WO2014189059A1

WO2014189059A1 - Road surface condition estimating device

Info

Publication number: WO2014189059A1
Application number: PCT/JP2014/063388
Authority: WO
Inventors: 俊樹磯貝; 須田　義大; 雅彦安藝; 中野　公彦; ティーラパットロジャナアーパー
Original assignee: 株式会社デンソー; 国立大学法人東京大学
Priority date: 2013-05-20
Filing date: 2014-05-20
Publication date: 2014-11-27
Also published as: JP2014228300A; JP6040096B2

Abstract

The present invention is provided with a light irradiation unit for irradiating laser light so as to scan a road surface, a light reception unit for acquiring reflection intensity values corresponding to each irradiated point on the road surface, and a road surface condition determination device for determining a road surface condition using the reflection intensity values. A set of reflection intensity values acquired for each irradiated point on a road surface (hereafter referred to as a "multipoint reflection intensity value set") defines the relationship between the irradiation positions in the far-near direction from the light irradiation unit and the reflection intensity values, which are dependent on the road condition. The road surface condition determination device is created through machine learning using, as learned data, a plurality of multipoint reflection intensity value sets acquired for a plurality of road surfaces having different known road surface conditions and determines and outputs a road surface condition using, as input, a multipoint reflection intensity value set acquired for a road surface having an unknown road surface condition.

Description

Road surface condition estimation device

The present invention relates to a road surface state estimating device for estimating the state of a road surface on which a vehicle travels.

As part of the development of energy-saving technology, the Energy ITS (Intelligent Transportation Systems) project is underway, and one of them is the technology development of automatic platooning (Patent Document 1, Non-Patent Documents 1 to 3). The purpose of this project is to reduce air resistance and save energy by shortening the inter-vehicle distance in platooning. The white line recognition algorithm (Patent Document 1, Non-Patent Document 2) in this project is used to control the lateral displacement of the front and rear vehicles in the platoon within a predetermined range in order to obtain fuel saving efficiency with a short inter-vehicle distance. .

In practical application of inter-vehicle distance control in automatic platooning with automatic steering and automatic speed control, braking control corresponding to sudden changes in road surface conditions is required. In automatic platooning, for example, a control gain tuned according to the road surface condition is created, and when the driving environment changes from a dry road surface to a wet, flooded, frozen road surface, etc. It is useful for ensuring safety to change the control gain instantaneously in response to the change.

In general, the driving force, braking force, and steering force of vehicles such as automobiles depend on the friction between the road surface and tires, and estimating the road surface condition is a useful technique for ensuring driving performance and improving safety. . As a road surface state estimation technique, various methods have been proposed so far, for example, a method utilizing laser radar reflection intensity, a method utilizing camera image, and the like (Non-Patent Documents 4 to 7). However, in particular, a method suitable for an autonomous driving vehicle has not been established.

JP2011-210165A

The present inventors have developed a new road surface state estimation technique using a laser radar used for white line recognition in automatic driving and platooning. Whereas the conventional road surface state estimation method using a laser radar employs a method that uses only the reflection intensity, the present invention uses a multipoint reflection intensity value to determine the positional relationship and the road surface state of laser light reflection. The mechanical properties that depend on The present invention was created in connection with the white line recognition technology, but the technical idea according to the present invention is not limited to automatic platooning or white line recognition technology, but is generalized as road surface state estimation technology. It can be applied to general estimation of road surface conditions.

The present invention provides a new road surface state estimation apparatus using machine learning.
One of the objects of the present invention is to provide a road surface state estimation device that can be satisfactorily applied to an autonomous driving vehicle.
Other objects and more specific objects of the present invention will become apparent from the description of the present specification and the drawings.

