Point cloud identification device, learning device, point cloud identification method and learning method

Abstract

A point cloud identification device (300) includes a point cloud acquisition unit (320) for acquiring point cloud information indicating N (N ≥ 2) points in k (k ≥ 2) dimensions, a model acquisition unit (310) for acquiring a model having a learning parameter, a rotation invariance conversion unit (330) for orthogonalizing each of basis vectors for each of points indicated in the point cloud information and calculating a rotation invariance feature using data after the orthogonalization, an inference unit (340) for identifying a point cloud indicated in the point cloud information using the rotation invariance feature and the model, and a result output unit (350) for outputting a classification result by identification by the inference unit.

Description

TECHNICAL FIELD

[0001] The present disclosure technology relates to a point cloud identification technology for identifying a point cloud indicated by point cloud information. BACKGROUND TO THE STATE OF THE ART

[0002] For example, a detection result of a sensor such as LiDAR or radar may be represented in the form of point cloud information containing a large number of points. There is a technique for extracting a feature of a point cloud represented by such point cloud information. Patent Literature 1 discloses a feature representation device that represents a feature of a point cloud. Specifically, a feature representation device of Patent Literature 1 is a feature representation device for the feature representation of three-dimensional point cloud data (point cloud information), and includes a distance field conversion unit that converts a set of points into a distance field representing the coordinates s X , s y , and s za spatial sampling point group s around the points, and a nearest neighbor distance ϕ(s) from the spatial sampling point s to a nearest neighbor point, a regular projection unit that receives a conversion to a standard coordinate system by performing a singular value decomposition of a matrix M containing the coordinates s x , s y , and s z of the spatial sampling point s and the nearest neighbor distance ϕ(s), and a parameterization unit that outputs a weight β as a feature vector of the three-dimensional point cloud data by training an expression learning machine that takes as input a coordinate L in of the spatial sample point s converted to the standard coordinate system and outputs the nearest neighbor distance ϕ(s).

[0003] The feature display device of Patent Literature 1 is configured to align, for example, a point cloud and a point cloud obtained by rotating the point cloud based on a nearest neighbor distance to a spatial sampling point group around the point cloud. REFERENCE LISTPATENT LITERATURE

[0004] Patent Literature 1: JP 2019-133545 A SUMMARY OF THE INVENTIONTECHNICAL PROBLEM

[0005] However, the feature display device of Patent Literature 1 has the problem that, depending on the point cloud, an error occurs at the time of positioning using the nearest neighbor distance between the point cloud and the spatial sampling point, and that a point cloud with the same shape may be identified as a point cloud with a different shape, and that the accuracy of identifying the point cloud tends to be low.

[0006] The present disclosure solves the above problem and an object is to improve the accuracy of identifying a point cloud. SOLUTION TO THE PROBLEM

[0007] A point cloud identification device according to the present disclosure comprises: a point cloud acquisition unit to acquire point cloud information indicating N (N ≥ 2) points in k (k ≥ 2) dimensions; a model acquisition unit to acquire a model that has a learning parameter; a rotation invariance conversion unit for performing orthogonalization of each of basis vectors for each of points specified in the point cloud information, and calculating a rotation invariance feature using data after the orthogonalization; an inference unit for identifying a point cloud specified in the point cloud information using the rotation invariance feature and the model; and a result output unit to output a classification result by identifying by the inference unit. ADVANTAGEOUS EFFECTS OF THE INVENTION

[0008] According to the present disclosure, it is possible to improve the accuracy of identifying a point cloud. BRIEF DESCRIPTION OF THE DRAWINGS Fig. 1 is a diagram illustrating an example of a configuration of a point cloud identification system including a point cloud identification device according to a first embodiment. Fig. 2 is a diagram showing an example of a configuration of a rotation invariance conversion unit in the point cloud identification device. Fig. 3 is a flowchart showing an example of processing by the point cloud identification device. Fig. 4 is a flowchart showing a concrete example of rotation invariance conversion processing in the point cloud identification device processing. Fig. Figure 5 is a diagram illustrating a rotation of a point cloud. Fig. Figure 6A is a diagram describing a relationship between the rotation of point clouds displaying similar shapes and subspaces. Fig. Figure 6B is a diagram describing a relationship between point clouds indicating different shapes and subspaces. Fig. 7 is a diagram illustrating an example of a configuration of a point cloud learning system 2 including a point cloud learning device according to a second embodiment. Fig. 8 is a flowchart showing an example of processing by the point cloud learning device. Fig. 9 is a diagram showing a first example of a hardware configuration for implementing functions of the point cloud identifying device or the point cloud learning device in the present disclosure. Fig. 10 is a diagram showing a second example of a hardware configuration for implementing the functions of the point cloud identification device or the point cloud learning device in the present disclosure. DESCRIPTION OF THE EMBODIMENTS

[0009] In order to further explain the present disclosure, embodiments of the present disclosure will be described below with reference to the accompanying drawings. First embodiment.

[0010] In a first embodiment, a form of point cloud identification device is explained.

[0011] Fig. 1 is a diagram illustrating an example of a configuration of a point cloud identification system 1 including a point cloud identification device 300 according to the first embodiment.

[0012] Fig. 2 is a diagram illustrating an example of a configuration of a rotation invariance conversion unit 330 in the point cloud identification device 300.

[0013] Fig. 3 is a flowchart illustrating an example of processing by the point cloud identification device 300.

[0014] Fig. 4 is a flowchart showing a concrete example of rotation invariance conversion processing in the processing of the point cloud identification device 300.

[0015] Fig. Figure 5 is a diagram showing a rotation of point clouds 1100 and 1200.

[0016] Fig. Figure 6A is a diagram explaining a relationship between the rotation of point clouds indicating the same shape and subspaces 2100 and 2200. Fig. Figure 6B is a diagram illustrating the relationship between point clouds indicating different shapes and subspaces 3100 and 3200.

[0017] An example of a configuration of the point cloud identification system 1 including the point cloud identification device 300 will be explained.

[0018] The point cloud identification system 1 is a system including a point cloud identification device 300 that identifies a point cloud.

[0019] The point cloud identification system 1, shown in Fig. 1, comprises a point cloud input device 100, a storage device 200, a point cloud identification device 300 and a result output device 900.

[0020] The point cloud input device 100 retrieves point cloud information (point cloud data) from, for example, a sensor not shown, and outputs the point cloud information to the point cloud identification device 300.

[0021] The sensor, which is not shown, is, for example, a LiDAR (Light Detection and Ranging) or a radar.

[0022] The point cloud information displays a plurality of (N ≥ 2) points representing position coordinates or features acquired by the sensor not shown.

[0023] The point cloud information represents each point contained in the point cloud, for example, in the form of a coordinate value in a k (k ≥ 2) dimension.

[0024] The storage device 200 comprises a storage unit 210.

