Method for detecting linear objects in a railway network, computer program and associated device

Abstract

The invention relates to a method for detecting linear objects in a railway network, comprising: for each point in a point cloud, calculation (22) of at least one corresponding geometric descriptor; assignment (24) of each point to a class, based on the corresponding geometric descriptors; detection (28) of planes defined by all or part of the points of a class of interest; consolidation (30) of the class of interest, including the exclusion of each point not belonging to a detected plane; agglomeration (32) of the points of the consolidated class of interest into disjoint sets of points, each representative of a respective linear object; determination (34) of the height of each linear object from the coordinate sets of the respective points; and association (38), to a predetermined target category, of each linear object whose height belongs to a corresponding predetermined height interval. Figure 1

Description

Title of the invention: Method for detecting linear objects on a railway network, computer program and associated device. Technical field

[0001] The present invention relates to a method for detecting linear objects in a railway network.

[0002] The invention also relates to a computer program and a device implementing such a method.

[0003] The invention applies to the field of control of structures, in particular linear objects, located on a railway network. State of the art

[0004] Fences are an element of railway line safety. More specifically, the installation of fences along railway tracks is an essential safety measure to protect passengers. Indeed, these fences are designed to prevent large animals (such as wild boar, roe deer, and red deer) from entering the tracks. Consequently, by limiting the risk of collisions between trains and wildlife, fences contribute to significantly reducing accidents, which is beneficial for passenger safety and the integrity of the railway service. Their presence also reduces the risk of human intrusion onto the tracks.

[0005] In general, the fences put in place by the railway network manager follow a precise set of specifications with regard, in particular, to their height, length and mesh.

[0006] However, private individuals and businesses located near railway lines may also install fences along said railway lines. In these cases, the characteristics of such fences are not guaranteed.

[0007] In addition, some fences are sensitive to weather events and can be knocked down or even uprooted by strong winds. Furthermore, they may need to be modified during occasional work.

[0008] Therefore, in the context of the monitoring and security of the tracks, having a thorough knowledge of the fence network is essential.

[0009] To assess the condition of the fence network, it has been proposed to carry out a fence inspection using field observation surveys, or from the observation of photos.

[0010] However, such a method does not give complete satisfaction.

[0011] Indeed, such a method is tedious and time-consuming. Furthermore, it requires the mobilization of significant human resources, both for carrying out observational surveys and for analyzing photos.

[0012] One object of the present invention is to remedy at least one of the drawbacks of the prior art.

[0013] Another object of the invention is to provide a method for detecting linear objects that is automatic and reliable. Description of the invention

[0014] To this end, the invention relates to a detection method of the aforementioned type, the method being implemented by computer and comprising the execution of a process including the steps: • from a cloud of points representing surfaces of objects present in a predetermined observation volume around a section of a railway line of the railway network, each point being associated with a respective set of coordinates in a predetermined three-dimensional frame, calculation, for each current point of the point cloud, of at least one corresponding geometric descriptor, from the sets of coordinates of the points located in a predetermined volume of interest around the current point; • for each point in the point cloud, assignment of said point, according to each corresponding calculated geometric descriptor, to a class among at least: a class of interest, representative of points likely to belong to a linear object, and a class of unassigned points; • from the coordinate sets of the points of the class of interest, detection of planes defined by all or part of said points; • consolidation of the interest class, including, for each point of the interest class not belonging to a detected plane, an exclusion of said point from said class; • from the coordinate sets of the points of the consolidated class of interest, agglomeration of said points into disjoint sets of points, each set being representative of a respective linear object; • for each linear object, determination of the corresponding height from the coordinate sets of the points in the respective set of points; and • association, to a predetermined target category, of each linear object whose determined height belongs to a predetermined height interval corresponding to the target category.

[0015] Indeed, by calculating the geometric descriptors of each point, information relating to the local geometry (shape, texture, etc.) in the vicinity of said point can be obtained. In this way, through the classification step, points whose local geometry resembles that of a point belonging to a linear object are identified. Thus, an automated implementation of the invention is possible.

[0016] Then, the implementation of the consolidation and filtering steps allows for the elimination of false positives, that is, points wrongly classified as belonging to a linear object. This results in a high reliability of the process according to the invention.

