METHOD FOR PROCUREMENT OF HIGH-RESOLUTION MAPS FROM POINT CLOUDS

Description

technical field

[0001] The present invention relates to a method for generating high-resolution maps from point clouds.

[0002] The invention relates to the field of autonomous vehicle localization combining GPS / GNSS data, a high-definition map and movement data, in particular from an inertial measurement unit. Previous technique

[0003] An autonomous vehicle needs to know its position relative to its environment in order to navigate. The problem of localization is solved using GPS / GNSS. However, in the absence of GPS / GNSS, other methods are necessary to help a vehicle determine its position. A high-definition (HD) onboard map fills in the information gaps while providing a reliable 3D description of the vehicle's environment.

[0004] Map generation devices are in high demand for generating high-definition maps, but for covering vast areas, current solutions for creating 3D maps are tedious and time-consuming and do not reduce the volume of data compared to the input data of measurements, whereas object recognition in image data and LiDAR data has progressed in recent years.

[0005] Autonomous mapping systems incorporate a minimum of three types of data: geolocation, 3D spatial measurements, and 360° visualization with optical imagery. High-quality geolocation measurements are the foundation for creating accurate and reliable maps. To achieve this, LiDAR scanners provide dense, high-precision 3D point clouds that can represent landforms, while panoramic images provide additional information such as texture, text, and other visual details.

[0006] This data together helps create a dense and highly detailed 3D map. Mobile mapping systems generate more than 1 gigabyte of data per 50 meters, making data management and processing complex and requiring significant data storage platforms, powerful computing resources, and video processing capabilities for data processing, visualization, navigation, and manipulation.

[0007] Processing and analyzing mobile mapping data to create high-definition maps is a complex and time-consuming process for several reasons: (1) the heterogeneous and unstructured nature of 3D point clouds, which are a relatively recent technology, makes them difficult to analyze (2) 2D images when taken alone lack spatial information (3) most classification and pattern extraction methods require significant human analyst intervention and, (4) standalone software used to produce HD maps is inherently designed as vertical solutions.

[0008] Furthermore, point clouds are large and complex computer objects to process.

[0009] Document EP 3 346 449 A1 gives an example of a computer-implemented point cloud preprocessing method that subdivides the cloud into cells.

[0010] In the field of neural networks, the document MATURANA DANIEL ET AL: “Voxnet: A 3D Convolutional Neural Network for real-time object recognition”, 2015 IEEE / RSJ INTERNATIONAL CONFERENCE ON INTELLIGENT ROBOTS AND SYSTEMS (IROS), IEEE, September 28, 2015 (2015-09-28), pages 922-928, XP032831749, DOI: 10.1109 / IROS.2015.7353481 [accessed 2015-12-11] discusses a model developed for recognizing indoor objects and scenes. It is an architecture composed of a convolutional neural network with the objective of generalizing the problem of segmenting and classifying objects in an indoor scene or classifying an entire object by consuming an entire point cloud as input. Such a process cannot be directly applied to a point cloud intended for 3D mapping due to its volume being one hundred to one thousand times denser.

[0011] The document "Multi-level 3D CNN for Learning Multi-scale Spatial Features" by S. Ghadai, X. Lee, A. Balu, S. Sarkar, and A. Krishnamurthy, arXiv, cs.CV 1805.12254, 2018, discusses the use of multiple scales, but with the goal of processing an entire object at a coarse scale and refining details at the object's edges using voxels at a finer resolution to bring out more detail and texture. This process is not intended to recognize different objects in a complex scene, but rather to understand a single object at a global level. It allows for the distinction of two or more types of objects, but it processes objects isolated from their context, which is not suitable for creating 3D maps where objects must be linked to their environment. Technical problem

[0012] For the generation of vector maps, the use of point clouds is complex because such clouds present an extremely large volume of data and are therefore difficult to process.

[0013] Furthermore, one step that particularly slows down the map generation process is the extraction of objects and their transformation into geospatial vectors. Optimization of these operations is being sought.