[0025] The storage unit 210 contains information used for identification processing to identify a point cloud. Specifically, the storage unit 210 contains, for example, a learning parameter as a model for identification.

[0026] The point cloud identification device 300 identifies the point cloud using a learned model in the k-dimensional point cloud identification. The learned model includes, for example, a model that was learned accordingly as in a second embodiment.

[0027] K-dimensional point cloud identification is used to perform class identification of point cloud data that specifies a point cloud represented in k dimensions, based on a feature such as the shape of the point cloud.

[0028] A configuration example of the point cloud identification device 300 will be explained.

[0029] The point cloud identification device 300 includes a model acquisition unit 310, a point cloud acquisition unit 320, a rotation invariance conversion unit 330, an inference unit 340, and a result output unit 350.

[0030] The model acquisition unit 310 acquires a model having a learning parameter.

[0031] The learning parameter is a parameter learned by the learning facility and used in the identification of the point cloud.

[0032] The point cloud acquisition unit 320 acquires k-dimensional point cloud information of N points.

[0033] Specifically, the point cloud acquisition unit 320 acquires point cloud information indicating N (N ≥ 2) points in k (k ≥ 2) dimensions.

[0034] The rotation invariance conversion unit 330 performs orthogonalization of each of basis vectors for each of points indicated in the point cloud information, and calculates a rotation invariance feature using data after the orthogonalization.

[0035] The rotation invariance feature is a feature that is unique to the shape of the point cloud, which is not affected even when the point cloud is rotated.

[0036] The inference unit 340 uses the rotation invariance feature and the model to identify the point cloud specified in the point cloud information.

[0037] Specifically, the inference unit 340 calculates a point cloud feature indicating a feature of the point cloud using the rotation invariance feature and the model, and identifies the point cloud based on the point cloud feature. The point cloud feature is a unique feature for each category of the point cloud and denotes or indicates a feature effective for classifying the point cloud into classes. The point cloud feature is a feature that can be identified as a category in a point cloud of a horse and as a category in a point cloud of a bird, for example, when the point cloud feature is identified by an animal species.

[0038] Furthermore, the inference unit 340 may be configured to extract the shape of the point cloud and identify the point cloud by a method that uses a filter to extract object-specific information from the shape of the point cloud. In this case, for example, when distinguishing between a horse and a bird, a filter is designed to extract a head shape, the number of feet, the presence or absence of wings, the presence or absence of a beak, and the like as features. The inference unit 340 extracts a point cloud feature using the filter and distinguishes a point cloud using the point cloud feature.

[0039] The result output unit 350 outputs a classification result by identification by the inference unit 340.

[0040] Specifically, the result output unit 350 outputs the classification result to the result output device 900.

[0041] An example of an internal configuration of the rotation invariance conversion unit 330 will be explained.

[0042] The Fig. The rotation invariance conversion unit 330 shown in Figure 2 comprises an orthogonalization layer unit 331 and a projection layer unit 332.

[0043] The orthogonalization plane unit 331 calculates a basis vector for each of k coordinates indicating a point (each of N points) included in the point cloud information, and performs conversion in such a manner that the k basis vectors are orthogonal to each other to generate an (N × k) orthonormal basis matrix.

[0044] The orthogonalization plane unit 331 performs orthogonalization processing, for example, according to the following method.

[0045] A point cloud P, which is specified in point cloud data (k-dimensional point cloud data) and specifies N (N ≥ 2) points N in k (k ≥ 2) dimensions, is expressed by an N × k matrix, for example, as given in Expression (1). In this case, each point in the point cloud P is represented by k coordinate values such as P N1 ,..., and P Nk expressed as given in expression (1). P=(p11…p1k⋮⋱⋮pN1…pNk)∈ℝN×k

[0046] It is assumed that the center of gravity of the point cloud P is transferred to the coordinate origin.

[0047] The orthogonalization plane unit 331 calculates a basis vector for each of the k coordinates.

[0048] The orthogonalization plane unit 331 performs orthogonalization orth on the point cloud P as shown in Expression (2) to perform a conversion so that the k basis vectors are orthogonal to each other.

[0049] The orthogonalization plane unit 331 obtains an orthonormal basis matrix X by the orthogonalization orth.

[0050] As the orthogonalization orth, the Gram-Schmidt orthogonalization method X = GramSchmit(P) can be used, a QR decomposition (X, r = QR(P)) can be used, a singular value decomposition (X, σ, V T = SVD(P)) can be used, or an eigenvalue decomposition (X, Λ = EIG(P)) can be used. orth(P)=X, stXTX=I_k

[0051] In expression (2), “Ik” stands for a k × k identity matrix.

[0052] The projection plane unit 332 calculates a projection matrix indicating the rotation invariance feature using the orthonormal basis matrix and the model.

[0053] The projection plane unit 332 extracts a projection matrix M, which is a k-dimensional rotation invariance feature (rotation invariance feature), from the orthonormal basis matrix obtained by the orthogonalization plane unit 331.

[0054] The projection matrix M is calculated using the following expression (3). M=XXT

[0055] If the projection matrix is calculated using a matrix XR obtained by a k-dimensional rotation (rotation R) of the orthonormal basis matrix X, the following expression (4) is obtained. XR(XR)T=XXT=M

[0056] For this reason, the projection matrix M is invariant with respect to the rotation R (rotation invariant).

[0057] The projection plane unit 332 calculates and outputs a matrix "Y" obtained by multiplying the projection matrix M by a learning parameter W included in the model, as shown in the following expression (5). In the present disclosure, the learning parameter W shown in expression (5) is a parameter that indicates a geometrically reasonable initial weight and eventually becomes a constant through learning. Y=MW

[0058] The projection matrix M is geometrically a projection matrix onto a space spanned by the projection matrix M.

[0059] "Y," expressed by expression (5), represents a point cloud (shape) in a case where "W" is projected onto a space spanned by "M" by multiplying the projection matrix M by "W" from the right. Thus, "Y" is shape information that is geometrically independent of rotation and denotes a rotation-invariance feature.

[0060] Using this information, the projection plane unit 332 can then obtain the same “Y” from the point cloud P before rotation (orthonormal basis matrix X) and a point cloud PR after rotation (orthonormal basis matrix XR), as will be described later with reference to the Fig. 6A and Fig. 6B.

[0061] In contrast, the projection plane unit 332 can obtain different “Y” and “Y_hat” (this means that “Y” obtained by Expression (5) is different) from the point cloud P1 (orthonormal basis matrix X1) and the point cloud P2 (orthonormal basis matrix X2) having different shapes.

[0062] The Fig. For example, the rotation invariance conversion unit 330 shown in Fig. 2 performs orthogonalization of each basis vector for each of the points specified in the point cloud information as explained above, and calculates the rotation invariance feature using the data after orthogonalization and the model (the learning parameter W indicates the first weight).