The road surface state estimation device adopted by the present invention includes a light projecting unit that irradiates a laser beam so as to scan the road surface, a light receiving unit that acquires a reflection intensity value corresponding to each irradiation point on the road surface, and a reflection intensity value And a road surface state discriminator that discriminates the road surface state using A set of reflection intensity values acquired corresponding to each irradiation point on the road surface (referred to as a “multi-point reflection intensity value set”) is the irradiation position and reflection in the perspective direction based on the light projecting part depending on the road surface condition. Defines the relationship with intensity values. The road surface state discriminator is generated by machine learning using a plurality of multipoint reflection intensity value sets respectively acquired for a plurality of road surfaces having different known road surface states as learning data. The road surface condition discriminator discriminates and outputs a road surface condition with a multipoint reflection intensity value set acquired for a road surface whose road surface condition is unknown as an input.

Although the laser beam irradiation / scanning direction is not limited based on the principle of the present invention, in one aspect, the road surface state estimation device is mounted on a vehicle traveling on a road surface, and the laser light crosses the road surface. Irradiated to scan in the direction.
In one aspect, the irradiation range of the laser beam includes a lane marking (white line in a typical example) extending in the traveling direction of the road surface, and the multipoint reflection intensity value acquired for a road surface whose road surface state is unknown. The set includes a reflection intensity value corresponding to the lane line, and the road surface state discriminator is a reflection obtained by removing the reflection intensity value corresponding to the lane line from the acquired multipoint reflection intensity value set. The road surface condition is determined using the intensity value set.
In this case, in one aspect, the determination is performed after the removed data is interpolated. Since data interpolation methods (interpolation, extrapolation) are known to those skilled in the art, description thereof is omitted. Alternatively, the road surface condition may be determined while the removed data is missing.
Further, a feature amount obtained from the reflection intensity value corresponding to the lane marking (for example, a difference in reflection intensity value between a white line portion and a portion other than the white line, a road surface roughness index in the white line portion, etc.) may be used. .

In one aspect, an analysis unit is provided that acquires a road surface roughness index depending on a road surface state by performing the multipoint reflection intensity value set analysis. The learning data includes a plurality of road surface roughness indices respectively acquired for a plurality of road surfaces having different known road surface conditions. The road surface state discriminator receives the multipoint reflection intensity value set and road surface roughness index acquired for a road surface whose road surface state is unknown, and determines and outputs the road surface state.
In one aspect, the analysis means is a discrete wavelet analysis.
Although the processing speed is slower than that of discrete wavelet analysis, it is possible to use continuous wavelet analysis. Further, the road surface roughness index may be acquired by Fourier analysis.
As described above, in one aspect, the “multi-point reflection intensity value set” is used as the first feature quantity and the “road surface roughness index” is used as the second feature quantity. In the present invention, the second feature quantity is preferable optional information. However, it is not essential information. It is not prohibited to use an index other than the road surface roughness index as the second feature amount depending on the road surface state. That is, as the second feature amount, the type and number of information can be arbitrarily selected, and a plurality of feature amounts may be adopted as the second feature amount.
Moreover, since the road surface state estimation technique has been studied for many years, various knowledge has been obtained, and it is optional to combine other estimation means and knowledge with the present invention.

In one aspect, the road surface state discriminator is composed of a naive Bayes estimator.
In one aspect, the said road surface state discriminator is comprised from the neural network.
As machine learning that can be applied to the road surface state discriminator of the present invention, any method of “machine learning with a teacher signal” can be applied. For example, clustering, discriminant analysis (linear discriminant function or Mahalanobis distance), logistic Regression analysis, support vector machines, decision trees, etc. are applicable.

In one aspect, the road surface state discriminator discriminates at least a dry state and a wet state.
In one aspect, the road surface condition discriminator further discriminates a flood condition.
In one aspect, the said road surface state discriminator further discriminate | determines the degradation state of a road surface.
In one aspect, six road surface states that are combinations of the above states are determined.
In a typical embodiment, the road surface is an asphalt pavement surface.
The road surface state and the road surface material that are the subject of the present invention are not limited to these. Since the present invention uses a road surface state discriminator generated by machine learning, the road surface state discriminator is not limited to the kind of road surface state and the material of the road surface, and the road surface state discriminator is obtained using learning data acquired for a road surface in a certain state. It only has to be generated.