[0017] Furthermore, thanks to the agglomeration step, the points belonging to the same linear object are associated with each other.

[0018] Finally, each linear object is identified as belonging or not to a target category, based on its height. This also promotes the fully automated nature of the process according to the invention.

[0019] Advantageously, the method according to the invention has one or more of the following characteristics, taken individually or in any technically feasible combination:

[0020] The target category includes fences, railway platforms, bridge arches, level crossing barriers, trains on track, houses and urban infrastructure and / or retaining walls;

[0021] for each point, the at least one corresponding geometric descriptor includes, for at least one axis of the three-dimensional frame: a linearity, a flatness, a dispersion, a verticality, an omnivariance and / or a curvature;

[0022] the process further includes, between the assignment step and the detection step, a regularization step comprising, for each current point, a modification of the class of said point according to a majority class among the classes of a predetermined number of points neighboring said current point;

[0023] The agglomeration step comprises, for each first point among the points of the consolidated interest class, the following sub-steps: • determining, among the points of the consolidated interest class, the number of second points distinct from the first point and located at a distance less than a predetermined distance from the first point; and • if the determined number of second points is greater than or equal to a predetermined threshold, association of the first point and said second points to the same linear object;

[0024] the method further comprises, for each linear object, a determination of corresponding limits along a horizontal axis orthogonal to the direction of the gravitational field and extending in the detected plane associated with said linear object;

[0025] each set of points obtained at the end of the agglomeration step is considered to be representative of a respective linear object if the number of respective points is within a predetermined range;

[0026] The method further comprises a filtering step involving the provision of each set of points obtained at the end of the agglomeration step as input to an artificial intelligence model previously trained on the basis of a training dataset comprising a plurality of point clouds, each point cloud being associated with a label indicating whether said point cloud is representative or not of a linear object, said set of points being considered as representative or not of a linear object according to a corresponding output of the artificial intelligence model;

[0027] The process further comprises, after the agglomeration step: • for each observation volume, an estimate, for each linear object, of a corresponding bounding box; and • for two distinct observation volumes, a fusion of point sets corresponding to linear objects whose bounding boxes overlap at least partially or are adjacent.

[0028] According to another aspect of the invention, a computer program is proposed comprising executable instructions which, when executed by computer, implement the steps of the process as defined above.

[0029] The computer program can be in any computer language, such as for example in machine language, in C, C++, JAVA, Python, etc.

[0030] According to another aspect of the invention, a computer device is proposed for the detection of linear objects on a railway network, configured to: • from a cloud of points representing surfaces of objects present in a predetermined observation volume around a section of a railway line of the railway network, each point being associated with a respective set of coordinates in a predetermined three-dimensional frame, calculate, for each current point of the point cloud, at least one corresponding geometric descriptor, from the sets of coordinates of the points located in a predetermined volume of interest around the current point; • for each point in the point cloud, assign said point, according to each corresponding calculated geometric descriptor, to a class among at least: a class of interest, representative of points likely to belong to a linear object, and a class of unassigned points; • from the coordinate sets of the points of the class of interest, detect planes defined by all or part of said points; • consolidate the class of interest, the consolidation including, for each point of the class of interest not belonging to a detected plane, an exclusion of said point from said class; • from the coordinate sets of the points of the consolidated class of interest, aggregate said points into disjoint sets of points, each set being representative of a respective linear object; • For each linear object, determine the corresponding height from the coordinate sets of the points in the respective set of points; and • associate, to a predetermined target category, each linear object whose determined height belongs to a predetermined height interval corresponding to the target category.

[0031] The device according to the invention can be any type of device such as a server, a computer, a tablet, a calculator, a processor, a computer chip, programmed to implement the method according to the invention, for example by executing the computer program according to the invention. Brief description of the figures

[0032] The invention will be better understood upon reading the following description, given solely by way of non-limiting example and made with reference to the accompanying drawings in which:

[0033] [Fig.1] is a flowchart of a process according to the invention;

[0034] [Fig. 2] is a schematic representation of a device implementing the method of [Fig. 1]; and

[0035] [Fig.3] is a schematic representation of an orthogonal three-dimensional frame used during the implementation of the process of [Fig.1].