[0063] In a case where the rotation invariance conversion unit 330 is configured as described above, the inference unit 340 identifies the point cloud indicated in the point cloud information using the projection matrix and the model indicating the rotation invariance feature.

[0064] An example of an internal configuration of the inference unit 340 is described.

[0065] The inference unit 340 extracts a point cloud feature based on a result of orthogonally projecting a component specified by the learning parameter onto a subspace spanned by the orthonormal basis matrix in an N-dimensional space using the projection matrix and the model that specify the rotation invariance feature, and identifies the point cloud using the point cloud feature. Specifically, class identification of the point cloud is performed.

[0066] The point cloud feature constructed by the inference unit 340 is evaluated with reference to the Fig. 5 and Fig. 6 described.

[0067] First, the rotation of the point cloud is explained.

[0068] In Fig. 5, a ◯ point cloud 1100 (white circle) represents a two-dimensional point cloud X before rotation, and the • point cloud 1200 (black circle) represents a two-dimensional point cloud XR(π) after rotation.

[0069] The two-dimensional point cloud X in Fig. 5 is represented by a matrix X, in which x 11 equal to -0.5, x 12 equals -1.625, x 21 equal to 0.5, x 22 equal to -0.625, x 31 equal to -1.5, x 32 equal to 3.375, x 41 equal to 1.5 and x 42 is equal to -1.125.

[0070] The two-dimensional point cloud XR(π) in Fig. 5 is represented by a matrix XR(π) in which xr 11equal to 0.5, xr 12 equal to 1.625, xr 21 equal to -0.5, xr 22 equal to 0.625, xr 31 equal to 1.5, xr 32 equals -3.375, xr 41 equal to -1.5 and xr 42 is equal to 1.125.

[0071] As in Fig. As shown in Figure 5, the position coordinates change when the point cloud data is rotated, and thus the information before and after rotation is different as numerical data.

[0072] Although the o-group (white circle point) and the •-group (black circle point) are expressed by different coordinate values due to rotation, because the object indicated by the point cloud is the same as the object, even if the coordinates before and after rotation change, the rotated point cloud can also be identified with the same accuracy (object of the same category) without changing the essential information.

[0073] In this context, the feature construction based on the projection matrix M can be performed by the inference unit 340 with reference to Fig. 6 can be described geometrically.

[0074] The orthonormal basis matrix X is a matrix in which the k basis vectors of each of the N points are arranged as in the following expression (6). X=(x11…x1k⋮⋱⋮xN1⋯xNk)

[0075] The orthonormal basis matrix X spans a k-dimensional subspace spanX in an N-dimensional space. That is, the subspace spanX is a space generated by the orthonormal basis matrix X.

[0076] In Fig. 6A, a subspace 2100 (spanX) represents a column space of an orthonormal basis matrix X obtained from the point cloud P. The orthonormal basis matrix X in Fig. 6A indicates that x 11 -0.5, x 12 -1.625, x 21 0.5, x 22 -0.625, x 31 -1.5, x32 3,375, x 41 1.5 and x 42 is -1.125.

[0077] In Fig. 6A, a subspace 2200 (spanXR) represents a column space of an orthonormal basis matrix XR obtained from the point cloud PR. The orthonormal basis matrix XR(π) in Fig. Figure 6A shows that xr 11 equal to 0.5, xr 12 equal to 1.625, xr 21 equal to -0.5, xr 22 equal to 0.625, xr 31 equal to 1.5, xr 32 equals -3.375, xr 41 equal to -1.5 and xr 42 is equal to 1.125.

[0078] The orthonormal basis matrix X and the orthonormal basis matrix XR are orthonormal basis matrices generated from point clouds of the same shape. In this case, the subspace 2100 (spanX) and the subspace 2200 (spanXR) coincide.

[0079] In Fig. 6B, subspace 3100 (spanX1) represents a column space of the orthonormal basis matrix X1 obtained from the point cloud P1. The orthonormal basis matrix X1 in Fig. 6B indicates that x1 11 -0.5, x1 12 -1.625, x1 21 0.5, x1 22 -0.625, x1 31 -1.5, x1 32 3,375, x1 41 1.5 and x1 42 is -1.125.

[0080] In Fig. 6B, the subspace 3200 (spanX2) represents a column space of the orthonormal basis matrix X2 obtained from the point cloud P2. The orthonormal basis matrix X2 in Fig. 6B indicates that x2 11 -4,676, x2 12 3.1419615, x2 21 3,892, x2 22 5.1419615, x2 31 1,892, x2 32 -7.4258845, x2 41 -1.108 and x2 42 - 0.8580385.

[0081] The orthonormal basis matrix X1 and the orthonormal basis matrix X2 are orthonormal basis matrices generated from point clouds with different shapes. In this case, the subspace 3100 (spanX1) and the subspace 3200 (spanX2) do not coincide.

[0082] Based on this fact, the inference unit 340 extracts a point cloud feature, which is a feature effective for identifying a point cloud, using an orthonormal basis matrix for each point cloud.

[0083] The inference unit 340 extracts a point cloud feature Z using, for example, the following expression (7). Z=Φ(X)=σ(WY)+B

[0084] In expression (7), “Y” is obtained by multiplying the projection matrix M by the parameter W, which indicates the first weight as expressed in expression (5).

[0085] In expression (7), the learning parameter W indicates a weight (second weight), and a learning parameter B (second learning parameter) indicates a bias.

[0086] Expression (7) indicates that a feature extraction function ϕ as a feature extractor has a learning parameter (a parameter W indicating a first weight, a parameter W indicating a second weight, and a parameter B indicating a bias) and a nonlinear conversion function σ for nonlinear conversion, and the point cloud feature Z is extracted by inputting the orthonormal basis matrix X into the feature extraction function ϕ.

[0087] The extraction of the point cloud feature Z by ϕ(X) in expression (7) corresponds to performing an orthogonal projection onto the subspace spanX (subspace 2100) spanned by the orthonormal basis matrix X.

[0088] When projecting orthogonally onto the subspace spanX (subspace 2100) and the subspace spanXR (subspace 2200), the parameter W is projected onto the same coordinates, since these subspaces coincide. This corresponds to the extraction of a rotationally invariant feature that is unique to the point cloud shape.

[0089] In contrast, orthogonal projections onto the subspace spanX1 (subspace 3100) and the subspace spanX2 (subspace 3200) project the parameter W onto different coordinates, since these subspaces do not coincide with each other.

[0090] In this way, the point cloud information can be identified using the point cloud shape-specific features that do not depend on the point cloud coordinates.