[0014] This is particularly detrimental to the development of automated guidance systems for autonomous vehicles. Description of the invention

[0015] In light of this prior art, the technical problems caused by point clouds in mapping, and the need to connect objects to their context, another approach was required. The present invention primarily uses a method based on image processing techniques, in other words, computer vision algorithms, to separate ground and non-ground elements. Indeed, in the present invention, the ground segmentation algorithm is a computer vision algorithm inspired by image processing, and not a neural network-based algorithm.

[0016] The present invention thus proposes an automated and generalized vector extraction method that operates in a 3D space, and which uses computer vision techniques and artificial intelligence techniques with convolutional neural networks and deep learning to perform a transformation of 2D and 3D objects into geospatial map vectors.

[0017] More specifically, the present invention proposes a method for generating high-definition 3D vector maps according to claim 1.

[0018] The process according to the invention, which combines several computer processing techniques, is optimized according to the type and volume of data to be processed.

[0019] Advantageously, the decimetric voxelization of the point cloud uses cubic voxels with a width of 20cm to 30cm.

[0020] The use of a computer vision algorithm at this stage is efficient while reducing the computational load on computer systems.

[0021] Preferably, the second and third classifications of object types are carried out on a computing system using voxelization, normalization and convolutional neural network classification algorithms.

[0022] Convolutional neural network systems and their associated deep learning technologies are suitable for processing more complex data.

[0023] The transformation of classified object types into unitary objects advantageously includes a method for extracting representative vector data of the unitary objects from the classified point cloud, said vector data extraction method comprising, from a selection of voxels, a graph construction algorithm including nodes of interest class and boundary class and an algorithm for extracting nodes constituting a vector representation of the objects.

[0024] The second classification of object types separating ground points from rail or track points preferably uses a centimeter voxelization algorithm with cubic voxels of a width of 2cm to 10cm prior to the execution of the convolutional neural network classification algorithm.

[0025] For non-ground points, the third classifications of object types advantageously include a plurality of voxelization, normalization, and classification operations by convolutional neural network according to iterative or parallel algorithms in order to classify object types such as vegetation object type, infrastructure object type, building object type, the set of voxelization, normalization and classification operations being followed by indexing operations of the classified point types for the generation of a classified point cloud.

[0026] The classification of objects buildings, vegetation and infrastructure advantageously uses three voxelization operations with cubic voxels of decametric, metric and decimetric size.

[0027] Using a multi-scale voxel structure allows the model to learn based on the type of object. This is an important implementation that reduces the load on computer systems and improves classification accuracy.

[0028] Vectorization operations may include for each object extraction algorithms which include the selection of voxels containing at least one point of the object to be extracted, a step of selecting adjacent voxels which contain at least one boundary point of the object, the construction of a graph with nodes and edges between nodes, said edges constructing the relationship between the boundary points and the points of the object, an ordering of the nodes of the graph with an assignment to the nodes of a value of increasing Euclidean distance from the boundary nodes, and an operation of extracting vector data representative of the object to be vectorized from the ordering carried out.

[0029] According to a first embodiment, the vector data extraction operation includes an object contour extraction.

[0030] According to a second alternative or complementary embodiment, the operation of extracting vector data from the object includes an extraction of skeleton data from the object.

[0031] According to a third alternative or complementary embodiment, the operation of extracting vector data from the object includes an extraction of the center and radius of the object.

[0032] According to one embodiment, said voxelization algorithm and vision algorithm of said first classification perform: a. a voxelization giving a voxelization data structure at a decimeter resolution for which a notion of voxel height is defined on the vertical axis of the volume, b. a detection of vertical or suspended objects which will determine voxel points classified as non-ground and candidate points of ground class, c. a connected component analysis giving an analysis of voxel connectivity according to which the largest connected component from this analysis is classified as ground.

[0033] Preferably, said vertical object detection includes vertical object segmentation.

[0034] This segmentation may include the detection of vertical objects by stacking non-empty voxels.

[0035] It can include detection of suspended objects by the presence of non-empty voxels without connectivity on the vertical axis.

[0036] The candidate points of ground class correspond to voxels that are not characterized as vertical or suspended objects.

[0037] Ground point detection may further include a statistical analysis on the distribution of minimum ground voxel elevations retaining voxels in a quartile lower than the third quartile of the minimum elevation distribution and eliminating other voxels as classified as non-ground.