[0036] It is understood that the embodiments described below are in no way limiting. In particular, variants of the invention may be conceived comprising only a selection of the features described below, isolated from the other features described, if this selection of features is sufficient to confer a technical advantage or to differentiate the invention from the prior art. This selection includes at least one preferably functional feature without structural details, or with only a portion of the structural details if this portion alone is sufficient to confer a technical advantage or to differentiate the invention from the prior art.

[0037] In particular, all the variants and all the embodiments described are combinable with each other if nothing prevents this combination from a technical point of view.

[0038] In the figures and in the rest of the description, elements common to several figures retain the same reference. Detailed description

[0039] A method 2 for detecting linear objects in a railway network is illustrated by [Fig. 1].

[0040] By "linear object", it is understood, for the purposes of the present invention, to be an object having a flat surface and / or extending in a plane, and arranged in the vicinity of a railway, that is to say at a distance less than a predetermined limit in relation to the railway.

[0041] Such a linear object is, for example, a fence, a railway platform, a bridge arch, a level crossing barrier, a train on a railway track, a house, urban infrastructure, or a retaining wall.

[0042] The detection method 2 is capable of being implemented by any computer device 4 ([Fig.2]) comprising a memory 6 and a processor 8 connected together.

[0043] More specifically, the memory 6 is configured to store a detection program 10 according to the invention. Furthermore, the processor 8 is configured to execute instructions from the detection program 10 in order to implement the detection method 2.

[0044] Memory 6 is further configured to store, in a corresponding memory location 12, a point cloud.

[0045] By "point cloud", it is understood, in the context of the present invention, as a set of points 14 ([Fig.3]), each associated with a respective set of coordinates in a predetermined three-dimensional frame 16.

[0046] In particular, the points 14 of the point cloud stored in memory location 12 are representative of object surfaces (in particular solid objects) present in a predetermined observation volume around a section of a railway track of the railway network.

[0047] Such a section may correspond to the entire railway line, or to one of a plurality of consecutive (possibly adjacent) sections of said railway line.

[0048] Preferably, points 14 are representative of solid surfaces seen radially by a moving object moving along the railway track.

[0049] For example, the point cloud was acquired using a telemetry device, such as a LiDAR (Light Detection and Ranging) device mounted on a railway vehicle traveling along the track. The point cloud may also have been acquired by other types of static or dynamic systems.

[0050] Furthermore, each point 14 of the point cloud is associated with a unique identifier. Such an identifier is preferably also indicative of the segment to which the volume in which said point 14 was detected is associated.

[0051] Preferably, the three-dimensional frame 16 has a vertical axis (Oz) oriented parallel to the direction of the gravitational field. Furthermore, the vertical axis (Oz) is orthogonal to a horizontal plane (xOy) comprising the other two axes (Ox) and (Oy) of the three-dimensional frame 16.

[0052] As previously stated, the computer device 4 is configured to implement the detection method 2.

[0053] More specifically, as illustrated by [Fig. 1], the detection process 2 comprises a calculation step 22, an assignment step 24, a detection step 28, a consolidation step 30, an agglomeration step 32, a determination step 34 and an association step 38.

[0054] Advantageously, the detection method 2 further includes an optional regularization step 26, between the assignment step 24 and the detection step 28.

[0055] Advantageously, the detection method 2 further includes an optional filtering step 36, subsequent to the agglomeration step 32.

[0056] Advantageously, the detection method 2 further comprises an optional fusion step 40, subsequent to the agglomeration step 32. In the case where the filtering step 36 is implemented, the fusion step 40 is carried out after said fusion step.

[0057] Calculation step 22

[0058] The computer device 4 is configured to calculate, during the calculation step 22, for each point 14 of the point cloud, at least one corresponding geometric descriptor.

[0059] More specifically, for each current point 14, the computer device 4 is configured to calculate the value of each corresponding geometric descriptor from the coordinate sets of the points 14 located in a predetermined volume of interest around said current point 14.