[0091] Note that in the description, the inference unit 340 is configured to receive "Y" calculated by multiplying the projection matrix (projection matrix M) by the learning parameter (parameter W indicating the first weight) included in the model and output by the projection plane unit 332. However, the same applies if, in the inference unit 340, it is configured to multiply the projection matrix (projection matrix M) by the learning parameter (parameter W indicating the first weight) included in the model. In such a configuration, the projection plane unit 332 outputs the projection matrix M as is.

[0092] The result output device 900 receives and outputs a classification result of the point cloud output from the point cloud identification device 300. The result output device 900 only needs to be a device that uses the classification result of the point cloud, and may be, for example, a display device that simply displays the result, or may be a control device for automated driving of a vehicle.

[0093] The processing by the point cloud identification device 300 is described with reference to the Fig. 3 and Fig. 4 explained.

[0094] After starting the processing, the point cloud identification device 300 executes the model acquisition processing (step ST110).

[0095] Specifically, in the point cloud identification device 300, the model acquisition unit 310 acquires a model having a learning parameter from the storage unit 210 of the storage device 200. The model acquisition unit 310 outputs the model having the learning parameter.

[0096] The point cloud identification device 300 executes the point cloud acquisition processing (step ST120).

[0097] Specifically, in the point cloud identification device 300, the point cloud acquisition unit 320 acquires point cloud information indicating N (N ≥ 2) points in k (k ≥ 2) dimensions from the point cloud input device 100. The point cloud acquisition unit 320 outputs the acquired point cloud information.

[0098] The point cloud identification device 300 performs rotation invariance conversion processing (step ST130).

[0099] Specifically, the rotation invariance conversion unit 330 in the point cloud identification device 300 performs orthogonalization of each basis vector for each of the points specified in the point cloud information, and calculates the rotation invariance feature using data after the orthogonalization.

[0100] A concrete example of rotation invariance conversion processing is explained.

[0101] After starting rotation invariance conversion processing, the rotation invariance conversion unit 330 executes the orthogonalization processing (step ST131) as shown in Fig. 4 shown.

[0102] Specifically, the orthogonalization plane unit 331 in the rotation invariance conversion unit 330 calculates a basis vector for each of k coordinates indicating a point (each of N points) included in the point cloud information, and performs conversion in such a manner that the k basis vectors are orthogonal to each other to generate an (N × k) orthonormal basis matrix.

[0103] The orthogonalization plane unit 331 uses, for example, the Gram-Schmidt orthogonalization method, QR decomposition, singular value decomposition, or eigenvalue decomposition to perform the conversion in such a way that the basis vectors are orthogonal to each other to generate an (N x k) orthonormal basis matrix. The orthogonalization plane unit 331 outputs an orthonormal basis matrix.

[0104] The rotation invariance conversion unit 330 executes projection processing (step ST132).

[0105] Specifically, in the rotation invariance conversion unit 330, the projection plane unit 332 acquires the orthonormal basis matrix from the orthogonalization plane unit 331. The projection plane unit 332 calculates a projection matrix indicating the rotation invariance feature using the orthonormal basis matrix. The projection plane unit 332 acquires the model from the model acquisition unit 310, multiplies the projection matrix (projection matrix M) by the learning parameter (learning parameter W indicating the first weight) included in the model, and outputs a result.

[0106] The point cloud identification device 300 performs inference processing (step ST140).

[0107] Specifically, in the point cloud identification device 300, the inference unit 340 first acquires the projection matrix from the projection plane unit 332, and acquires the model (the learning parameter W indicating a second weight and the learning parameter B indicating a bias) from the model acquisition unit 310. Next, the inference unit 340 uses the rotation invariance feature and the model to identify the point cloud indicated in the point cloud information.More specifically, as described above, the inference unit 340 extracts a point cloud feature based on a result of orthogonally projecting a component specified by the learning parameter onto a subspace spanned by the orthonormal basis matrix in an N-dimensional space using the projection matrix and the model specifying the rotation invariance feature, and identifies the point cloud using the point cloud feature.

[0108] Note that in a case where the projection plane unit 332 in the rotation invariance conversion unit 330 is configured to output the projection matrix without multiplying the learning parameter (parameter W), the projection matrix is multiplied by the learning parameter (parameter W) in the inference unit 340.

[0109] The point cloud identification device 300 executes result output processing (step ST150).

[0110] Specifically, the result output unit 350 in the point cloud identification device 300 outputs a classification result by identifying the inference unit 340 to the result output device 900.

[0111] The point cloud identification device 300 determines whether the processing sequence is terminated (step ST160).

[0112] When it is determined not to end the processing (step ST160: “No”), the point cloud identification device 300 proceeds to the processing of step ST110 and repeats the processing of step ST110.

[0113] When it is determined to end the processing (step ST160 “YES”), the point cloud identification device 300 ends the processing.

[0114] An example of the effect of the point cloud identification device of the present disclosure will be described.

[0115] For example, in automated driving technology, to detect the near environment of a vehicle, other vehicles and obstacles in the periphery are identified using point cloud data. In this case, by processing the point cloud data in real time and identifying the point cloud, it is possible to check for the presence of a vehicle ahead to avoid a collision, or to detect an obstacle, debris, or the like that the automated driving device is approaching. Here, there is a case where the point cloud is rotated to detect a state where the point cloud is viewed from different angles, and separate processing such as data augmentation through rotation and attitude adjustment is usually required.

[0116] In contrast, in the present disclosure, as described above, the configuration in which the point cloud data expressed by the orthonormal basis matrix is treated as a single subspace, and the projection matrix is treated as data for acquiring rotation invariance was explained. Since the point cloud in the present disclosure is expressed as rotation invariance data, learning only a single position is equivalent to learning each position individually, and it is possible to acquire an identification model that is robust to any rotation, thereby not only making learning more efficient but also achieving the same identification accuracy before and after rotation.

[0117] The point cloud identification device according to the present disclosure is configured as follows.

[0118] Point cloud identification device, comprising: a point cloud acquisition unit to acquire point cloud information indicating N (N ≥ 2) points in k (k ≥ 2) dimensions; a model acquisition unit to acquire a model that has a learning parameter; a rotation invariance conversion unit for performing orthogonalization of each of basis vectors for each of points specified in the point cloud information, and calculating a rotation invariance feature using data after the orthogonalization; an inference unit for identifying a point cloud specified in the point cloud information using the rotation invariance feature and the model; and a result output unit to output a classification result by identifying by the inference unit.

[0119] Thus, the present disclosure has the effect that a point cloud identification device can be provided which improves the accuracy of identifying a point cloud.

[0120] The point cloud identification method according to the present disclosure is configured as follows.

[0121] Point cloud identification method, comprising: a point cloud acquisition step of causing a point cloud acquisition unit to acquire point cloud information indicating N (N ≥ 2) points in k (k ≥ 2) dimensions; a model acquisition step of causing a model acquisition unit to acquire a model having a learning parameter; a rotation invariance conversion step of causing a rotation invariance conversion unit to perform orthogonalization of each of basis vectors for each of points specified in the point cloud information, and to calculate a rotation invariance feature using the information after the orthogonalization; an inference step of causing an inference unit to identify a point cloud specified in the point cloud information using the rotation invariance feature and the model; and a result output step of causing a result output unit to output a classification result by identification by the inference unit.