[0038] Preferably, the point cloud is acquired using one or more acquisition vehicles traveling on roads to be used by said autonomous vehicles.

[0039] Point cloud acquisition is advantageously achieved using LIDAR devices mounted on at least one acquisition vehicle.

[0040] According to a particular embodiment, the LIDAR devices acquire data on two planes orthogonal at 45° to the axis of movement of said acquisition vehicle.

[0041] The invention further relates to a mapping device comprising one or more acquisition vehicles equipped with LIDAR devices, a point cloud dating and geolocation unit, one or more point cloud storage units and one or more processing units comprising means for calculating the classification and vectorization operations of the invention.

[0042] The invention also relates to a computer program comprising instructions for implementing the above process, when said instructions are executed by one or more processors of a computing system.

[0043] The invention further relates to a computerized guidance system for an autonomous vehicle comprising a 3D vectorized map obtained by the above process, wherein the 3D vectorized map is loaded into a computer of a guidance device of said autonomous vehicle and combined at the level of said computer with satellite positioning data of said vehicle, the vectorized data comprising data of roads, vegetation, infrastructure, buildings; and wherein the system comprises an algorithm for evaluating the reliability of satellite positioning signals using said vectorized data. Brief description of the drawings

[0044] Other features, details, and advantages of the invention will become apparent upon reading the detailed description, which describes examples of embodiments of the invention with reference to the accompanying drawings, which depict: [ Fig. 1] represents steps of a first part of an example of a process applicable to the invention; [ Fig. 2 ] represents steps of a second part of an example of a process applicable to the invention; [ Fig. 3A] ; [Fig. 3B ] ; ] Fig. 3C] ; [Fig. 3D ] represent steps in an example graph construction; [ Fig. 4 ] represents a greyscale graphical illustration of a node scheduling result; [ Fig. 5 ] represents an example of a soil segmentation algorithm applicable to the invention. Description of implementation methods

[0045] The drawings and description below relate to examples of implementation that may serve to better understand the present invention, but also to contribute to its definition, if necessary.

[0046] Object recognition in images using artificial intelligence is an established technology, and a new challenge is object recognition in point clouds for computing architectures that incorporate machine learning. When creating high-definition maps, point cloud data must be transformed into vector-based symbolic representations accompanied by rich object descriptions. In HD maps, data and their specifications are stored in XML formats such as OGC's GML or GeoJSON. Consequently, point cloud data, such as mobile mapping data, needs to be vectorized.

[0047] Vectorization is necessary because the aforementioned HD map formats are preferred as they are lightweight and can be stored on embedded computing devices. To date, the final stage of automated conversion of mobile mapping data into vector map formats has not yet been achieved. Current vectorization methods rely on semi-automatic extractions assisted by a human operator, explicitly guided or drawn using CAD applications. Data acquisition:

[0048] The mobile land mapping system used to collect data for this disclosure is a system using LIDAR-type devices mounted on at least one acquisition vehicle.

[0049] In particular, LiDAR devices can be configured to acquire data on two planes orthogonal to each other at 45° to the axis of movement of the acquisition vehicle. The device is complemented by a point cloud timing and geolocation unit, and one or more point cloud storage units.

[0050] It should be noted that due to the large amount of data obtained during the acquisitions, it is necessary to segment the paths to, for example, process point clouds of a reasonable volume, for example, on the order of 50 million points for a distance traveled of 100m to 500m.

[0051] An example of a usable system is the Road Scanner 4 (RS4) system from SITECO. The RS4 system is a topographic-class system that produces a dense, high-fidelity 3D point cloud.

[0052] Integrated into the localization system is a high-precision, dual-frequency GNSS antenna and a Landins ixBlue-C inertial measurement unit with an odometer. This hybrid geolocation optimizes the accuracy and coverage of geoposition information. The RS4 system is equipped with two 2D Zoller-Fröhlich (Z+F) scanners capable of rotating at 200 revolutions per second with 1 million points per second. The two Z+F lasers are positioned so that the laser scanners are orthogonal to minimize masking effects and maximize the scanning area. Object recognition in a 3D point cloud:

[0053] The invention includes a software organization which takes as input data the point clouds from the acquisitions.

