JP6829626B2 - Reference plane generation method and device from point cloud data - Google Patents

Description

The present invention relates to a method and an apparatus for generating a reference plane from the point cloud data in order to analyze planar point cloud data acquired from a road or the like by a measuring device.

In general, roads have ruts and irregularities over time due to the passage of vehicles and the like, and these need to be repaired. In order to repair such a road, the road is inspected and data on the road surface properties, that is, data on the state of unevenness of the road surface is acquired. Data on these road surface properties are obtained by measurement by a measurer or by traveling a road surface measuring vehicle along a measurement route of a road to be measured. The road surface measuring vehicle is equipped with a measuring device that irradiates the road surface with scan light to measure the height of each point on the road surface.

Patent Document 1 describes a means for detecting a moving distance and a means for projecting light in a device that projects light toward a plane while moving a moving body in a longitudinal direction of the plane and measures a step on the plane based on the result of the projection. The configuration includes means for imaging a light irradiation line, cross-sectional data calculation means for acquiring height data, vertical data calculation means, and three-dimensional data calculation means. With the above configuration, every time the moving body moves a predetermined distance, light is projected from the moving body toward the plane so that one irradiation line is formed on the plane along the transverse direction of the plane, and by the above-mentioned various means. A technique for acquiring an uneven profile in real time is described.

In such a road surface condition vehicle, the height of the road surface is measured by scanning light diagonally forward of the vehicle by the scanner of the measuring device while acquiring the position by GNSS (Global Navigation STtellite System). It is performed by irradiating in a spiral shape and receiving reflected light from surrounding structures. Here, the measuring device includes a plurality of measuring elements, for example, 32 measuring elements, and the measuring elements are rotationally driven to perform sequential scanning to acquire the distance to the surrounding structure and obtain the point group data over the entire circumference. ..

As a method of evaluating the unevenness of the road point cloud, a reference plane is calculated from the road point cloud, and how far each point of the road point cloud is from the reference plane is used as an index.

As a method for generating (fitting) a reference plane from the above-mentioned road point cloud, the least squares method (LS: Least Square Fitting: see Non-Patent Document 1) and principal component analysis (PCA: Principal Component Analysis: see Non-Patent Document 2). Is widely known. This principal component analysis, which is also called classical principal component analysis, has higher accuracy than the least squares method and is the mainstream method for plane fitting to a point cloud (see Non-Patent Document 2).

C. Gauss. Theoria Motus Corporum Coelestiumin Sectionibus Conicis Solem Ambientum Perthes, Hamburg, 1809. K. Klasing et al. Comparison of surfacenormal estimation methods for range sensing applications. In IEEE Int. Conf. Robot Autom., Pages 1977-1982. IEEE Press, 2009. D. Kraus and V. Panaretos. Dispersion operators resistant 2ndorder functional data analysis. Biometrika, 99 (4): 813-832, 2012. H. Cardot and A. Godichon. Fast estimation of the median covariation matrix with application toonline robust principal components analysis .. arXiv: 1504.02852, pages 1-46, 2016.

Japanese Unexamined Patent Publication No. 10-288516

However, the least squares method and the classical principal component analysis described above are known to be vulnerable to outliers. The point cloud data of the road obtained by the measuring device is due to the noise of the laser scanner itself used to acquire the point cloud data, road signs, electric poles, other vehicles, curbs, guard rails, etc. other than the road. There is a point cloud. Such noise and point clouds are outliers that have coordinates that are far from the road surface, and have an effect when fitting the reference plane to the road point cloud.

Therefore, the present invention provides point cloud data that can generate a reference plane without being affected by the noise from the point cloud data in which noise that is distributed almost in a plane and has extremely large outliers exists. It is an object of the present invention to provide a method for generating a reference plane and an apparatus.

The invention according to claim 1, which solves the above-mentioned problems, uses a planar point group data indicating the position of each point on the road acquired by a measuring device moved along a measurement path as an evaluation standard for the point group data. In this method of generating a reference plane, the median value of the point group data is obtained, the distance between each point of the point group data and the median value is calculated, and the larger the distance is, the smaller the value is. This is a method for generating a reference plane from point group data, which is characterized in that a reference plane is generated by obtaining the contribution ratio to be obtained and multiplying the contribution ratio.

