A complex road boundary detection method and device based on point cloud voxel processing

Abstract

The application discloses a complex road boundary detection method and device based on point cloud voxel processing and belongs to the field of road detection. The method comprises the following steps: acquiring point cloud image data of articles around a road vehicle driving area, and performing voxel segmentation; adding a fixed point in the voxel to expand the voxel features; inputting a single-layer full connection layer to obtain a high-dimensional feature vector matrix; performing global feature transformation on all point cloud high-dimensional feature vector matrices in the voxel; selecting the global features corresponding to the k nearest point clouds as the center to aggregate; taking different values of k to obtain voxel features at different scales, and performing fusion; performing regression on the fused features to obtain related information of the articles around the vehicle driving area; calculating the boundary position of the vehicle driving area to obtain the width of the vehicle driving area, and determining whether the vehicle driving area is passable. The complex road boundary is accurately detected.

Description

Technical Field

[0001] This invention belongs to the field of road detection, and more specifically, relates to a method and apparatus for detecting complex road boundaries based on point cloud voxel processing. Background Technology

[0002] Complex roads refer to roads with varied environments, such as urban non-arterial roads and rural streets. These roads lack lane lines and clear road boundaries, and are further affected by shadows, watermarks, and other factors, making it difficult to distinguish between road areas and non-road areas. The varied road types, complex environmental backgrounds, shadows, watermarks, and changing weather all pose challenges to the detection of vehicle driving area boundaries on complex roads, and are also one of the main research directions in current road recognition technology.

[0003] Patent CN110244321A discloses a method for detecting passable road areas based on 3D LiDAR. This method uses 3D LiDAR to acquire point cloud data; combines this with vehicle odometer information to eliminate motion distortion in the point cloud; based on the point cloud data after motion distortion elimination, data points above a certain height above the LiDAR are removed; the RANSAC algorithm is used for ground segmentation to distinguish between ground point clouds and obstacle point clouds; the distinguished obstacle point clouds are rasterized, and the data points closest to the vehicle in each grid are extracted. These data points combined are the boundary points of the passable area. However, this method only operates on the nearest data points, losing all other information; and it lacks feature detection, failing to provide classification information and motion direction, making it prone to false detections and incomplete information judgment.

[0004] Patent CN112395956A discloses a method and system for detecting passable areas in complex environments. This system constructs a point cloud deep learning detection subsystem for passable areas, including a road type recognition network and a passable area detection network and modules, directly connected through the underlying interface of the deep learning framework. Point cloud data collected by an onboard LiDAR is input into the point cloud deep learning detection subsystem for passable area detection. However, the entire system uses only a single loss function for its networks and modules, and this function is dependent on parameter settings, making the system difficult to apply to multiple scenarios and unable to effectively identify changing roads.

[0005] Patent CN114155509A discloses a method and system for extracting passable areas of a road surface. This system extracts the road surface from point cloud data and uses a deep learning semantic segmentation model to segment the passable areas of the road surface from the image data. However, this system relies on high-precision maps and vehicle posture information for auxiliary determination, and it cannot effectively identify complex roads with changing conditions (such as those under construction or with temporary parking).

[0006] Road vehicle driving area boundary detection is a fundamental functional unit in the architecture of intelligent driving technology. The development of this field has followed an evolution from rule-based geometric methods to data-driven deep learning methods. Rule-based geometric methods, traditional methods, mainly include those based on elevation / slope, grid and reflection intensity, and model fitting. Their core problem lies in their weak generalization ability and inability to understand the semantic information of complex roads (such as road barriers, sidewalks, and curbs), making them unsuitable for the requirements of intelligent driving. Data-driven deep learning methods are currently the mainstream approach, mainly including projection-based, point-based, and voxel-based methods. Their characteristics include the ability to automatically learn features from data, strong semantic understanding, and better generalization.

[0007] Among them, the voxel-based method processes point clouds in a regular 3D mesh. The point cloud is divided into a voxel mesh, features are extracted within each voxel, and then processed using a 3D CNN. The introduction of sparse convolution allows it to perform calculations only on non-empty voxels, greatly improving efficiency and making high-resolution 3D voxel networks possible. However, it suffers from problems such as quantization errors and high memory consumption at high resolutions.