[0054] According to this disclosure, the point cloud is processed in such a way as to classify the points of the cloud so that each point of a 3D cloud will be classified as notable objects or elements, including ground, infrastructure, vegetation, buildings and roads or rails.

[0055] An important aspect of this process is to use multiple voxelization steps with voxel dimensions adapted to the type of object to be extracted.

[0056] Furthermore, a problem with recognition systems is the processing time required; the method of the invention is optimized by using a vision algorithm (computer vision) for ground object type detection and one or more convolutional neural network algorithms for more complex object types to detect.

[0057] In this context, the figure 1 represents a classification process.

[0058] The first step consists of separating points representative of the ground from points other than the ground (for example, points aligned along vertical axes). This initial classification is performed using a first voxelization with a resolution of decimeter voxels, for example, on the order of 20 to 30 cm, by means of a vision algorithm. During this classification, the spatial coordinates of the points and their nature are associated with the points detected as ground or non-ground points.

[0059] An example of a vision or image processing algorithm adapted to ground segmentation is given in figure 5 for which soil segmentation consists of several steps starting from a voxelization. This algorithm, starting from the initial point cloud 50, uses a voxelization 70a, giving a voxelization data structure at a decimeter resolution, for which the voxels have the notion of height on the vertical axis of the volume.

[0060] Once the cloud is referenced by the voxel structure, a segmentation of vertical objects is performed. For this, according to the example, vertical objects are detected by the stacking of non-empty voxels, and suspended objects by the presence of non-empty voxels without connectivity on the vertical axis, at the level of vertical object detection.

[0061] This detection process will identify voxel points classified as non-ground (840) and candidate ground points (830) corresponding to voxels not characterized as vertical or suspended objects. Continuing with this example, to refine the detection, a connected component analysis (850, 860) is performed among these candidate ground voxels, providing an analysis of voxel connectivity. The largest connected component resulting from this analysis is classified as ground (880).

[0062] Next, to further refine the semantic segmentation of the soil, a statistical analysis 890 is then made on the distribution of minimum altitudes of all the soil voxels (output of connected component analysis) keeping the voxels which are of a quartile lower than the third quartile of the distribution of minimum altitudes 900, 920 the other voxels 910 being eliminated here also as classified as non-soil.

[0063] Such a multi-level filtering algorithm is, in the case of large point clouds, simpler and more efficient than a process using neural networks. Such an algorithm requires less development time, does not require training neural networks or training datasets.

[0064] .

[0065] For the ground points 102, a second voxelization step 110 is applied. This voxelization uses smaller cells or voxels, a few centimeters smaller, to classify rails and increase the resolution of the processed portion of the point cloud. This voxelization is followed by normalization 120, which makes the point coordinates easier to manipulate (in the initial point cloud, the points are located in absolute coordinates relative to the Earth's center and must be converted to more easily manipulated values). Once the data is normalized, a classification algorithm 130 using a convolutional neural network is applied to the voxelized cloud to distinguish the voxelized ground points from the voxelized rail or track points.

[0066] To use a neural network, a deep learning phase, a technique well-known in the field, will allow for the calibration of the parameters necessary for the process to function and for differentiating ground points from rail or track points. Based on the classification, a projection of the voxels onto the point cloud is applied to index the ground points (150) and rail points (160) within the cloud.

[0067] On the side of the points classified as non-ground 103, it is necessary to distinguish between types of objects of different volumes and shapes.

[0068] We seek to distinguish between infrastructure type objects, such as poles, catenaries, footbridges; building type objects and vegetation type objects.

[0069] The classification of objects buildings, vegetation and infrastructure described advantageously uses three voxelization operations with cubic voxels of dimensions adapted according to the type of object to be determined, for example dimensions: decametric, metric and decimetric with for example voxel dimensions determined by deep learning between 10m and 30m for decametric voxels, between 1m and 3m for metric voxels and between 20cm and 30cm for decimetric voxels.

[0070] At the end of the classification process, operations are performed to assign to the points in the cloud the object class to which they belong.

[0071] These operations include so-called voxel projection operations 140, 240, point indexing 150, 160, 250, classification update 300 which will resolve the types or classes of objects when points are found in more than one class by assigning to the point the most probable class among the classes resulting from the classification and a classification of the point cloud 310.