The present invention provides a new road surface state estimation apparatus using machine learning.
The road surface state estimation device according to the present invention can be mounted on a vehicle, can estimate a road surface state in real time while acquiring reflection intensity data while the vehicle is traveling, and can be applied well to an autonomous driving vehicle. is there.
In the present invention, by acquiring learning data of a predetermined road surface state and generating a road surface state discriminator, as in an embodiment described later, friction such as asphalt road surface deterioration state, water lubrication state, water film presence, etc. The state depending on the coefficient can be determined using the reflection intensity acquired by the laser radar.
The present invention can also estimate a road surface condition by using an output of a white line detecting laser radar device in a vehicle that recognizes a white line with a laser radar and performs automatic traveling.

The figure on the left shows the relationship between reflection intensity and lateral distance. The right figure is a figure which shows the aspect of laser beam irradiation. It is the schematic which shows the flow of the whole road surface state estimation apparatus which concerns on this embodiment. It is a block diagram which shows the production | generation of a road surface state estimator, and discrimination | determination of a road surface state. It is a figure which shows the reflective characteristic relevant to a dry asphalt road surface / wet asphalt road surface. It is a figure which shows the reflective characteristic relevant to the absorption of the laser beam on a road surface. It is a figure which shows the reflective characteristic relevant to a new asphalt road surface / aged asphalt road surface. The reflection characteristics depending on the road surface condition are arranged in dry, wet, submerged, vehicle distal region and vehicle proximal region for aged asphalt and new asphalt, respectively. FIG. 8 is a graph showing the relationship between the reflection intensity and the scanning position (position of each irradiation point) in FIG. It is a figure explaining decomposition | disassembly of a signal. A is a low frequency signal and D is a high frequency signal. Fig. 4 shows multi-stage signal decomposition using a filter bank. The frequency bands separated in each decomposition stage are shown. It is a figure explaining decomposition | disassembly of the reflection intensity (dry state) by discrete wavelet analysis. It is a figure which shows the whole flow of the road surface state estimation apparatus which concerns on this embodiment. It is a schematic diagram of a test device. It is a figure explaining acquisition of the reflection intensity value by a laser radar. It is a figure which shows the example of measurement data. It is a figure which shows the relationship between reflection intensity and a horizontal direction distance about the measured reflection intensity value (multipoint reflection intensity value set). It is a figure which shows the road surface roughness parameter | index acquired about the dry road surface and the wet road surface. It is a figure which shows the structural example of a neural network.

[A] Outline of the Present Invention A road surface state estimation apparatus according to the present invention is a light projecting unit that irradiates a laser beam so as to scan on a road surface, and a light receiving unit that acquires a reflection intensity value corresponding to each irradiation point on the road surface. And a road surface state discriminator that discriminates the road surface state using the reflection intensity value. The road surface state estimation apparatus includes a two-dimensional scan type laser radar (LIDAR), a computer (an input unit, an output unit, a storage unit such as a RAM and a ROM, a processing unit mainly composed of a CPU) and the like as hardware configurations. ) And.

The laser radar includes a light projecting unit and a light receiving unit. The light projecting unit includes a laser diode, a driving unit for emitting laser light from the laser diode, and an optical system for scanning (such as a mirror). The light receiving unit includes a photodetector (photodiode) that detects reflected light from the measurement target (road surface), and the reflection intensity is acquired from the output of the photodiode. Since the configuration of the laser radar itself is well known to those skilled in the art, a detailed description thereof will be omitted.

The road surface state discriminator can be composed of a computer. The road surface state discriminator is generated by machine learning using multiple multipoint reflection intensity value sets acquired for multiple road surfaces with different known road surface states as learning data, and acquired for road surfaces with unknown road surface states Using the set multi-point reflection intensity value as an input, the road surface condition is determined and output.