[0008] To address the practical needs of intelligent driving in existing technologies, a vehicle driving area boundary detection solution that can adapt to complex roads is required to cope with actual road conditions such as road construction, illegal parking, pedestrians, and non-motorized vehicles, and to achieve a more lightweight, flexible, and adaptable detection method. Summary of the Invention

[0009] To address the shortcomings of related technologies, the present invention aims to provide a method and apparatus for detecting complex road boundaries based on point cloud voxel processing, which is intended to solve the problem that traditional autonomous driving recognition methods are unable to quickly and accurately determine whether a vehicle's driving area is passable in some complex road conditions.

[0010] To achieve the above objectives, the present invention provides a method for detecting complex road boundaries based on point cloud voxel processing, comprising:

[0011] S1. Obtain point cloud image data of objects around the road vehicle driving area through a lidar sensor, and perform voxel segmentation processing on the point cloud image data.

[0012] S2. Add a fixed point within the voxel, and make the distance from the fixed point included in any point cloud image data within the voxel;

[0013] S3. Input the feature-amplified point cloud image data into a single-layer fully connected mapping network function to obtain a high-dimensional feature vector matrix; perform global feature transformation on all point cloud high-dimensional feature vector matrices within the voxel through max pooling to obtain global features under the three-dimensional receptive field.

[0014] S4. Using the global feature as the center, calculate the Euclidean distance from the point cloud within the voxel to the global feature, sort them according to the distance, and select the global features corresponding to the k nearest point clouds for aggregation to obtain the voxel features at the corresponding scale. The value of k is a positive integer k1, k2, ..., k c ,k1 to k c All are different, resulting in c voxel features, which are then fused together. Regression is then performed on the fused features to obtain relevant information about items surrounding the vehicle's driving area. This relevant information includes: category information, size information, location information, and orientation information; where c is greater than or equal to 2.

[0015] S5. Based on the relevant information of the objects around the vehicle's driving area, calculate the boundary position of the vehicle's driving area, obtain the width of the vehicle's driving area, compare it with the minimum width required for driving vehicles, and determine whether the vehicle's driving area is passable.

[0016] Optionally, the point cloud image data is represented as a two-dimensional matrix Q, Q=N×M; where N is the number of points in the point cloud, and M corresponds to the three-dimensional coordinate information (x,y,z) and intensity information r of each point.

[0017] Optionally, voxel segmentation is performed on the point cloud image data, including:

[0018] The point cloud image data is discretized into W×H grids in the xy plane to obtain the corresponding voxels; where W is the number of grids in the x-axis direction, H is the number of grids in the y-axis direction, and each grid is a voxel P.

[0019] Optionally, the step of adding a fixed point within a voxel and making any point cloud image data within the voxel include the distance from the data to the fixed point includes:

[0020] Based on the feature vector (x, y, z, r) of the point cloud image data, the distance from the point cloud image data within a voxel to a fixed point of the voxel is added to obtain a 7-dimensional feature vector (x, y, z, r, x', y', z'); where the fixed point is the voxel center point or the center of the voxel's projection on the xy plane, (x, y, z) are the three-dimensional coordinate information of each point cloud, r is the intensity information, and (x', y', z') are the three-dimensional coordinate information of each point cloud to the fixed point.

[0021] Optionally, the step of selecting and aggregating the global features corresponding to the k nearest point clouds to obtain voxel features at the corresponding scale includes:

[0022] Depending on the different detection ranges of the lidar, k can be set to multiple values, k=i j; where the value of i is negatively correlated with the current detection range of the lidar, and the values ​​of j are 1, 2, 3, and 4 respectively.

[0023] Secondly, the present invention also provides a complex road boundary detection device based on point cloud voxel processing, comprising:

[0024] The voxel segmentation module is used to acquire point cloud image data of objects around the road vehicle driving area through a lidar sensor, and to perform voxel segmentation processing on the point cloud image data.