[0072] Once the point cloud has been classified, it can be stored in non-volatile memory for various uses, including the creation of a vector map as will be seen below. Object vectorization:

[0073] The computerized vectorization process uses the method described in figure 2and will allow the vectorization of classified objects such as rails, buildings, infrastructure elements, and vegetation to produce the autonomous navigation aid map. This method takes into account the membership of points in an interest class and the adjacent points that correspond to a boundary class.

[0074] To automatically extract an object of a classified type from a 3D point cloud, the points in the point cloud are analyzed to determine the relationships of a point with other surrounding points, as well as the distances between these points. This organization can be considered as a graph where the points constitute the nodes and the distances between a point and each of its neighbors constitute the edges.

[0075] In such a graph, the boundaries of an object will consist of all points outside a class, corresponding to a type of object, and which are direct neighbors of the points belonging to that class. These points constitute a boundary class.

[0076] First, the process involves a new voxelization (410) in which the voxel size takes into account the size of the objects to be extracted. From this voxelization, a selection (420) of voxels containing at least one point of a class of interest—that is, the class of the object to be extracted—is performed. Next, the process involves a selection of adjacent voxels that contain at least one point of a boundary class (430). The two sets of classes are analyzed to determine their proximity and distance to each other in the context of voxel construction.

[0077] The detailed vectorization operations in figure 2The process begins with a reading of the point cloud (400), followed by a new voxelization of the point cloud with voxels of a size appropriate to the objects to be extracted. For each object to be extracted, extraction algorithms are used, which include the selection (420) of voxels containing at least one point of the object to be extracted or class of interest, and a selection step (430) of voxels adjacent to the class of interest that contain at least one point of the boundary class of the class of interest. Then, an algorithm selects an arbitrary point from each selected voxel (440), this point becoming a so-called representative point.From the representative points a calculation algorithm will construct a graph with nodes made up of the representative points and edges connecting the adjacent voxel nodes 450, then the process includes a node ordering operation 460 including an iterative operation which will first assign a zero value to the nodes of the border class 600, 610 and then assign to each node adjacent to the border class node a value corresponding to the Euclidean distance of that node to the border node.

[0078] THE figures 3A to 3D illustrate simple graphs corresponding to these operations.

[0079] In figure 3A The points of the interest class 720 and the points of the adjacent voxels 710 identified in the selection steps 420, 430 and connected by edges 750 during the graph construction step 450 are represented. figure 3B, the nodes of border class 730 are identified around the nodes of class of interest. In figure 3C a node ordering is performed in which the border nodes 735 are of rank zero, the interest class nodes 725 contiguous to the border nodes of rank 1, the following nodes 740 of rank 2, the rank of the nodes increasing according to the distance from the border nodes.

[0080] In 3D figure , the Euclidean distance of the nodes to the boundary nodes is assigned to them in correspondence with their rank.

[0081] In the case of larger objects, higher-order nodes are present and processed.

[0082] There figure 4gives a greyscale example of a scheduling result for a road marking where the Euclidean distance 700 corresponding to the order of the nodes is transformed into colours according to the rainbow from blue 701 for an original distance of zero corresponding to the points of the border class to red 702 for the highest ranks in the original figure, passing through green 703 and yellow 704 for the intermediate values.

[0083] Several types of extraction are possible, depending in particular on the data format in which the map is created.

[0084] According to steps 490 and 495 illustrated in figure 2Vector data extraction involves extracting the object's outline. In this case, the useful nodes are the first-order nodes after the boundary nodes, and the associated algorithm is a search and selection algorithm for all nodes sharing an edge with a boundary class node. These are the first-order nodes on the figures 3C and 3D .

[0085] According to steps 500 and 505 of the flowchart of the figure 2 The operation to extract vector data from the object includes extracting skeleton data from the object. For this operation, the highest-order nodes are used.

[0086] This extraction is particularly useful for the vectorization of rails or road marking lines and includes the construction of a minimal tree according to an STP (Spanning tree protocol) algorithm, a minimum weight spanning tree algorithm, then for lines 515 the calculation of the longest path of the skeleton 525.