Similarly, the invention according to claim 2 is characterized in that the contribution rate is inversely proportional to the distance in the reference plane generation method from the point cloud data according to claim 1.

Similarly, in the invention according to claim 3, in the reference plane generation method from the point group data according to claim 1, principal component analysis (PCA: Principal Component Analysis) is used to generate the reference plane, and PCA is calculated. The feature is that the geometric median covariance matrix (MCM) is used, the geometric median covariance matrix is updated by the weighted stochastic gradient descent method, and this update is repeated until convergence.

Similarly, in the invention according to claim 4, in the reference plane generation method from the point cloud data according to claim 3, when the update of the geometric median covariance matrix is repeated, the calculation is performed once for all the point clouds. It is characterized by randomly shuffling the reading order of all points each time it finishes.

Similarly, the invention according to claim 5 generates a reference plane as an evaluation standard of the point group data from the planar point group data indicating the position of each point on the road acquired by the measuring device moved along the measurement path. The means for obtaining the median value of the point group data, the means for obtaining the distance between each point of the point group data and the median value, and the larger the distance to each point, the smaller the value. It is a reference plane generation device for point group data, which comprises means for obtaining a contribution rate and means for generating a reference plane by multiplying the contribution rate.

According to the method for generating a reference plane to the point cloud data according to the present invention and the apparatus, the point cloud data is affected by the noise from the point cloud data which is distributed almost in a plane and has noise having extremely large outliers. A reference plane can be generated without any need.

That is, according to the reference plane generation method from the point group data according to claim 1 and the reference plane generation device for the point group data according to claim 5, the measurement device is moved along the measurement path. When generating a reference plane that serves as an evaluation standard for the point group data from the planar point group data indicating the position of each point on the road, the median value of the point group data is obtained, and each point and the median value of the point group data The distance to and from is obtained, the contribution rate that becomes smaller as the distance increases at each point is obtained, and the reference plane is generated by multiplying the contribution rate.
As a result, noise with large outliers is multiplied by a small contribution factor to reduce the effect on reference plane generation.

Further, according to the reference plane generation method from the point cloud data according to claim 2, the contribution rate is inversely proportional to the distance.
As a result, the noise with a large outlier becomes a small value in inverse proportion to the distance, and the influence on the generation of the reference plane becomes small.

Further, according to the reference plane generation method from the point group data according to claim 3, principal component analysis (PCA: Principal Component Analysis) is used to generate the reference plane, and the principal component analysis (PCA) is calculated. Using the Median Covariance Matrix (MCM), the geometric median covariance matrix is updated by the weighted stochastic gradient descent method, and this update is repeated until convergence.
As a result, the noise with a large outlier becomes a small value in inverse proportion to the distance, and the influence on the generation of the reference plane becomes small.

Then, according to the reference plane generation method from the point cloud data according to claim 4, when the geometric median covariance matrix is repeatedly updated, all points are obtained every time the calculation is completed for all the point clouds. Randomly shuffle the reading order of.
As a result, it is possible to prevent falling into a local solution.

It is a block diagram which shows the structure of the generation apparatus which concerns on embodiment of this invention. It is a schematic view which shows the measurement state of the road surface by the road surface measuring apparatus, (a) is a side view, (b) is a plan view. It is a schematic diagram which shows the structure arranged on the road. The schematic state of the point cloud data is shown, (a) is a perspective view schematically showing the distribution of the point cloud data in the unit region, and (b) is a schematic view showing the measurement data and the reference plane. It is a schematic diagram which shows the median value and the distance of a point cloud data. It is a flowchart which shows the process of the reference plane generation method from the point cloud data which concerns on embodiment of this invention. It is a figure which shows the mathematical expression for performing a plane fitting. The processing result of the reference plane generator to the point cloud data according to the embodiment of the present invention is shown, (a) is a schematic diagram showing the generated plane, and (b) is the plane in which noise is generated from the point cloud. It is a graph which showed the inclination of the plane generated from the point cloud with noise with respect to by the processing method.