[0060] Preferably, for each point 14, at least one corresponding geometric descriptor includes linearity, flatness, dispersion, verticality, omnivariance, and / or curvature. Even more preferably, the computer device 4 is configured to calculate each geometric descriptor for at least one axis of the three-dimensional coordinate system 16. This results in up to eighteen geometric descriptors per point 14.

[0061] In particular, for each current point 14, the computer device 4 is configured to calculate the value of each geometric descriptor corresponding to starting from the covariance matrix extracted from current point 14 as a function of the coordinate set of points 14 located in the predetermined volume of interest around current point 14.

[0062] Preferably, each geometric descriptor is associated with a corresponding volume of interest, which may be different from the volume of interest associated with another geometric descriptor. For example, depending on the geometric descriptor chosen, the volume of interest is a sphere with a radius of 15 cm, 25 cm, or 40 cm.

[0063] The calculation of such geometric descriptors is described in the publication: Weinmann, M., Jutzi, B., & Mallet, C. (2013), “Feature relevance assessment for the semantic interpretation of 3D point cloud data”, ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences, 2, 313-318.

[0064] Assignment step 2 4

[0065] The computer device 4 is also configured to assign, during the assignment step 24, each point 14 of the point cloud to a corresponding class.

[0066] More specifically, the computer device 4 is configured to assign each point 14 to the corresponding class according to each associated geometric descriptor, obtained at the end of the calculation step 22.

[0067] More specifically, the computer device 4 is configured to assign each point 14 to one of at least: a class of interest, representative of points likely to belong to a linear object, or a class of unassigned points.

[0068] Preferably, at least one class includes, in addition, a "soil" class, representative of points likely to belong to a soil (on which rests, in particular, the or each linear object), and / or a "vegetation" class, representative of points likely to belong to plants present in the vicinity of the railway.

[0069] Obviously, each point 14 is assigned to only one class.

[0070] For example, the computer device 4 is configured to carry out such an assignment by implementing a first classification model, in particular a first supervised classification model, such as a random forest.

[0071] In this case, the first classification model has been previously trained on the basis of a first training dataset comprising a plurality of points, each associated with corresponding geometric descriptors. Furthermore, each point in the first training dataset is associated with a label indicating the class to which that point belongs. In this case, the label associated with each point forms the expected output of the first classification model for that point and the corresponding geometric descriptors as inputs.

[0072] Regularization step _ 26

[0073] Advantageously, the computer device 4 is configured so as to, for each current point 14, determine, during the regularization step 26, a majority class among the classes associated with a predetermined number N of points near said current point (N being a non-zero natural number).

[0074] In other words, for each current point 14, the computer device 4 is configured to identify the N neighboring points and to determine the majority class among the classes associated with said N neighboring points. Such a determination of the majority class is also likely to include the class of the current point.

[0075] Furthermore, for each point 14, the computer device 4 is configured to modify the class of said point 14 according to the corresponding determined majority class. In particular, the computer device 4 is configured to assign each point 14 to the corresponding determined majority class.

[0076] Such a step is advantageous, insofar as it leads to a local standardization of the classes, which is likely to correct erroneous outcomes of the assignment step 24.

[0077] Detection step 28

[0078] In addition, the computer device 4 is configured to detect, during the detection step 28, planes defined by all or part of the points of the class of interest, from the respective coordinate sets.

[0079] For example, in order to detect such plans, the computer device 4 is configured to implement the RANSAC algorithm (from the English "RANdom SAmpling Consensus"), which is known to a person skilled in the art.

[0080] Such a step exploits the fact that the linear objects to be detected are planar, or have planar external surfaces.

[0081] Consolidation step 30

[0082] In addition, the computer device 4 is configured to consolidate the class of interest, during the consolidation step 30.

[0083] More specifically, the computer device 4 is configured to exclude from the class of interest each point 14 not belonging to a plane detected at the end of the detection step 28.

[0084] By "point not belonging to a plane", it is understood, for the purposes of the present invention, to be a point located at a distance greater than a predetermined maximum distance.

[0085] This results in a consolidated interest class.

[0086] For example, each point excluded from the class of interest is assigned to the class of unassigned points.

[0087] Agglomeration stage 32

[0088] The computer device 4 is also configured to aggregate, during the aggregation step 32, the points of the consolidated interest class into sets disjoint sets of points, based on their respective coordinate sets. In this case, each set of points obtained represents a respective linear object.