[0087] For point vectorizations of 520, we are looking for the node with the maximum overall number of radii of 530.

[0088] According to steps 540, 545 of the figure 2 Vectorization involves extracting the radius and center of an object 540 which uses a calculation of the distances from the highest rank nodes to the nearest zero-order node 545.

[0089] These different types of extraction can be combined or performed in parallel for maps of different types.

[0090] This point cloud vectorization algorithm operates in 2D and 3D spaces and allows the extraction of various geospatial vector shapes: points, lines, and polygons. The solutions chosen for the different classifications and voxelizations improve the quality of the point cloud classification, thus enhancing the point cloud vectorization algorithm. However, the extraction of broken lines requires an additional algorithm to determine the relationships between the classified markers of these lines.

[0091] Convolutional neural network systems and deep learning models used to segment and classify point clouds can be disseminated in cloud computing to increase computing resources horizontally (horizontal scaling) and enable the processing of the large volume of data typical of the mobile terrestrial mapping system.

Claims

1. A computer-implemented method for generating a high-definition 3D vector map for driving assistance for one or more autonomous vehicles, successively comprising: a. - one or more generations of at least one initial cloud of points (50) of a route to be mapped by means of one or more acquisition vehicles, b. - the generation (10) of a classified cloud of points (310) from said initial cloud of points, said generation comprising: i. - a first classification of object types: ground and non-ground from said initial cloud of points (50) comprising a decimetric voxelization algorithm (100) of the cloud of points using cubic voxels of a decimetric width and a computer vision algorithm (100) executed on a computing system to classify on the one hand ground points (102) relating to a horizontal plane defining the ground and on the other hand non-ground points (200), said voxelization algorithm and vision algorithm comprising:

1. a detection of vertical or suspended objects adapted to determine voxel points classified as non-ground (840) and candidate points of ground class (830) corresponding to the voxels which are not characterized as vertical or suspended objects, 2. among the candidate voxels corresponding to candidate points of ground class, a connected components analysis (850, 860) providing an analysis of the connectivity of the voxels, the largest connected component resulting from this analysis being classified as ground (880), a statistical analysis on the distribution of minimum altitudes of all the voxels classified as ground at the output of the connected components analysis, keeping the voxels that belong to a quartile lower than the third quartile of the distribution of minimum altitudes (900, 920), the other voxels (910) being eliminated as classified as non-ground, ii. - a second classification (130) of a road or rail object type (132) from the ground object type, iii. - a succession of third classifications (210, ..., 270) of vegetation, building, infrastructure object types from the non-ground object type; c. - from the classified cloud of points, the identification (20) of unitary objects from the classified object types and the vectorization of the unitary objects, d. - the storage of the vectorized data in non-volatile memory.

2. The method for generating a 3D vector map according to claim 1, wherein the decimetric voxelization algorithm (100) of the cloud of points uses cubic voxels with a width of 20cm to 30cm.

3. The method for generating a 3D vector map according to claim 1 or 2, wherein the second and third classifications of object types are carried out on a computing system from voxelization (110, 210), normalization (120, 220) and classification (130, 230) algorithms by a convolutional neural network.

4. The method for generating a 3D vector map according to any one of the preceding claims, wherein the transformation (20) of the classified object types into unitary objects comprises a method for extracting vector data representative of the unitary objects from the classified cloud of points, said method for extracting vector data comprising, from a selection of voxels (410, 420, 430, 440), a graph construction algorithm (450) comprising nodes of interest class and of border class and an algorithm for extracting constituent nodes of a vector representation of the objects.

5. The method for generating a 3D vector map according to any one of the preceding claims, wherein the second classification of object types separating ground points (131) of rail or road points (132) uses a centimetric voxelization algorithm (110) with cubic voxels of a width from 2cm to 10cm prior to the execution of the classification algorithm by convolutional neural network.

6. The method for generating a 3D vector map according to any one of the preceding claims, wherein, for the non-ground points (200), the third classifications of object types comprise a plurality of voxelization (210), normalization (220), classification (230) operations by convolutional neural network according to iterative or parallel algorithms so as to classify the object types such as vegetation object type, infrastructure object type, building object type, the set of voxelization (110, 210), normalization (120, 220) and classification (130, 230) operations being followed by indexing operations (150, 160, 250, 260) of the types of classified points for the generation of a classified cloud of points (310).