From the continuous reflection intensity values acquired corresponding to each continuous irradiation point on the road surface (the scanning interval in the embodiment described later is 0.0045 m. The resolution may be further increased). The multipoint reflection intensity value set (in the embodiment described later, the reflection intensity value at 171 points) is the irradiation position and reflection intensity in the perspective direction (along the scanning line) with reference to the light projecting portion depending on the road surface state. Defines the relationship with the value. In the present invention, a large number of reflection intensity values (multi-point reflection intensity value sets) at a large number of irradiation points are used for machine learning, and measurement data comprising the same multi-point reflection intensity value set is used as an input to classify road surface conditions. One of the features is that

[B] Road Surface State Estimation Using Laser Radar [B-1] Outline of Road Surface State Estimation Method An embodiment of the present invention will be described in detail. The purpose of this embodiment is to perform dry / wet determination in real time, and to estimate the road surface condition using a laser radar used for white line recognition in automatic driving and platooning. In the present embodiment, a road surface state estimation algorithm for determining dryness and wetness of a road surface by machine learning using laser radar reflection intensity and road surface roughness component information of the reflection intensity is proposed, and a road surface state estimation experiment is performed using this algorithm. went.

As shown in the figure on the right side of FIG. 1, a laser radar is mounted on the vehicle, and the road surface is irradiated while scanning the laser beam obliquely downward in the lateral width direction of the vehicle from the light projecting unit. The laser beam is scanned and transmitted over a predetermined range including the width of the white line in a direction crossing the road surface toward the side of the vehicle. The laser reflected light reflected by the road surface is received by the light receiving unit, and the reflection intensity corresponding to each irradiation position of the laser light is acquired. By utilizing this reflection intensity information, an integrated system that performs both white line recognition and road surface recognition is constructed.

In the technical development of automated driving platooning, white line recognition is performed by measuring the reflection intensity of a white line using a laser radar (Patent Document 1, Non-Patent Document 2). This white line recognition uses a difference in reflection intensity to detect a white line because the white line and road asphalt have different reflectivities. Since the white line is mixed with glass beads, the reflection intensity of the white line part is larger than the reflection intensity of the asphalt part due to retroreflection of the glass beads. Therefore, when the horizontal distance is taken along the horizontal axis (scanning direction) and the reflection intensity is taken along the vertical axis, the left side of FIG. 1 is obtained. The lateral distance X can be calculated by X = H * tan (laser radar mounting angle + scanning angle) where the laser radar installation height is H. For details of white line recognition, Patent Literature 1 and Non-Patent Literature 2 can be referred to. In this embodiment, the road surface state can be estimated simultaneously using information acquired in white line recognition, and no new device or component is required.

As shown in FIG. 2, the road surface state estimation method according to the present embodiment uses a laser radar reflection intensity and a road surface roughness component extracted from the reflection intensity, and wavelet decomposition is used to extract the road surface roughness component. The machine learning is used for road surface discrimination. Aiming to distinguish six types of road surface conditions: dry, wet, and submerged, and two types of road surface quality: new asphalt and aged asphalt. An overall view of a road surface state discriminator using machine learning is shown in FIG.

The road surface state estimation method according to the present embodiment considers the positional relationship and the mechanical properties of laser light reflection by using multipoint reflection intensity values included in the measurement range. Depending on the road surface condition, it is conceivable that the reflection intensity value approaches at different places between different road surfaces. Therefore, the feature amount is further increased using the road surface roughness component as an index, and this road surface roughness index is classified by the spatial frequency of the road surface. Therefore, what was calculated using wavelet was defined.

The road surface state is estimated by using a naive Bayes estimator as a machine learning method for the feature quantity prepared in this way. While traveling, the number of feature values of multipoint reflection intensity values used for road surface state estimation may differ depending on the presence or absence of a white line (if there is a white line, feature quantity data corresponding to the white line is removed, as will be described later) ) The naive Bayes estimator can be applied well even if there is a partial loss of feature data.