[0025] The feature amplification module is used to add fixed points within a voxel and to include the distance from the fixed point in any point cloud image data within the voxel.

[0026] The global feature processing module is used to input the feature-amplified point cloud image data into a single-layer fully connected mapping network function to obtain a high-dimensional feature vector matrix; and to perform global feature transformation on all point cloud high-dimensional feature vector matrices within the voxel through max pooling to obtain global features under the three-dimensional receptive field.

[0027] The feature regression module is used to calculate the Euclidean distance from the point cloud within a voxel to the global feature, centered on the global feature, and sort the points according to the distance. It then selects the global features corresponding to the k nearest point clouds and aggregates them to obtain the voxel features at the corresponding scale. The value of k is a positive integer k1, k2, ..., kk. c ,k1 to k c All are different, resulting in c voxel features, which are then fused together. Regression is then performed on the fused features to obtain relevant information about items surrounding the vehicle's driving area. This relevant information includes: category information, size information, location information, and orientation information; where c is greater than or equal to 2.

[0028] The judgment module is used to calculate the boundary position of the vehicle driving area based on relevant information of objects around the vehicle driving area, obtain the width of the vehicle driving area, compare it with the minimum width required for driving vehicles, and determine whether the vehicle driving area is passable.

[0029] Thirdly, the present invention also provides an electronic device, the electronic device comprising: a lidar, an adapter bracket, a visual terminal, at least one processor, and a memory communicatively connected to the processor;

[0030] The lidar is mounted on an adapter bracket, which is fixed in the direction of vehicle travel. The lidar is used to acquire point cloud data information of objects around the vehicle in the direction of travel.

[0031] The processor is installed inside the vehicle and connected to the lidar via a connecting cable for point cloud data processing.

[0032] The memory stores a computer program that can be executed by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the complex road boundary detection method based on point cloud voxel processing as described in any one of the first aspects;

[0033] The visual terminal is installed inside the vehicle and connected to the processor via a connecting cable, and is used to display the vehicle's driving area.

[0034] Fourthly, the present invention also provides a computer-readable storage medium storing computer instructions for causing a processor to execute and implement the complex road boundary detection method based on point cloud voxel processing as described in any one of the first aspects.

[0035] Compared with the prior art, the above-described technical solutions conceived in this invention can achieve the following beneficial effects:

[0036] 1. This invention provides a method for detecting complex road boundaries based on point cloud voxel processing. Utilizing the perception capabilities of general sensors, it performs real-time detection of objects around vehicles. By combining the characteristics of point cloud sparsity at different distances using LiDAR, a detection distance threshold is introduced. Depending on the detection distance, different numbers of nearest neighbor points are searched, optimizing voxel features and improving detection efficiency while reducing the computational requirements of the computing unit. This solution performs well in detecting vehicle driving area boundaries on complex roads that do not rely on lane lines, enhancing vehicle passability under various road conditions and facilitating the application and promotion of related technologies.

[0037] 2. This invention provides a method for detecting complex road boundaries based on point cloud voxel processing. According to the different detection distances of the lidar, different point cloud data are selected for feature aggregation processing, which reduces the hardware resource requirements while ensuring the ability to perceive the vehicle driving environment. Attached Figure Description

[0038] Figure 1 This is a flowchart illustrating a complex road boundary detection method based on point cloud voxel processing provided in an embodiment of the present invention.

[0039] Figure 2 This is a schematic diagram of the hardware system composition provided in an embodiment of the present invention.

[0040] Figure 3 This is a schematic diagram of voxelization feature amplification of point cloud data according to the present invention.

[0041] Figure 4 This is a point cloud data acquisition image before processing in this invention.

[0042] Figure 5This is a point cloud data segmentation and recognition data map after processing by the present invention. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0044] The following description, in conjunction with a preferred embodiment, illustrates the content involved in the above embodiments.

[0045] Example 1

[0046] like Figure 1 As shown, this invention provides a method for detecting complex road boundaries based on point cloud voxel processing, including:

[0047] S1. Obtain point cloud image data of objects around the road vehicle driving area through a lidar sensor, and perform voxel segmentation processing on the point cloud image data.