[0089] Preferably, in order to carry out such an agglomeration, the computer device 4 is configured so as to determine, for each first point among the points 14 of the consolidated interest class, the number of second points, among the other points of the consolidated interest class, which are located at a distance less than a predetermined distance from the first point.

[0090] Furthermore, in this case, if the determined number of second points is greater than or equal to a predetermined threshold, the computer device 4 is configured to associate the first point and said second points with the same linear object.

[0091] For example, to achieve this agglomeration, the computer device 4 is configured to implement a data partitioning algorithm such as the DBSCAN algorithm (from the English "Density-Based Spatial Clustering of Applications with Noise"), known to those skilled in the art.

[0092] Such an algorithm aims to group geometrically close points into sets. To define proximity, the algorithm considers that the neighborhood of each first point is, for example, 20 cm (value of the predetermined distance in the example), and that it must, for example, be composed of at least 20 points (value of the predetermined threshold in the example) to continue propagating the set.

[0093] In this way, the 14 geometrically distant points are not included in the same set, and each set of points is associated with a respective linear object.

[0094] Advantageously, each set of points obtained at the end of the agglomeration step 32 is considered representative of a respective linear object only if the number of respective points is within a predetermined range. In this way, false positives are likely to be eliminated.

[0095] Such a characteristic is advantageous, insofar as it leads to the exclusion of objects whose dimensions would not be compatible with the expected dimensions of a linear object.

[0096] Determination step 34

[0097] In addition, the computer device 4 is configured so as to determine, for each linear object (i.e. for each set of points obtained at the end of the agglomeration step 32), during the determination step 34, the corresponding height.

[0098] More specifically, for each linear object, the computer device 4 is configured to determine the corresponding height from the coordinate sets of the points 14 of the respective set of points.

[0099] Such a height corresponds, for example, to the greatest vertical coordinate difference, that is to say along the vertical direction (Oz), between the points 14 of the set of points considered.

[0100] Alternatively, for a given linear object, the height is advantageously taken to be equal to the value associated with a predetermined quantile of the distribution of the vertical coordinates of the points of the corresponding set of points, for example equal to the value of the ninth decile.

[0101] Such a feature is advantageous, insofar as it allows additional filtering of classification and acquisition noise, and reduces the impact of shape effects and noise.

[0102] Preferably, the computer device 4 is configured so as to determine, for each linear object, corresponding limits along a horizontal axis (called "local axis") extending in the detected plane associated with said linear object.

[0103] To do this, the computer device 4 is, for example, configured to determine a local reference frame associated with each linear object.

[0104] In particular, such a local reference frame comprises two non-collinear axes (including the local axis) extending in the detected plane associated with the linear object, and a third axis orthogonal to the detected plane.

[0105] For example, for each linear object, the computer device 4 is configured to apply a principal component analysis to the set of corresponding points in order to obtain the rotation matrix between the three-dimensional frame 16 and the associated local reference frame.

[0106] Furthermore, the computer device 4 is advantageously configured to identify the two points having the minimum and maximum coordinates along the local axis. The two identified points materialize the limits of the linear object along the local axis.

[0107] Obviously, the determination of the height of a linear object can also be carried out in the local reference frame, from the coordinate sets of the points 14 of the respective set of points in said local reference frame.

[0108] The computer device 4 is also configured to convert, in the three-dimensional reference frame 16, the coordinates of the two identified points, as the start and end coordinates of the linear object.

[0109] Filtering step 36

[0110] The computer device 4 is also configured to filter, during the optional filtering step 36, the sets of points obtained at the end of the agglomeration step 32.

[0111] More specifically, the computer device 4 is configured to provide each set of input points to a previously artificial intelligence model trained on the basis of a second training dataset comprising a plurality of point clouds, each point cloud being associated with a label indicating whether said point cloud is representative or not of a linear object.

[0112] In this case, each set of points is considered representative or not of a linear object depending on a corresponding output from the artificial intelligence model. In other words, the computer device 4 is configured to exclude each set of points for which the output of the artificial intelligence model indicates that said set of points is not representative of a linear object.