Hereinafter, a reference plane generation method and an apparatus from the point cloud data according to the embodiment of the present invention will be described. In the reference plane generation method from the point group data according to the present invention, the median value of the planar point group data is obtained, the distance between each point of the point group data and the median value is obtained, and the larger the distance becomes, the larger the distance becomes. The contribution rate, which is a small value, is obtained, and each point is multiplied by the contribution rate to generate correction point group data, and a reference plane as an evaluation standard is generated based on the correction point group data.

In recent years, various mathematical models that extend the concept of median to tensors have been proposed, and the (geometric) median covariance matrix (MCM) is drawing attention for adaptation to PCA (non-patent). Reference 3). In the reference plane generation method from the point cloud data according to the present embodiment, the reference plane is efficiently estimated by applying the robust "weighted MCM-PCA" to the outliers to the large-scale data and performing the calculation. .. The outliers are assumed to have a large distance from the center of the geometry, and their reciprocals are used as the weights of MCM-PCA.

In this embodiment, a weighting coefficient is introduced into a known MCM-PCA (see Non-Patent Document 4), and the geometric median covariance matrix is updated by a weighted stochastic gradient descent method. This update is repeated until it converges, but the reading order of all points is randomly shuffled every time the calculation is completed for all point groups. This avoids falling into a local solution.

Hereinafter, the reference plane generator according to the embodiment of the present invention will be described. First, a general procedure from acquisition of point cloud data to analysis will be described. 1 is a block diagram showing a configuration of a generator according to an embodiment of the present invention, FIG. 2 is a schematic view showing a road surface measurement state by a road surface measuring device, FIG. 1A is a side view, and FIG. 2B is a plan view. , FIG. 3 is a schematic view showing a structure arranged on a road.

The point cloud data is acquired by the road surface measuring device 300. The road surface measuring device 300 constitutes an MMS (Mobile Mapping System: Mobile Mapping System). That is, the road surface measuring device 300 scans a predetermined range including the road, and points group data (road point group data) about the road and the structures around the road that the road surface measuring device 300 has passed through, and the road surface measuring device 300. The locus point sequence data showing the movement locus and the reference plane generation device 100 are output.

The reference plane generation device 100 sets a unit area UA as a unit for sequentially processing along the trajectory based on the acquired road point cloud data, and the area point cloud which is the road point cloud data for each unit area UA. Data is extracted, and a reference plane CP in this unit area is generated from the area point cloud data. The generated reference plane CP is sent to the road surface evaluation device 200 together with the road point cloud data and the locus point data. The road surface evaluation device 200 determines the state of the road by analyzing the reference plane CP, the road point cloud data, and the locus point sequence data sequentially acquired.

As shown in FIG. 2A, the road surface measuring device 300 is mounted on a traveling vehicle and acquires high-precision actual measurement data of the structure 400 by a scanner and an image. This measurement is performed around the road of the scanner 310, and the structure 400 includes a road 410, road accessories such as signals and road signs 420, a building 430, and a standing tree 440, and is a road surface measuring device. The 300 acquires point group data and images for these.

When evaluating a road, it is not possible to accurately set the reference plane of the road on which the point cloud data obtained for something other than the road exists. Therefore, the reference plane generation device 100 generates the reference plane by performing a process of reducing the influence of the point cloud data acquired from other than the road surface.

As shown in FIG. 2A, the road surface measuring device 300 is mounted on the vehicle 340 traveling on the road 410. The road surface measuring device 300 includes a scanner 310 which is a measuring device, an all-around camera 320, a GNSS (Global Navigation Satellite System) device 330, an attitude detection device of the road surface measuring device 300, an accelerometer, and the like. The road surface measuring device 300 spirally scans and irradiates the scan light La diagonally forward of the vehicle 340 with the scanner 310 while acquiring the position by the GNSS device, and receives the reflected light Lb from the structure 400, for example, the road 410. To do.