[0122] Thus, the present disclosure has the effect of providing a point cloud identification method that improves the accuracy of identifying a point cloud.

[0123] The point cloud identification device according to the present disclosure is further configured as follows.

[0124] Point cloud identification device in which the rotation invariance conversion unit has: an orthogonalization plane unit for calculating a basis vector for each of k coordinates indicating a point included in the point cloud information, and performing conversion in such a manner that the k basis vectors are orthogonal to each other to generate an orthonormal basis matrix; and a projection plane unit for calculating a projection matrix indicating the rotation invariance feature using the orthonormal basis matrix, and the inference unit identifies a point cloud specified in the point cloud information using the projection matrix specifying the rotation invariance feature and the model.

[0125] Thus, the present disclosure has the effect that a point cloud identification device can be provided which is capable of efficiently calculating a rotation invariance feature of a point cloud.

[0126] Furthermore, the present disclosure shows a similar effect to the above-described effect when the above-described configuration is applied to the point cloud identification method.

[0127] The point cloud identification device according to the present disclosure is further configured as follows.

[0128] In the point cloud identification facility, extracts the inference unit, using the projection matrix, which specifies the rotation invariance feature, and the model, extracting a point cloud feature based on a result of orthogonally projecting a component specified by the learning parameter onto a subspace spanned by the orthonormal basis matrix in an N-dimensional space, and identifying a point cloud using the point cloud feature.

[0129] Thus, the present disclosure has the effect of making it possible to provide a point cloud identification device that identifies point cloud information using the point cloud shape-specific features that do not depend on the coordinates of the point cloud.

[0130] Furthermore, the present disclosure shows a similar effect to the above-described effect when the above-described configuration is applied to the point cloud identification method. Second embodiment.

[0131] A second embodiment describes a form of learning device.

[0132] In the second embodiment, a detailed description of the same contents as those already explained will be omitted if necessary.

[0133] Fig. 7 is a diagram illustrating an example of a configuration of a point cloud learning system 2 including a point cloud learning device 400 according to the second embodiment.

[0134] The point cloud learning system 2 includes a point cloud learning device 100, a storage device 200A, and a point cloud learning device 400.

[0135] The point cloud input device 100 is similar to the point cloud input device 100 explained above, so a detailed description thereof is omitted here.

[0136] The storage device 200A comprises a storage unit 210A.

[0137] The storage unit 210A contains information used for the identification processing of identifying a point cloud. Specifically, the storage unit 210A contains, for example, a learning parameter as a model for identification. The model is appropriately trained and updated by the point cloud learning device 400.

[0138] The point cloud identification device 400 trains a rotation invariance model in a k-dimensional point cloud identification, and identifies a point cloud using the learned model.

[0139] K-dimensional point cloud identification is used to perform class identification of point cloud data that specifies a point cloud represented in k dimensions, based on a feature such as the shape of the point cloud.

[0140] A configuration example of the point cloud learning device 400 is explained.

[0141] The point cloud learning device 400 includes a model learning unit 410, a point cloud acquisition unit 420, a rotation invariance conversion unit 430, an inference unit 440, a result output unit 450, an evaluation unit 460, and a model update unit 470.

[0142] The model acquisition unit 410 acquires a model having a learning parameter.

[0143] The learning parameter is a parameter trained by the point cloud learning device 400 and is a parameter used when the point cloud is identified.

[0144] The point cloud acquisition unit 420 acquires k-dimensional point cloud information of N points.

[0145] The point cloud acquisition unit 420 acquires point cloud information indicating N (N ≥ 2) points in k (k ≥ 2) dimensions, similarly to the point cloud acquisition unit 320 described above.

[0146] Similar to the rotation invariance conversion unit 330 described above, the rotation invariance conversion unit 430 performs orthogonalization of each basis vector for each of the points specified in the point cloud information, and calculates a rotation invariance feature using data after the orthogonalization.

[0147] Similar to the inference unit 340 described above, the inference unit 440 uses the rotation invariance feature and the model to identify the point cloud specified in the point cloud information. Specifically, the class identification of the point cloud is performed.

[0148] The result output unit 450 outputs a classification result by identification by the inference unit.

[0149] Specifically, the result output unit 450 outputs the classification result to the evaluation unit 460.

[0150] In addition, the result output unit 450 can be configured to further output the classification result to an external device. In this case, the external device can be, for example, the result output device 900 described in the first embodiment.

[0151] The evaluation unit 460 evaluates the model using the classification result.

[0152] Various evaluation methods can be considered as an evaluation method in the evaluation unit 460, provided that the evaluation method is an index for measuring an error between an output label and a correct answer label. For example, a basic evaluation method includes an evaluation method using a cross-entropy error function or a quadratic error function.

[0153] The evaluation unit 460 outputs an evaluation result to the model update unit 470.

[0154] The model updating unit 470 updates the model using the evaluation result by the evaluation unit 460.

[0155] Specifically, the model updating unit 470 updates the learning parameter as the model stored in the storage unit 210A of the storage device 200A by rewriting the learning parameter.

[0156] For example, when using the evaluation result by the cross-entropy error function, the model update unit 470 updates the parameter W in a direction in which the value of the error function is minimized. The direction in which the function value is minimized is a gradient direction in which the function decreases when the error function is differentiated by the model parameter W. Note that at this time, the model update unit 470 may perform the update until the value of the error function converges, or stop the update midway when a predetermined condition is met.

[0157] Furthermore, the model update unit 470 performs the update using a general optimization method, such as a stochastic gradient descent method or a Newton method. Note that at the time of gradient calculation, the gradients of all model parameters are calculated sequentially, for example, using a method called backpropagation.

[0158] An example of an internal configuration of the rotation invariance conversion unit 430 will be explained.

[0159] Here, the internal configuration of the rotation invariance conversion unit 430 is obtained by replacing the orthogonalization plane unit 331 and the projection plane unit 332 in the rotation invariance conversion unit 330 already described in Fig. 2, can be replaced by an orthogonalization plane unit 431 and a projection plane unit 432, and is not shown.

[0160] The rotation invariance conversion unit 430 includes an orthogonalization plane unit 431 and a projection plane unit 432.

[0161] Similar to the orthogonalization plane unit 331 described above, the orthogonalization plane unit 431 calculates a basis vector for each of k coordinates indicating a point (each of N points) in the point cloud information, and performs conversion in such a manner that the k basis vectors are orthogonal to each other to generate an (N × k) orthonormal basis matrix.