[0048] S2. Add a fixed point within the voxel, and make the distance from the fixed point included in any point cloud image data within the voxel;

[0049] S3. Input the feature-amplified point cloud image data into a single-layer fully connected mapping network function to obtain a high-dimensional feature vector matrix; perform global feature transformation on all point cloud high-dimensional feature vector matrices within the voxel through max pooling to obtain global features under the three-dimensional receptive field.

[0050] S4. Using the global feature as the center, calculate the Euclidean distance from the point cloud within the voxel to the global feature, sort them according to the distance, and select the global features corresponding to the k nearest point clouds for aggregation to obtain the voxel features at the corresponding scale. The value of k is a positive integer k1, k2, ..., k c ,k1 to k c All are different, resulting in c voxel features, which are then fused together. Regression is then performed on the fused features to obtain relevant information about items surrounding the vehicle's driving area. This relevant information includes: category information, size information, location information, and orientation information; where c is greater than or equal to 2.

[0051] S5. Based on the relevant information of the objects around the vehicle's driving area, calculate the boundary position of the vehicle's driving area, obtain the width of the vehicle's driving area, compare it with the minimum width required for driving vehicles, and determine whether the vehicle's driving area is passable.

[0052] like Figure 2 As shown, the implementation of the complex road boundary detection method based on point cloud voxel processing provided by this invention requires the use of a vehicle hardware system. Point cloud image information of the road conditions ahead is acquired by a sensor installed in front of the vehicle in the direction of travel. The point cloud data information is then processed by a computing unit, and the processed result, i.e., whether the area ahead is passable, is displayed by a visual terminal installed in the vehicle. The visual terminal should be installed within the driver's line of sight, and the scanning angle of the LiDAR is usually greater than 120 degrees, covering the driving area.

[0053] After acquiring point cloud data through sensors, the point cloud images are used to achieve rapid 3D modeling of objects near complex roads for clustering, classification and recognition. Deep learning training is used to enhance the reasoning and judgment of vehicle driving area boundaries in complex road conditions, providing effective detection of vehicle driving area boundaries in autonomous driving under complex road conditions.

[0054] Optionally, the point cloud image data is represented as a two-dimensional matrix Q, Q=N×M; where N is the number of points in the point cloud, and M corresponds to the three-dimensional coordinate information (x,y,z) and intensity information r of each point.

[0055] Two-dimensional matrix transformation facilitates adaptation to the tensor computation requirements of large datasets in deep learning.

[0056] Optionally, voxel segmentation is performed on the point cloud image data, including:

[0057] The point cloud image data is discretized into W×H grids in the xy plane to obtain the corresponding voxels; where W is the number of grids in the x-axis direction, H is the number of grids in the y-axis direction, and each grid is a voxel P.

[0058] By using rasterization, point cloud image data is standardized, making it easier to adapt to deep learning models for batch data processing.

[0059] Optionally, the step of adding a fixed point within a voxel and making any point cloud image data within the voxel include the distance from the data to the fixed point includes:

[0060] Based on the feature vector (x, y, z, r) of the point cloud image data, the distance from the point cloud image data within a voxel to a fixed point of the voxel is added to obtain a 7-dimensional feature vector (x, y, z, r, x', y', z'); where the fixed point is the voxel center point or the center of the voxel's projection on the xy plane, (x, y, z) are the three-dimensional coordinate information of each point cloud, r is the intensity information, and (x', y', z') are the three-dimensional coordinate information of each point cloud to the fixed point.

[0061] The acquired point cloud image data is represented as a two-dimensional matrix Q. The distance from a point cloud to a fixed point on its corresponding voxel is (x', y', z'). By increasing the distance from the point cloud image data within a voxel to a fixed point on its corresponding voxel, the four-dimensional feature vector of the point cloud image data is amplified to a seven-dimensional feature vector, represented as a 7-dimensional feature vector g: (x, y, z, r, x', y', z'). The amplified feature vector of each point cloud data is represented as: It includes the three-dimensional coordinates, intensity information, and distance from the point cloud to a fixed point of its corresponding voxel for each point cloud.