The road surface measuring device 300 acquires road measurement data (road point cloud data) based on the time until reception. Therefore, the locus T of the scan light La in the structure 400 has a spiral shape. Note that FIG. 2B shows only the scan light La irradiated on the road 410. 32 measuring elements are arranged in the scanner 310. This measuring element includes a light emitting element and a light receiving element, and the measuring light is emitted in a pulse shape from the light emitting element, and the light receiving element receives the reflection of the measured light by the structure 400. However, this reflected light may include reflected light from the road 410, the accessory 420, the building 430, and the standing tree 440, noise due to sunlight, and noise due to the device.

Further, the road surface measuring device 300 simultaneously acquires an image of the road over the entire circumference by the all-around camera 320. The GNSS device 330 captures the radio waves of the artificial satellite, acquires the plane position and altitude of the road surface measuring device 300, and acquires the traveling path of the road surface measuring device 300, that is, the measuring path. Then, the coordinates of the road surface measuring device 300 are acquired at regular time intervals, for example, locus points at intervals of 100 times / second, and the locus point sequence data is output as coordinates.

The reference plane generation device 100 sequentially transmits the long road point cloud data to the road surface evaluation device 200 for the reference plane CP for each unit area in the designated range, and the road point cloud data and the trajectory data. The road surface evaluation device 200 analyzes this point cloud data for each unit region UA and evaluates the road surface properties.

FIG. 4 shows a schematic state of the point cloud data, FIG. 4A is a perspective view schematically showing the distribution of the point cloud data in the unit region, and FIG. 5 is a schematic showing the median value and the distance of the point cloud data. It is a figure. As shown in FIG. 4A, the road surface evaluation device 200 receives the road point group data, the reference plane CP, and the locus point sequence data for each unit area UA, and as shown in FIG. 4B, the road surface evaluation device 200 receives the road point group data. The amount of distance from the reference plane CP of the point group data at each position is calculated, and this value is displayed as an image to evaluate the road.

Hereinafter, the reference plane generator 100 will be described in detail. As shown in FIG. 1, the reference plane generation device 100 includes a unit area cutting unit 110, an initial value calculation unit 120, a distance calculation unit 130, a contribution rate calculation unit 140, a median / covariance matrix update unit 150, and a plane. A generation unit 160 is provided.

The reference plane generation device 100 according to the present embodiment has a CPU (Central Processing Unit) as a processing device, a RAM (Random Access Memory) and a ROM (Read Only Memory) as a main storage device, and an HDD (Hard Disk Drive) as an auxiliary storage device. It is configured as a computer equipped with such as. In the reference plane generation device 100, the means realized by the reference plane generation device 100 by executing the program by the CPU, that is, the unit area cutting unit 110, the initial value calculation unit 120, the distance calculation unit 130, the contribution rate calculation unit 140, The functions of the median / covariance matrix update unit 150 and the plane generation unit 160 are realized. The reference plane generation device 100 can be realized by a notebook-type personal computer in which a program for realizing a reference plane generation method from point cloud data is installed.

The unit area cutting unit 110 cuts out the road point cloud data acquired from the road surface measuring device 300 for each unit area UA and outputs it as area point cloud data. The size of the unit area UA can be set as needed, and can be, for example, 4 m in the vehicle width direction and 3 m in the measurement direction.

The initial value calculation unit 120 calculates the geometric median of the area point cloud data. The geometric median value C is a value that is the center of the geometric coordinate values of the distributed road point cloud data for the three-dimensional coordinates based on the geodetic coordinate system. This is saved as the initial value C1 of the geometric median. Further, the initial value V1 of the geometric median covariance matrix is calculated and stored using the obtained geometric median value C. The output coordinate system of the point cloud data differs depending on the geodetic system, and there are various methods such as the case of longitude / latitude, the method of setting the reference point as the origin, the x direction as the east, the y direction as the north, and the z direction as the altitude.

The distance calculation unit 130 obtains the distance Wn between the geometric median Cn and each point Pn of the area point cloud data (see FIG. 5). Note that the point cloud data may be obtained in a system other than the Cartesian coordinates (Cartesian coordinates) such as polar coordinates, and the distance between points is not limited to the Pythagorean theorem and is determined by a predetermined method.