[0162] Similar to the projection plane unit 332 described above, the projection plane unit 432 calculates a projection matrix indicating a rotation invariance feature using an orthonormal basis matrix.

[0163] The orthogonalization processing by the orthogonalization plane unit 431 is similar to the orthogonalization processing by the orthogonalization plane unit 331 described above, and its more detailed description is omitted here.

[0164] In a case where the rotation invariance conversion unit 430 is configured in this way, the inference unit 440 identifies the point cloud indicated in the point cloud information using the projection matrix and the model indicating the rotation invariance feature.

[0165] An internal configuration of the inference unit 440 is explained below.

[0166] The inference unit 440 extracts a point cloud feature based on a result of orthogonally projecting a component specified by the learning parameter onto a subspace spanned by the orthonormal basis matrix in an N-dimensional space using the projection matrix and the model that specify the rotation invariance feature, and identifies the point cloud using the point cloud feature. Specifically, class identification of the point cloud is performed.

[0167] The point cloud feature constructed by the inference unit 440 is similar to the point cloud feature constructed by the inference unit 340 described above, so a more detailed description is omitted here.

[0168] In this way, in the present disclosure, the point cloud information can be identified based on the point cloud shape-specific features that do not depend on the point cloud coordinates.

[0169] The processing of the point cloud identification device is described with reference to Fig. 8 explained.

[0170] Fig. 8 is a flowchart illustrating an example of processing by the point cloud learning device 400.

[0171] After starting the processing, the point cloud learning device 400 first determines whether learning is performed (step ST100).

[0172] Specifically, the point cloud learning device 400 selects whether to learn the learning parameter from the beginning. For example, the point cloud learning device 400 checks predefined setting information and determines whether to learn from the beginning or to use the learning parameters of the previously saved model.

[0173] If it is determined that learning should be performed (step ST100 “YES”), the point cloud learning device 400 proceeds to the processing of step ST120.

[0174] In the case of learning from scratch, the point cloud learning device 400 may initialize the learning parameters randomly or may initialize the learning parameters using any commonly used initialization method.

[0175] When the point cloud learning device 400 determines not to perform learning (step ST100 "NO"), the point cloud learning device 400 performs model acquisition processing (step ST110).

[0176] After determining to perform learning (step ST100 “YES”) or executing the model acquisition processing (step ST110), the point cloud learning device 400 executes the point cloud acquisition processing (step ST120).

[0177] Specifically, in the point cloud learning device 400, the point cloud acquisition unit 420 acquires point cloud information indicating N (N ≥ 2) points in k (k ≥ 2) dimensions from the point cloud input device 100. The point cloud acquisition unit 420 outputs the acquired point cloud information.

[0178] The point cloud learning device 400 executes rotation invariance conversion processing (step ST130).

[0179] Specifically, the rotation invariance conversion unit 430 in the point cloud learning device 400 performs orthogonalization of each of basis vectors for each of points specified in the point cloud information, and calculates the rotation invariance feature using data after the orthogonalization.

[0180] A concrete example of the rotation invariance conversion processing will be explained. It should be noted that the rotation invariance conversion processing by the point cloud learning device 400 is similar to the rotation invariance conversion processing of Fig. 4, and further illustration is omitted here.

[0181] After starting the rotation invariance conversion processing, the rotation invariance conversion unit 430 executes the orthogonalization processing (step ST131) as shown in Fig. 4 shown.

[0182] Specifically, the orthogonalization plane unit 431 in the rotation invariance conversion unit 430 calculates a basis vector for each of k coordinates indicating a point (each of N points) in the point cloud information, and performs the conversion in such a manner that the k basis vectors are orthogonal to each other to generate an (N × k) orthonormal basis matrix. The orthogonalization plane unit 431 outputs an orthonormal basis matrix.

[0183] The rotation invariance conversion unit 430 executes projection processing (step ST132).

[0184] Specifically, in the rotation invariance conversion unit 430, the projection plane unit 432 acquires the orthonormal basis matrix from the orthogonalization plane unit 431. The projection plane unit 432 calculates a projection matrix indicating the rotation invariance feature using the orthonormal basis matrix.

[0185] The point cloud learning device 400 performs inference processing (step ST140).

[0186] Specifically, in the point cloud learning device 400, the inference unit 440 first acquires the projection matrix from the projection plane unit 432 and acquires the model from the model acquisition unit 410. Next, the inference unit 440 uses the rotation invariance feature and the model to identify the point cloud specified in the point cloud information.

[0187] More specifically, the inference unit 440 extracts a point cloud feature based on a result of orthogonally projecting a component specified by the learning parameter onto a subspace spanned by the orthonormal basis matrix in an N-dimensional space using the projection matrix and the model indicating the rotation invariance feature, and identifies the point cloud using the point cloud feature.

[0188] The point cloud learning device 400 executes result output processing (step ST150).

[0189] Specifically, the result output unit 450 in the point cloud learning device 400 outputs the classification result by identifying the inference unit 440 to the result output device 900.

[0190] The point cloud learning device 400 determines whether to terminate the processing sequence (step ST160).

[0191] When it is determined to end the processing (step ST160 “YES”), the point cloud learning device 400 ends the processing.

[0192] When it is determined not to end the processing (step ST160 "NO"), the point cloud learning device 400 executes the model evaluation processing (step ST170).

[0193] Specifically, the evaluation unit 460 in the point cloud learning device 400 evaluates the model using the classification result. The evaluation unit 460 outputs an evaluation result to the model update unit 470.

[0194] The point cloud learning device 400 executes model update processing (step ST180).

[0195] The model updating unit 470 updates the model using the evaluation result by the evaluation unit 460.

[0196] Specifically, the model updating unit 470 updates the learning parameter as the model stored in the storage unit 210A of the storage device 200A by rewriting the learning parameter.

[0197] After executing the model update processing (step ST180), the point cloud learning device 400 proceeds to the processing of step ST110 and repeats the processing from step ST110.

[0198] The learning device (point cloud learning device) in the sense of the present disclosure is configured as follows.

[0199] Learning facility, comprising: a point cloud acquisition unit to acquire point cloud information indicating N (N ≥ 2) points in k (k ≥ 2) dimensions; a model acquisition unit to acquire a model that has a learning parameter; a rotation invariance conversion unit for performing orthogonalization of each of basis vectors for each of points specified in the point cloud information, and calculating a rotation invariance feature using data after the orthogonalization; an inference unit for identifying a point cloud specified in the point cloud information using the rotation invariance feature and the model; a result output unit for outputting a classification result by identification by the inference unit; an evaluation unit to evaluate the model based on the classification result; and a model update unit to update the model using an evaluation result by the evaluation unit.

[0200] Thus, the present disclosure has the effect that a point cloud learning device can be provided which improves the accuracy of identification of a point cloud.

[0201] The learning method (point cloud learning method) in the sense of the present disclosure is configured as follows.