[0062] By amplifying the feature vectors, the geometric understanding and contextual information representation of point cloud image data can be enhanced.

[0063] Furthermore, the feature-amplified point cloud image data is clustered into voxels to extract voxel features; specifically, the point cloud data is mapped through a single-layer fully connected layer network function f to obtain a high-dimensional feature vector h. In this process, the 7-dimensional feature vector g of each point cloud is transformed into a feature vector h. Then, a max-pooling operation is performed to transform the global features, obtaining the global features within the 3D receptive field. The global features corresponding to the k nearest points are then aggregated to represent the voxel features at the corresponding scale, as follows: , where h is the high-dimensional feature vector; k is the number of nearest points; j is the index of different voxels; and max-pooling is the max pooling operation.

[0064] Based on the above scheme, optionally, the step of selecting the global features corresponding to the k nearest point clouds for aggregation to obtain voxel features at the corresponding scale includes:

[0065] Depending on the different detection ranges of the lidar, k can be set to multiple values, k=i j ; where the value of i is related to the current detection distance, and the values ​​of j are 1, 2, 3, and 4 respectively.

[0066] For example, for a lidar with a detection range of less than 10 meters, i is set to 3; for a lidar with a detection range of more than 10 meters, i is set to 2. When i is 2, k takes values ​​of 2, 4, 8, and 16 respectively; when i is 3, k takes values ​​of 3, 9, 27, and 81 respectively. Different values ​​are used depending on the current situation to obtain the features aggregated at the corresponding k points. The features with different k values ​​are then concatenated and input into a single-layer fully connected mapping network function f to integrate the feature information under different k values, resulting in the concatenated features. For example, for a lidar with a detection range of less than 10 meters, when i is 3, the resulting concatenated features are represented as follows: .

[0067] LiDAR includes horizontal and vertical scanning. The sparsity of the data points varies with the detection distance. The number of nearest points is controlled by the k value. The larger the k value, the more points correspond to a feature.

[0068] The features obtained above are then fused using feature learning to achieve feature fusion at different k values. Regression is then performed on the fused features at different k values ​​to obtain information such as the category, size, location, and orientation of the surrounding objects.

[0069] Finally, based on the relevant information of the surrounding objects, the width of the vehicle's driving area is obtained, and it is compared with the minimum width required for driving vehicles to determine whether it is passable; thus providing intelligent judgment guidance for the driver and assisting the driver in judging whether the current driving behavior needs to be adjusted.

[0070] like Figure 4 , Figure 5 As shown in the figure, this embodiment provides a method for detecting the boundary of a complex road vehicle driving area based on voxel segmentation of point cloud images, and demonstrates its detection effect. Figure 4 To obtain the display effect of point cloud image data, Figure 5 To process the input point cloud data using this solution, we can display the objects around the vehicle's driving area (as shown by the Car icon in the figure) and analyze the vehicle's driving area.

[0071] like Figure 5 As shown, this demonstrates the perception of road information in the vehicle's direction of travel, including the presence of walls or parked vehicles. Compared to existing technologies, this solution reduces the demand for hardware resources while maintaining a certain perception speed and accuracy.

[0072] This invention provides a method for detecting complex road boundaries based on point cloud voxel processing. Utilizing the perception capabilities of general sensors, it performs real-time detection of objects around the vehicle. By combining the point cloud sparsity patterns of LiDAR at different distances, a detection distance threshold is introduced. Depending on the detection distance, different numbers of nearest neighbor points are searched, optimizing voxel features and improving detection efficiency while reducing the computational requirements of the computing unit. This solves the technical problem of traditional autonomous driving recognition methods' inability to accurately determine the vehicle's driving area in complex urban and suburban road conditions. It can effectively detect the vehicle's driving area boundary in urban and suburban roads, completing the processing of complex roads. This solution has excellent results in detecting vehicle driving area boundaries on complex roads that do not rely on lane lines, enhancing the vehicle's ability to travel in different road conditions and facilitating the application and promotion of related technologies.