The contribution rate calculation unit 140 calculates the contribution rate. The contribution rate is set to a smaller value as the distance increases for each point, for example, the reciprocal of the distance (1 / Wn). As the contribution rate, an arbitrary function can be used in which the value decreases as Wn increases. This time, the contribution rate is calculated by multiplying the reciprocal of the distance from the median of each point by an arbitrary constant coefficient, but an exponent or logarithm or a more complicated function can also be used. .. The function representing the contribution rate may have the property that "the contribution rate decreases as the distance from the median increases".

The median / covariance matrix update unit 150 updates the geometric median covariance matrix V while imparting a contribution rate by a weighted stochastic gradient descent method, and repeats this update until convergence. As a result, a reference plane CP in which the influence of the noise component is reduced can be generated. Here, the "weighted stochastic gradient descent method" introduces weighting into the stochastic gradient descent method, which is known as a solution of optimization problems, to reduce the influence on outliers and avoid falling into a local solution. It is a process.

The plane generator 160 uses the median / covariance matrix updated by the median / covariance matrix updater 150 instead of the geometric mean and covariance matrix used in the classical PCA. , Calculate the geometric mean CP.

The above processing flow can be summarized as follows. FIG. 6 is a flowchart showing the processing of the reference plane generation method from the point cloud data according to the embodiment of the present invention. Here, the unit area UA is defined and the area point cloud data is acquired. The reference plane generation device 100 reads the unit area point cloud data (point cloud data P: point N) from the road surface measuring device 300 (step ST1). Then, the initial value calculation unit 120 calculates the geometric median value C from the coordinates of all the points in the point cloud data P. This is saved as the initial value C1 of the geometric median. Further, the initial value V1 of the geometric median covariance matrix is calculated and saved using the obtained geometric median value C (step ST2).

Then, the processes of step ST4 and step ST5 are repeated according to n = 1; n ≦ N; n ++ (steps ST3 to ST7). That is, the distance calculation unit 130 calculates the distance Wn between the point Pn of the point cloud data P and the geometric median Cn (step ST4). Then, the contribution rate un is calculated by the contribution rate calculation unit 140 (step ST5), and the geometric median value Cn is calculated by the median / covariance matrix update unit 150 based on the weighted probabilistic gradient descent method for multiplying the contribution rate un. The geometric median covariance matrix Vn is updated (step ST6).

Then, the geometric median values Cn and Vn are updated for N points, and the obtained geometric median value is CN and the geometric median covariance matrix is VN. Then, the obtained VN and CN are stored (step ST8). Then, the initial value V1 of the geometric median covariance matrix and the root mean square of each component of VN are calculated (step ST9), and if this is not less than the predetermined convergence determination threshold value ε, the point group is subjected to step ST10. Randomly shuffle the read order of each point. Then, CN and VN are substituted into the initial values C1 and V1, respectively (step ST11), and steps ST3 to ST7 are repeated. This gives a median and covariance matrix in which the effects of outliers are corrected by the contribution of each point.

Then, the plane generation unit 160 generates a reference plane CP by performing plane fitting by, for example, PCA, using the geometric median and the geometric median covariance matrix updated until convergence (step ST12), and saves the result. (Step ST13). Other methods can be used for plane fitting.

Next, a calculation method for a series of processes up to the generation of the reference plane will be described. FIG. 7 shows a mathematical formula for updating the geometric median and the geometric median covariance matrix. Equations 1 to 3 obtain the initial value V1 of the geometric median covariance matrix and the initial value C1 of the geometric median value, and correspond to the initial value calculation unit 120.

Equation 4 in FIG. 7 is a distance calculation unit 130, equation 5 is a contribution ratio calculation unit 140, and equations 6 to 9 are calculation formulas in which the processing of the median / covariance matrix update unit 150 is repeated for each point. n is the number of a point in the point cloud, and data is read in no particular order from n = 1 to n = N (N is the number of all points), and the value is updated for each point. This is repeated until n = N, and the VN and CN at that stage are stored.

The initial values V1 and C1 are substituted as VN and CN, and the update process is repeated for all points in the same manner. At this time, the reading order of the points is shuffled at random. By randomly shuffling the reading order for each round of the iterative process in this way, it is possible to avoid falling into an inappropriate local solution and converge to the optimum solution. When converged, the difference in the sum of squares of the differences between the components of V1 and VN becomes very small. When it falls below the predetermined convergence test threshold value ε, it is regarded as having converged and the iterative calculation is terminated.