[0202] Learning methods, including: a point cloud acquisition step of causing a point cloud acquisition unit to acquire point cloud information indicating N (N ≥ 2) points in k (k ≥ 2) dimensions; a model acquisition step of causing a model acquisition unit to acquire a model having a learning parameter; a rotation invariance conversion step of causing a rotation invariance conversion unit to perform orthogonalization of each of basis vectors for each of points specified in the point cloud information, and to calculate a rotation invariance feature using the information after the orthogonalization; an inference step of causing an inference unit to identify a point cloud specified in the point cloud information using the rotation invariance feature and the model; a result output step of causing a result output unit to output a classification result by identification by the inference unit; an evaluation step of causing an evaluation unit to evaluate the model using the classification result; and a model updating step of causing a model updating unit to update the model using an evaluation result by the evaluation unit.

[0203] Thus, the present disclosure has the effect of providing a learning method that improves the accuracy of identifying a point cloud.

[0204] The learning device (point cloud learning device) in the sense of the present disclosure is further configured as follows.

[0205] Learning institution in which the rotation invariance conversion unit has: an orthogonalization plane unit for calculating a basis vector for each of k coordinates indicating a point included in the point cloud information, and performing conversion in such a manner that the k basis vectors are orthogonal to each other to generate an orthonormal basis matrix; and a projection plane unit for calculating a projection matrix indicating the rotation invariance feature using the orthonormal basis matrix, and the inference unit identifies a point cloud specified in the point cloud information using the projection matrix specifying the rotation invariance feature and the model.

[0206] Thus, the present disclosure has the effect of providing a learning device capable of efficiently calculating a rotation invariance feature of a point cloud.

[0207] Furthermore, the present disclosure achieves a similar effect to the above-described effect when the above-described configuration is applied to the above-described learning method.

[0208] The learning device (point cloud learning device) in the sense of the present disclosure is further configured as follows.

[0209] In the learning facility, the inference unit has: extracts the inference unit, using the projection matrix, which specifies the rotation invariance feature, and the model, extracting a point cloud feature based on a result of orthogonally projecting a component specified by the learning parameter onto a subspace spanned by the orthonormal basis matrix in an N-dimensional space, and identifying a point cloud using the point cloud feature.

[0210] Thus, the present disclosure has the effect of making it possible to provide a learning device that identifies point cloud information using the point cloud shape-specific features that do not depend on the coordinates of the point cloud.

[0211] Furthermore, the present disclosure achieves a similar effect to the above-described effect when the above-described configuration is applied to the above-described learning method.

[0212] Here, a hardware configuration implementing the functions of the point cloud identification device 300 and the point cloud learning device 400 according to the present disclosure will be described.

[0213] Fig. 9 is a diagram showing a first example of a hardware configuration for implementing the functions of the point cloud identification device 300 and the point cloud learning device 400 according to the present disclosure.

[0214] Fig. 10 is a diagram showing a second example of a hardware configuration for implementing the functions of the point cloud identification device 300 and the point cloud learning device 400 according to the present disclosure.

[0215] The point cloud identification device 300 and the point cloud learning device 400 of the present disclosure are implemented by hardware as shown in Fig. 9 or Fig. 10 shown.

[0216] As in Fig. 9, the point cloud identification device 300 and the point cloud learning device 400 include, for example, a processor 10001, a memory 10002, an input / output interface 10003, and a communication circuit 10004.

[0217] For example, the processor 10001 and the RAM 10002 are built into a computer.

[0218] The working memory 10002 stores a program that causes the computer to operate as a model acquisition unit 310, a point cloud acquisition unit 320, a rotation invariance conversion unit 330, an inference unit 340, a result output unit 350, a model acquisition unit 410, a point cloud acquisition unit 420, a rotation invariance conversion unit 430, an inference unit 440, a result output unit 450, an evaluation unit 460, a model update unit 470, and a control unit (not shown).When the processor 10001 reads and executes the program stored in the memory 10002, the functions of the model acquisition unit 310, the point cloud acquisition unit 320, the rotation invariance conversion unit 330, the inference unit 340, the result output unit 350, the model acquisition unit 410, the point cloud acquisition unit 420, the rotation invariance conversion unit 430, the inference unit 440, the result output unit 450, the evaluation unit 460, the model update unit 470, and the control unit (not shown) are implemented.

[0219] In addition, a storage unit, which is not shown, is realized by the working memory 10002 or another working memory, which is not shown.

[0220] The processor 10001 uses, for example, a central processing unit (CPU), a graphics processing unit (GPU), a microprocessor, a microcontroller, a digital signal processor (DSP), or the like.

[0221] The memory 10002 may be, for example, a non-volatile or volatile semiconductor memory such as a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), a flash memory, or the like, a magnetic disk such as a hard disk or a flexible disk, an optical storage disk such as a compact disc (CD) or a digital versatile disc (DVD), or a magneto-optical disk.

[0222] The processor 10001 and the memory 10002 are connected in a state where data can be transferred to each other. Furthermore, the processor 10001 and the memory 10002 are connected in a state where data can be transferred to other hardware via an input / output interface 10003.

[0223] Furthermore, a communication unit (not shown) is implemented by the communication circuit 10004.

[0224] Alternatively, the functions of the model acquisition unit 310, the point cloud acquisition unit 320, the rotation invariance conversion unit 330, the inference unit 340, the result output unit 350, the model acquisition unit 410, the point cloud acquisition unit 420, the rotation invariance conversion unit 430, the inference unit 440, the result output unit 450, the evaluation unit 460, the model update unit 470 and the control unit, which is not shown, can be implemented by a Fig. 10 may be implemented by a dedicated processing circuit 20001.

[0225] The processing circuit 20001 uses, for example, a single circuit, a composite circuit, a programmed processor, a parallel programmed processor, an application-specific integrated circuit (ASIC), a programmable logic device (PLD), a field-programmable gate array (FPGA), a system-on-a-chip (SoC), a system-large-scale integration (LSI), or the like.

[0226] In addition, a memory unit, which is not shown, is realized by the main memory 20002 or another main memory, which is not shown.

[0227] The memory 20002 may be, for example, a non-volatile or volatile semiconductor memory such as a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), a flash memory, or the like, a magnetic disk such as a hard disk or a flexible disk, an optical storage disk such as a compact disc (CD) or a digital versatile disc (DVD), or a magneto-optical disk.

[0228] The processing circuit 20001 and the memory 20002 are connected in a state where data can be transferred to each other. Furthermore, the processor 20001 and the memory 20002 are connected in a state where data can be transferred to other hardware via an input / output interface 20003.

[0229] Furthermore, a communication unit (not shown) is implemented by the communication circuit 20004.