[0073] Example 2

[0074] The present invention also provides a complex road boundary detection device based on point cloud voxel processing, comprising:

[0075] The voxel segmentation module is used to acquire point cloud image data of objects around the road vehicle driving area through a lidar sensor, and to perform voxel segmentation processing on the point cloud image data.

[0076] The feature amplification module is used to add fixed points within a voxel and to include the distance from the fixed point in any point cloud image data within the voxel.

[0077] The global feature processing module is used to input the feature-amplified point cloud image data into a single-layer fully connected mapping network function to obtain a high-dimensional feature vector matrix; and to perform global feature transformation on all point cloud high-dimensional feature vector matrices within the voxel through max pooling to obtain global features under the three-dimensional receptive field.

[0078] The feature regression module is used to calculate the Euclidean distance from the point cloud within a voxel to the global feature, centered on the global feature, and sort the points according to the distance. It then selects the global features corresponding to the k nearest point clouds and aggregates them to obtain the voxel features at the corresponding scale. The value of k is a positive integer k1, k2, ..., kk. c ,k1 to k c All are different, resulting in c voxel features, which are then fused together. Regression is then performed on the fused features to obtain relevant information about items surrounding the vehicle's driving area. This relevant information includes: category information, size information, location information, and orientation information; where c is greater than or equal to 2.

[0079] The judgment module is used to calculate the boundary position of the vehicle driving area based on relevant information of objects around the vehicle driving area, obtain the width of the vehicle driving area, compare it with the minimum width required for driving vehicles, and determine whether the vehicle driving area is passable.

[0080] The present invention provides a complex road boundary detection device based on point cloud voxel processing, which is used to execute a complex road boundary detection method based on point cloud voxel processing in Embodiment 1, and has the same or similar beneficial effects.

[0081] Example 3

[0082] like Figure 2 As shown, the present invention also provides an electronic device, which includes: a lidar, an adapter bracket, a visual terminal, at least one processor, and a memory communicatively connected to the processor;

[0083] The lidar is mounted on an adapter bracket, which is fixed in the direction of vehicle travel. The lidar is used to acquire point cloud data information of objects around the vehicle in the direction of travel.

[0084] The processor is installed inside the vehicle and connected to the lidar via a connecting cable for point cloud data processing.

[0085] The memory stores a computer program that can be executed by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the complex road boundary detection method based on point cloud voxel processing as described in any one of Embodiments 1;

[0086] The visual terminal is installed inside the vehicle and connected to the processor via a connecting cable, and is used to display the vehicle's driving area.

[0087] The sensor includes a lidar, a mounting bracket, a connector, and a connecting cable.

[0088] The computing unit includes an AI computing unit, a general computing unit, a control unit, a security processing unit, etc.

[0089] The visual terminal includes a display screen, operation buttons, etc. The operation buttons are located within the driver's control range and are controlled by the operation buttons.

[0090] Example 4

[0091] The present invention also provides a computer-readable storage medium storing computer instructions for causing a processor to execute the complex road boundary detection method based on point cloud voxel processing as described in any one of Embodiments 1.

[0092] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for detecting complex road boundaries based on point cloud voxel processing, characterized in that, include: S1. Obtain point cloud image data of objects around the road vehicle driving area through a lidar sensor, and perform voxel segmentation processing on the point cloud image data. S2. Add a fixed point within the voxel, and make the distance from the fixed point included in any point cloud image data within the voxel; S3. Input the feature-amplified point cloud image data into a single-layer fully connected mapping network function to obtain a high-dimensional feature vector matrix; perform global feature transformation on all point cloud high-dimensional feature vector matrices within the voxel through max pooling to obtain global features under the three-dimensional receptive field. S4. Using the global feature as the center, calculate the Euclidean distance from the high-dimensional features of the point cloud within the voxel to the global feature, sort them according to the distance, and select the global features corresponding to the k nearest point clouds for aggregation to obtain the voxel features at the corresponding scale. The value of k is a positive integer k1, k2, ..., k c ,k1 to k c All are different, resulting in c voxel features, which are then fused together. Regression is then performed on the fused features to obtain relevant information about items surrounding the vehicle's driving area. This relevant information includes: category information, size information, location information, and orientation information; where c is greater than or equal to 2. S5. Based on the relevant information of the objects around the vehicle's driving area, calculate the boundary position of the vehicle's driving area, obtain the width of the vehicle's driving area, compare it with the minimum width required for driving vehicles, and determine whether the vehicle's driving area is passable.