Next, the result of the processing by the reference plane generator 100 will be described. FIG. 8 shows the processing result of the reference plane generator for the point cloud data according to the embodiment of the present invention, (a) is a schematic diagram showing the generated plane, and (b) is noise from the point cloud. It is a graph which showed the inclination of the plane generated from the point cloud with noise with respect to the generated plane by the processing method.

In the figure shown in FIG. 8A, the processing target is the point cloud data P distributed in the square unit region UA, and the point cloud data P includes a large amount of impulse noise (10% of the total: the amplitude is random). Is added. The reference plane CP was obtained under these conditions. The normal of the reference plane CP is indicated by N.

8 (a) and 8 (i) show the case where the reference plane generator 100 according to the embodiment processed, and FIG. 8 (ii) shows the case where the process was performed by the conventional method. In the reference plane generator 100 according to the embodiment for exactly the same point cloud data Pn, the reference plane CP is substantially along the unit region UA plane, whereas in the conventional method, the reference plane CP is the unit region UA. It can be seen that it is arranged in the direction intersecting with and is greatly affected by noise.

8 (b) and 8 (i) compare the results of the present embodiment and the conventional example when the number of impulse noises is changed and the amplitude of the noise is changed in the same (ii). In each graph, the vertical axis shows the difference in normal angle from the ideal plane without noise. In the graph, "W-MCM" is the process according to the present embodiment, and "Our" is the process according to the embodiment. Therefore, in order to compress the memory consumption, a part of the point cloud data is temporarily transferred from the RAM to the HDD during the process. It was saved and implemented. Although the calculation method is the same, it shows that it can handle very large point cloud data that cannot be stored in RAM.

According to each figure, embodiments of the present invention are known existing methods such as Least Square Fitting (LSF), Ridge Regression (RR), Principal Component Analysis (PCA), and Weiszfeld method. Even if the number of noises and noise amplitude are increased compared to each processing method of PCA (MCM-PCA: MedianCovariation Matrix PCA) using geometric median, the arrangement angle of the generated reference plane CP does not change and noise. It can be seen that the effect of

100: Reference plane generation device 110: Unit area cutting unit 120: Initial value calculation unit 130: Distance calculation unit 140: Contribution rate calculation unit 150: Geometric center value / covariance matrix update unit 160: Plane generation unit 200: Road surface evaluation Device 300: Road surface measuring device 310: Scanner 320: All-around camera 330: GNSS device 340: Vehicle 400: Structure

Claims (5)

It is a method of generating a reference plane as an evaluation standard of the point cloud data from the planar point cloud data indicating the position of each point of the road acquired by the measuring device moved along the measurement path.
Find the median of the point cloud data
Find the distance between each point in the point cloud data and the median,
The contribution rate, which becomes smaller as the distance increases, is obtained for each of the points.
A method for generating a reference plane from point cloud data, which comprises multiplying the contribution rate to generate a reference plane. The method for generating a reference plane from point cloud data according to claim 1, wherein the contribution rate is inversely proportional to the distance. Principal component analysis (PCA) is used to generate the reference plane, and a geometric median covariance matrix (MCM) is used to calculate the principal component analysis (PCA). The method for generating a reference plane from the point group data according to claim 1, wherein the geometric median covariance matrix is updated by the target gradient descent method, and this update is repeated until convergence. The point according to claim 3, wherein when the update of the geometric median covariance matrix is repeated, the reading order of all points is randomly shuffled every time the calculation is completed once for all the point clouds. How to generate a reference plane from point cloud data. A device that generates a reference plane as an evaluation standard for the point cloud data from the planar point cloud data indicating the position of each point on the road acquired by the measuring device moved along the measurement path.
A means for obtaining the median value of the point cloud data,
A means for obtaining the distance between each point of the point cloud data and the median,
A means for obtaining a contribution rate that becomes a smaller value as the distance increases at each of the points.
A means of multiplying the contribution ratio to generate a reference plane, and
A reference plane generator for point cloud data, which comprises.