7. The method for generating a 3D vector map according to claim 6, wherein the classification of building, vegetation and infrastructure objects uses three voxelization operations with cubic voxels of increasing precision: decametric, metric and decimetric.

8. The method for generating a 3D vector map according to any one of the preceding claims, wherein the vectorization operations comprise for each object extraction algorithms which include the selection of voxels (420) comprising at least one point of the object to be extracted, a step (430) of selecting adjacent voxels which contain at least one border point of the object, the construction of a graph (450) with nodes and edges between nodes, said edges constructing the relationship between the border points and the points of the object, an ordering of the nodes of the graph with an assignment (600, ..., 640) to the nodes of an increasing Euclidean distance value from the border nodes (600, 610), and an extraction operation (490, 500, 540) of vector data representative of the object to be vectorized from the ordering carried out.

9. The method for generating a 3D vector map according to claim 8, wherein the operation of extracting vector data comprises an extraction (490) of a contour of the object.

10. The method for generating a 3D vector map according to claim 8 or 9, wherein the operation of extracting vector data of the object comprises an extraction (500) of skeleton data of the object.

11. The method for generating a 3D vector map according to claim 8, 9 or 10, wherein the operation of extracting vector data of the object comprises an extraction (540) of a center and radius of the object.

12. The method for generating a 3D vector map according to any one of the preceding claims, wherein said voxelization algorithm and vision algorithm of said first classification perform: a. a voxelization (70a) providing a voxelization data structure at a decimetric resolution for which a notion of voxel height is defined on the vertical axis of the volume, b. said detection of vertical or suspended objects (820) which will determine voxel points classified as non-ground (840) and candidate points of ground class (830), c. a connected components analysis (850, 860) providing an analysis of the connectivity of the voxels according to which the largest connected component resulting from this analysis is classified as ground (880).

13. The method for generating a 3D vector map according to claim 12, wherein said detection of vertical objects comprises a segmentation (810) of the vertical objects comprising a detection of the vertical objects by the stacking of non-empty voxels and of suspended objects by the presence of non-empty voxels without connectivity on the vertical axis, the candidate points of ground class corresponding to the voxels which are not characterized as vertical or suspended objects.

14. The method for generating a 3D vector map according to claim 12 or 13, comprising a statistical analysis (890) on the distribution of minimum altitudes of ground voxels, conserving the voxels that belong to a quartile lower than the third quartile of the distribution of minimum altitudes (900, 920) and eliminating the other voxels (910) as classified as non-ground.

15. The method for generating a 3D vector map according to any one of the preceding claims, wherein the cloud of points is acquired by means of one or more acquisition vehicles traveling on lanes to be taken by said autonomous vehicles.

16. The method for generating a 3D vector map according to claim 15, wherein the acquisition of the cloud of points is carried out by means of LIDAR devices mounted on at least one acquisition vehicle.

17. The method for generating a 3D vector map according to claim 16, wherein the LIDAR devices perform a data acquisition on two planes orthogonal at 45° to the axis of displacement of said acquisition vehicle.

18. A mapping device comprising one or more acquisition vehicles provided with LIDAR devices, a dating and geolocation unit for the cloud of points, one or more storage units for the cloud of points and one or more processing units comprising calculation means for carrying out the classification and vectorization operations of the method for generating a 3D vector map according to claims 1 to 17.

19. A computer program comprising instructions for implementing the method for generating a 3D vector map according to one of claims 1 to 17, when said instructions are executed by one or more processors of a computing system.

20. A computerized guidance system for an autonomous vehicle comprising a vectorized 3D map obtained by the method for generating a 3D vector map of any one of claims 1 to 17 for which the vectorized 3D map is loaded into a computer of a guidance device of said autonomous vehicle and combined at the level of said computer with satellite positioning data of said vehicle, the vectorized data comprising data on roads, vegetation, infrastructure, buildings; and for which the system comprises an algorithm for assessing the reliability of the satellite positioning signals using said vectorized data.