[0113] For example, such an artificial intelligence model is a second classification model, in particular a neural network called “PointNet”.

[0114] Such a neural network is described by QI, Charles R., SU, Hao, MO, Kaichun, et al., in the publication "Pointnet: Deep learning on point sets for 3d classification and segmentation", Proceedings of the EEE conference on computer vision and pattern recognition, 2017, p. 652-660.

[0115] A PointNet neural network generally requires that the point sets supplied to it as input be of the same size. Therefore, the implementation of the second classification model is generally preceded by preprocessing including resampling and / or compression of each point set. For example, after such preprocessing, each resulting point set is contained within a cube with sides of one meter, and comprises at most 4048 points.

[0116] Association step 38

[0117] In addition, the computer device 4 is configured to, during the association step 38, associate with a predetermined target category each linear object whose determined height belongs to a predetermined height interval corresponding to the target category.

[0118] For example, the computer device 4 is configured to associate linear objects with a height between 1 m and 3.5 m with a representative target category of fences.

[0119] Merge step 40

[0120] Preferably, the computer device 4 is also configured to estimate, during the optional fusion step 40, for each observation volume, and for each linear object detected in said observation volume, a bounding box corresponding to said linear object.

[0121] In addition, the computer device 4 is configured so as to, for two separate observation volumes, merge the sets of points associated with linear objects whose bounding boxes overlap at least in part or are adjacent.

[0122] It follows that the merged point sets are representative of the same linear object.

[0123] Such a characteristic is advantageous, insofar as the same linear object is likely to extend through a plurality of successive observation volumes.

[0124] Operation

[0125] The operation of the computer device 4 will now be described.

[0126] During a preliminary configuration step, a point cloud is stored in memory location 12 of memory 6.

[0127] Then, during the calculation step 22, the computer device 4 calculates, for each point 14 of the point cloud, at least one corresponding geometric descriptor.

[0128] Then, during the assignment step 24, the computer device 4 assigns each point 14 of the point cloud to a corresponding class according to each associated geometric descriptor obtained at the end of the calculation step 22.

[0129] Preferably, during the regularization step 26, the computer device 4 determines, for each current point 14, a majority class from among the classes associated with the N neighboring points of said current point. Furthermore, for each point 14, the computer device 4 modifies the class of said point 14 according to the corresponding majority class determined.

[0130] Then, during the detection step 28, the computer device 4 detects the planes defined by all or part of the points of the class of interest, from the respective sets of coordinates.

[0131] Then, during the consolidation step 30, the computer system 4 consolidates the class of interest. More precisely, the computer system 4 excludes from the class of interest each point 14 not belonging to a plane detected at the end of the detection step 28. A consolidated class of interest is thus obtained.

[0132] Then, during the agglomeration step 32, the computer device 4 aggregates the points of the consolidated class of interest into disjoint sets of points, according to their respective coordinate sets. In this case, each set of points obtained is representative of a respective linear object.

[0133] Then, during the determination step 34, the computer device 4 determines, for each linear object, the corresponding height.

[0134] Preferably, during the optional filtering step 36, the computer device 4 filters the sets of points obtained at the end of the agglomeration step 32.

[0135] Then, during the association step 38, the computer device 4 associates with a predetermined target category each linear object whose determined height belongs to a predetermined height interval corresponding to the target category.

[0136] Preferably, during the optional merging step 40, the computing device 4 estimates, for each observation volume and for each linear object detected in said observation volume, a bounding box corresponding to said linear object. Furthermore, the computing device 4 merges the point sets associated with linear objects whose bounding boxes at least partially overlap or are adjacent.

[0137] Of course, the invention is not limited to the examples just described.