[0230] Note that the functions of the model acquisition unit 310, the point cloud acquisition unit 320, the rotation invariance conversion unit 330, the inference unit 340, the result output unit 350, the model acquisition unit 410, the point cloud acquisition unit 420, the rotation invariance conversion unit 430, the inference unit 440, the result output unit 450, the evaluation unit 460, the model update unit 470, and the control unit, which is not shown, may be implemented by different processing circuits, or may be implemented collectively by one processing circuit.

[0231] Alternatively, some functions of the model acquisition unit 310, the point cloud acquisition unit 320, the rotation invariance conversion unit 330, the inference unit 340, the result output unit 350, the model acquisition unit 410, the point cloud acquisition unit 420, the rotation invariance conversion unit 430, the inference unit 440, the result output unit 450, the evaluation unit 460, the model update unit 470, and the control unit, which is not shown, may be implemented by the processor 10001 and the memory 10002, and the remaining functions may be implemented by the processing circuit 20001.

[0232] It should be noted that the present disclosure can freely combine the respective embodiments, modify any component of each of the embodiments, or omit any component in each of the embodiments within the scope of the invention. For example, the first embodiment and the second embodiment can be combined such that the result output unit outputs the point cloud identification result via the evaluation unit to the result output device and the model update unit. In this way, both the effects of the first embodiment and the second embodiment can be achieved. INDUSTRIAL APPLICABILITY

[0233] Both the point cloud identification device and the learning device according to the present disclosure can improve the accuracy of the identification of the point cloud and are therefore suitable for use in the identification of the point cloud in technologies such as vehicle control and driving assistance. LIST OF REFERENCE SYMBOLS

[0234] 1: Point cloud identification system, 2: Point cloud learning system, 100: Point cloud input device, 200, 200A: Storage device, 210, 210A: Storage unit, 300: Point cloud identification device, 310: Model acquisition unit, 320: Point cloud acquisition unit, 330: Rotation invariance conversion unit, 331: Orthogonalization plane unit, 332: Projection plane unit, 340: Inference unit, 350: Result output unit, 400: Point cloud learning device, 410: Model acquisition unit, 420: Point cloud acquisition unit, 430: Rotation invariance conversion unit, 431: Orthogonalization plane unit, 432: Projection plane unit, 440: Inference unit, 450: Result output unit, 460: Evaluation unit, 470: Model update unit, 900: Result output device QUOTES CONTAINED IN THE DESCRIPTION

[0000] This list of documents submitted by the applicant was generated automatically and is included solely for the convenience of the reader. This list is not part of the German patent or utility model application. The DPMA assumes no liability for any errors or omissions. Cited patent literature

[0000] JP 2019-133545 A

[0004]

Claims

[1] Point cloud identification device, comprising: a point cloud acquisition unit to acquire point cloud information indicating N (N ≥ 2) points in k (k ≥ 2) dimensions; a model acquisition unit to acquire a model that has a learning parameter; a rotation invariance conversion unit for performing orthogonalization of each of basis vectors for each of points specified in the point cloud information, and calculating a rotation invariance feature using data after the orthogonalization; an inference unit for identifying a point cloud specified in the point cloud information using the rotation invariance feature and the model; and a result output unit to output a classification result by identification by the inference unit. [2] The point cloud identification device according to claim 1, wherein the rotation invariance conversion unit comprises: an orthogonalization plane unit for calculating a basis vector for each of k coordinates indicating a point included in the point cloud information, and performing conversion in such a manner that the k basis vectors are orthogonal to each other to generate an orthonormal basis matrix; and a projection plane unit for calculating a projection matrix indicating the rotation invariance feature using the orthonormal basis matrix, and an inference unit that identifies a point cloud specified in the point cloud information using the projection matrix specifying the rotation invariance feature and the model. [3] Point cloud identification device according to claim 2, wherein the inference unit, using the projection matrix, which specifies the rotation invariance feature, and the model, extracting a point cloud feature based on a result of orthogonally projecting a component specified by the learning parameter onto a subspace spanned by the orthonormal basis matrix in an N-dimensional space, and identifying the point cloud using the point cloud feature. [4] Learning facility comprising: a point cloud acquisition unit to acquire point cloud information indicating N (N ≥ 2) points in k (k ≥ 2) dimensions; a model acquisition unit to acquire a model that has a learning parameter; a rotation invariance conversion unit for performing orthogonalization of each of basis vectors for each of points specified in the point cloud information, and calculating a rotation invariance feature using data after the orthogonalization; an inference unit for identifying a point cloud specified in the point cloud information using the rotation invariance feature and the model; an output unit for outputting a classification result by identification by the inference unit; an evaluation unit to evaluate the model using the classification result; and a model update unit to update the model using an evaluation result by the evaluation unit. [5] The learning device according to claim 4, wherein the rotation invariance conversion unit comprises: an orthogonalization plane unit for calculating a basis vector for each of k coordinates indicating a point included in the point cloud information, and performing conversion in such a manner that the k basis vectors are orthogonal to each other to generate an orthonormal basis matrix; and a projection plane unit for calculating a projection matrix indicating the rotation invariance feature using the orthonormal basis matrix, and an inference unit that identifies a point cloud specified in the point cloud information using the projection matrix specifying the rotation invariance feature and the model. [6] Learning device according to claim 5, wherein the inference unit, using the projection matrix, which specifies the rotation invariance feature, and the model, extracting a point cloud feature based on a result of orthogonally projecting a component specified by the learning parameter onto a subspace spanned by the orthonormal basis matrix in an N-dimensional space, and identifying the point cloud using the point cloud feature. [7] Point cloud identification method, comprising: Acquiring, by a point cloud acquisition unit, point cloud information indicating N (N ≥ 2) points in k (k ≥ 2) dimensions; Acquiring, by a model acquisition unit, a model that has a learning parameter; Performing, by a rotation invariance conversion unit, an orthogonalization of each of basis vectors for each of points specified in the point cloud information, and calculating, by the rotation invariance conversion unit, a rotation invariance feature using information after the orthogonalization; Identifying, by an inference unit, a point cloud specified in the point cloud information using the rotation invariance feature and the model; and Outputting, by a result output unit, a classification result by identification by the inference unit. [8] Learning methods, including: Acquiring, by a point cloud acquisition unit, point cloud information indicating N (N ≥ 2) points in k (k ≥ 2) dimensions; Acquiring, by a model acquisition unit, a model that has a learning parameter; Performing, by a rotation invariance conversion unit, an orthogonalization of each of basis vectors for each of points specified in the point cloud information, and calculating, by the rotation invariance conversion unit, a rotation invariance feature using information after the orthogonalization; Identifying, by an inference unit, a point cloud specified in the point cloud information using the rotation invariance feature and the model; Outputting, by a result output unit, a classification result by identification by the inference unit; Evaluating, by an evaluation unit, the model using the classification result; and Updating, by a model updating unit, the model using an evaluation result by the evaluation unit.

Images