the middle sampling points coordinates Perform a translation to obtain the translated [number]. sampling points coordinates ; S4.3: Calculation All translated sampling points are respectively mapped to the world coordinate system of the autonomous vehicle. origin of coordinates The distances are calculated, and the furthest distance is selected from them. Afterwards, Divide by Thus, the normalized first... sampling points ; S4.4: For the first sampling points After performing random rotation, random scaling, and random translation operations on the three-dimensional coordinates, the preprocessed first-order coordinates are obtained. This process generates a preprocessed non-ground point cloud from a number of non-ground points. .

6. The obstacle avoidance method for unmanned vehicles based on 3D point cloud depth detection according to claim 1, characterized in that, Step S5 includes: S5.1: Using the hierarchical feature extraction network PointNet++ to process preprocessed non-terrestrial point clouds Multi-scale feature learning and downsampling are performed to generate a set of seed points containing local geometric information and contextual semantic information. ; S5.2: with Any j-th seed point Using the center as the reference point, a radius neighborhood search algorithm is employed to determine... The set of K neighboring points within the neighborhood of . Features of its neighboring points ;in, express The k-th neighboring point within the neighborhood of ; express Features; S5.3: Calculate the j-th seed point using equation (1) query vector The key vector of the k-th neighboring point and the value vector of the k-th neighboring point : (1) In equation (1), express Features , , These are three linear transformation weight matrices to be learned; S5.4: Calculate the j-th seed point using equation (2) With the k-th neighbor point Attention weights between : (2) In equation (2), is the dimension of the vector; T denotes transpose; S5.5: Obtained using equation (3) Neighborhood enhancement features and will and After concatenation, a nonlinear transformation is performed to obtain the enhanced features of the j-th seed point. ; = (3) S5.6: Calculation right The offset is calculated to obtain K offsets, and the neighborhood point with the smallest offset is selected as the nearest neighbor. polling stations and to All polling stations Clustering is performed to obtain The clustering results and cluster centers are analyzed, and several center proposal points are randomly selected from the cluster centers. For each selected center proposal point... Find all the polling stations in its neighborhood. ; Based on the geometric and contextual semantic information of the seed point inherited from the voting point, The geometric semantic features are aggregated into each central proposal point. Comprehensive proposal features Each central proposal point and its corresponding comprehensive proposal features A binary tuple that merges into a single candidate target The 3D object detection results are obtained by predicting the bounding boxes of all candidate objects using a multilayer perceptron. The position and parameters.

7. The obstacle avoidance method for unmanned vehicles based on 3D point cloud depth detection according to claim 1, characterized in that, Step S6 includes: S6.1: Obtain the autonomous vehicle in the world coordinate system The pose parameters and inherent geometric dimensions are given below, the pose parameters including: center coordinates. and heading angle The inherent geometric dimensions include: the length of the autonomous vehicle. Width of driverless cars The height of driverless cars ; S6.2: Define a scaling factor based on the inherent geometric dimensions of the autonomous vehicle. Thus, the length of the virtual extended body can be obtained using equation (4). Width of the virtual extension and the height of the virtual extension Thus, a virtual extended body is constructed at time t. ,in, This represents the coordinates of the center point of the virtual extended body at time t. The heading angle of the virtual extended body at time t is: ， ， (4) S6.3: The virtual extension is updated in real time as the autonomous vehicle moves. In This ensures that it always moves in sync with the driverless car and maintains the same direction of travel as the driverless car.

8. The obstacle avoidance method for unmanned vehicles based on 3D point cloud depth detection according to claim 1, characterized in that, Step S7 includes: S7.1: Calculate the overlap region of the two in space at time t using equation (5). : (5) S7.2: If Then it represents the virtual extended body at time t. With 3D bounding box set If one or more of them overlap, S7.3 is executed; otherwise, the driverless vehicle is controlled to continue driving along the original path. S7.3: Control the unmanned vehicle to stop and wait within the set waiting time, and determine whether the position and parameters of the 3D bounding box in the overlapping area have changed during the waiting time; if they have changed, execute S7.4; otherwise, replan the path. S7.4: Based on the virtual extended body at time t Reassess the changed environmental information. If the value is greater than 0, and less than 0, it means that the original planned path is safe, and the driverless car continues to drive along the original planned path; otherwise, the path is replanned. S7.5: If the waiting time exceeds the threshold Then determine the 3D bounding box set. With a static bounding box, the autonomous vehicle replans its path.

9. An electronic device, comprising a memory and a processor, characterized in that, The memory is used to store a program that supports a processor in executing the method of any one of claims 1-8, the processor being configured to execute the program stored in the memory.

10. A computer-readable storage medium storing a computer program thereon, characterized in that, The computer program is executed by the processor to perform the steps of the method according to any one of claims 1-8.

Images