2. The method as described in claim 1, characterized in that, The point cloud image data in step S1 is represented as a two-dimensional matrix Q, Q=N×M; where N is the number of points in the point cloud, and M corresponds to the three-dimensional coordinate information (x,y,z) and intensity information r of each point.

3. The method as described in claim 2, characterized in that, Voxel segmentation of the point cloud image data includes: The point cloud image data is discretized into W×H grids in the xy plane to obtain the corresponding voxels; where W is the number of grids in the x-axis direction, H is the number of grids in the y-axis direction, and each grid is a voxel P.

4. The method as described in claim 1, characterized in that, Increasing the distance from the point cloud image data within a voxel to a fixed point on the voxel expands the four-dimensional feature vector of the point cloud image data to a seven-dimensional feature vector, including: Based on the feature vector (x, y, z, r) of the point cloud image data, the distance from the point cloud image data within a voxel to a fixed point of the voxel is added to obtain a 7-dimensional feature vector (x, y, z, r, x', y', z'); where the fixed point is the voxel center point or the center of the voxel's projection on the xy plane, (x, y, z) are the three-dimensional coordinate information of each point cloud, r is the intensity information, and (x', y', z') are the three-dimensional coordinate information of each point cloud to the fixed point.

5. The method as described in claim 1, characterized in that, By taking different values ​​for k, features at different scales are obtained, including: Depending on the different detection ranges of the lidar, k can be set to multiple values, k=i j ; where the value of i is negatively correlated with the current detection range of the lidar, and the values ​​of j are 1, 2, 3, and 4 respectively.

6. A complex road boundary detection device based on point cloud voxel processing, characterized in that, include: The voxel segmentation module is used to acquire point cloud image data of objects around the road vehicle driving area through a lidar sensor, and to perform voxel segmentation processing on the point cloud image data. The feature amplification module is used to add fixed points within a voxel and to include the distance from the fixed point in any point cloud image data within the voxel. The global feature processing module is used to input the feature-amplified point cloud image data into a single-layer fully connected mapping network function to obtain a high-dimensional feature vector matrix; and to perform global feature transformation on all point cloud high-dimensional feature vector matrices within the voxel through max pooling to obtain global features under the three-dimensional receptive field. The feature regression module is used to calculate the Euclidean distance from the high-dimensional features of the point cloud within a voxel to the global feature, centered on the global feature. The features are then sorted by distance, and the global features corresponding to the k nearest point clouds are aggregated to obtain the voxel features at the corresponding scale. The value of k is a positive integer k1, k2, ..., k. c ,k1 to k c All are different, resulting in c voxel features, which are then fused together. Regression is then performed on the fused features to obtain relevant information about items surrounding the vehicle's driving area. This relevant information includes: category information, size information, location information, and orientation information; where c is greater than or equal to 2. The judgment module is used to calculate the boundary position of the vehicle driving area based on relevant information of objects around the vehicle driving area, obtain the width of the vehicle driving area, compare it with the minimum width required for driving vehicles, and determine whether the vehicle driving area is passable.

7. An electronic device, characterized in that, The electronic device includes: a lidar, an adapter bracket, a visual terminal, at least one processor, and a memory communicatively connected to the processor; The lidar is mounted on an adapter bracket, which is fixed in the direction of vehicle travel. The lidar is used to acquire point cloud data information of objects around the vehicle in the direction of travel. The processor is installed inside the vehicle and connected to the lidar via a connecting cable for point cloud data processing. The memory stores a computer program that can be executed by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the complex road boundary detection method based on point cloud voxel processing as described in any one of claims 1-5; The visual terminal is installed inside the vehicle and connected to the processor via a connecting cable, and is used to display the vehicle's driving area.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that cause a processor to execute the complex road boundary detection method based on point cloud voxel processing as described in any one of claims 1-5.

Images