Claims

1. Demands Method (2) for detecting linear objects in a railway network, the method being implemented by computer and comprising the execution of a process including the steps: • from a point cloud (14) representing surfaces of objects present in a predetermined observation volume around a section of a railway line of the railway network, each point (14) being associated with a respective set of coordinates in a predetermined three-dimensional frame (16), calculation (22), for each current point (14) of the point cloud, of at least one corresponding geometric descriptor, from the sets of coordinates of the points located in a predetermined volume of interest around the current point (14); • for each point (14) of the point cloud, assignment (24) of said point, according to each corresponding calculated geometric descriptor, to a class among at least: a class of interest, representative of points likely to belong to a linear object, and a class of unassigned points; • from the coordinate sets of the points (14) of the class of interest, detection (28) of planes defined by all or part of said points; • consolidation (30) of the interest class, comprising, for each point (14) of the interest class not belonging to a detected plan, an exclusion of said point (14) from said class; • from the coordinate sets of the points of the consolidated interest class, agglomeration (32) of said points into disjoint sets of points, each set being representative of a respective linear object; • for each linear object, determination (34) of the corresponding height from the coordinate sets of the points in the respective set of points; and • association (38), to a predetermined target category, of each linear object whose determined height belongs to a predetermined height range corresponding to the target category.

2. A method according to claim 1, wherein the target category includes fences, railway platforms, bridge arches, level crossing barriers, trains on track, houses and urban infrastructure and / or retaining walls.

3. A method according to claim 1 or 2, wherein, for each point (14), at least one corresponding geometric descriptor comprises, for at least one axis of the three-dimensional frame (16): a linearity, a flatness, a dispersion, a verticality, an omnivariance and / or a curvature.

4. A method according to any one of claims 1 to 3, further comprising, between the assignment step (24) and the detection step (28), a regularization step (26) comprising, for each current point, a modification of the class of said point (14) according to a majority class among the classes of a predetermined number of points neighboring said current point.

5. A method according to any one of claims 1 to 4, wherein the agglomeration step (32) comprises, for each first point among the points of the consolidated interest class, the following substeps: • determining, among the points of the consolidated interest class, the number of second points distinct from the first point and located at a distance less than a predetermined distance from the first point; and • if the determined number of second points is greater than or equal to a predetermined threshold, associating the first point and said second points with the same linear object

6. A method according to any one of claims 1 to 5, further comprising, for each linear object, a determination (34) of corresponding limits along a horizontal axis orthogonal to the direction of the gravitational field and extending in the detected plane associated with said linear object.

7. A method according to any one of claims 1 to 6, wherein each set of points obtained at the end of the agglomeration step (32) is considered to be representative of a respective linear object if the number of respective points is within a predetermined range.

8. A method according to any one of claims 1 to 7, further comprising a filtering step (36) comprising supplying each set of points obtained at the end of the agglomeration step as input to an artificial intelligence model previously trained on the basis of a training dataset comprising a plurality of point clouds, each point cloud being associated with a label indicating whether said point cloud is representative or not of a linear object, said set of points being considered as representative or not of a linear object according to a corresponding output of the artificial intelligence model.

9. A method according to any one of claims 1 to 8, further comprising, after the agglomeration step (32): • for each observation volume, an estimation, for each linear object, of a corresponding bounding box; and • for two separate observation volumes, a merging (40) of the point sets corresponding to linear objects whose bounding boxes at least partially overlap or are adjacent.

10. A computer program comprising executable instructions which, when executed by a computer, implement the steps of the process according to any one of claims 1 to Q

11. y. Computer device (4) for detecting linear objects of a railway network, configured to: • from a point cloud (14) representing surfaces of objects present in a predetermined observation volume around a section of a railway track of the railway network, each point being associated with a respective set of coordinates in a predetermined three-dimensional frame, calculate (22), for each current point of the point cloud, at least one corresponding geometric descriptor, from the sets of coordinates of the points located in a predetermined volume of interest around the current point; for each point in the point cloud, assign (24) said point, according to each corresponding calculated geometric descriptor, to a class among at least: a class of interest, representative of points likely to belong to a linear object, and a class of unassigned points; from the coordinate sets of the points of the class of interest, detect (28) planes defined by all or part of said points; consolidate (30) the class of interest, the consolidation including, for each point of the class of interest not belonging to a detected plane, an exclusion of said point from said class; from the coordinate sets of the points of the consolidated class of interest, agglomerate (32) said points into disjoint sets of points, each set being representative of a respective linear object; for each linear object, determine (34) the corresponding height from the coordinate sets of the points in the respective set of points; and associate (38), to a predetermined target category, each linear object whose determined height belongs to a predetermined height interval corresponding to the target category.

Images