[B-2] Road surface state estimation feature quantity (1) Laser radar reflection intensity (first feature quantity)
It is known that the laser radar reflection intensity varies depending on the road surface quality / state and the incident angle on the road surface. First, the mechanical properties of the laser radar reflection intensity are summarized. The laser radar scans the road surface downward in the lateral direction with respect to the traveling direction of the vehicle. We consider three points: reflection of the laser beam by the water film, absorption by the water film, and reflection by the road surface.

First, the reflection characteristics of a laser beam by a water film will be described. On the dry road surface, the laser light is reflected in all directions by the scattering effect, but on the road surface covered with the water film, the reflected light has directionality due to regular reflection in addition to the scattering effect. Therefore, although the reflection intensity is low at the distal side (the side far from the light source), strong laser light reflection due to regular reflection occurs at the proximal side (side near the light source). This reflection intensity characteristic is shown in FIG.

Second, the absorption and reflection characteristics to the water film will be described. Reflection intensity is high because there is no absorption into the water film on the dry road surface. However, on the road surface covered with the water film, laser light is absorbed into the water film regardless of whether it is distal or proximal. The reflection intensity is lower than that of the road surface. This reflection intensity characteristic is shown in FIG.

Thirdly, the reflection characteristics of the road surface will be described. At the beginning of asphalt laying, the particles were large and the road surface was uneven, but the asphalt particle unevenness was leveled as the vehicle traveled. Along with this, the laser light is diffused by unevenness at the beginning of asphalt laying and the reflection intensity is low, but the reflection intensity is increased by reducing the laser light dissipation due to the unevenness by becoming asphalt over time. This reflection intensity characteristic is shown in FIG.

反射 Thus, the reflection state varies depending on the mechanical properties of the road surface. For aged asphalt and new asphalt, organize in dry, wet, submerged, and vehicle distal and vehicle proximal areas. First, when comparing a new asphalt and an aged asphalt, as described above, the new asphalt has a higher reflection intensity than the aged asphalt. Also, compared to dry road surfaces, the wet road surface absorbs laser light into the water film, resulting in a decrease in reflection intensity.However, when the water film becomes thicker and becomes a flooded road surface, the reflection intensity approaches that of regular road reflection. To rise. In addition to these characteristics, the reflection intensity tends to gradually decrease as the distance from the laser beam emitting portion of the laser radar increases. However, the road surface condition (flooding / aging asphalt, wet / In the case of new asphalt, submersion / new asphalt), the farther away, the more the reflection intensity decreases. FIG. 7 summarizes these relationships. FIG. 8 is a graph showing the relationship between the reflection intensity and the scanning position. In FIG. 8, the farthest data within the effective data range is taken as the origin, and the range up to the laser radar position is taken as the range of the horizontal axis. Thus, the reflection intensity can be expressed as a function of position by the mechanical properties of the road surface.

(2) Road surface roughness index (second feature value)
The difference in the laser reflection angle is caused by the unevenness of the road surface roughness, and as a result, the reflection intensity value also includes a wavelength component corresponding to the road surface wavelength in the scanning direction. More specifically, scanning is performed according to the characteristics of the spatial frequency (with the addition of laser reflection characteristics): “The larger the macro, the greater the undulation of the entire road surface”, and “The smaller the surface, the micro (particle) unevenness of the road surface”. It is measured as the difference in size of the reflection intensity value in the direction (see FIG. 15). This wavelength component is extracted by subjecting the reflection intensity value to discrete wavelet decomposition (FIGS. 9A to 9D). It is also possible to acquire a road surface roughness index using continuous wavelet analysis. By using discrete wavelet analysis, the amount of calculation can be reduced compared to continuous wavelet analysis. The reduction in the calculation amount leads to a reduction in processing cost. In the discrete wavelet analysis according to the present embodiment, the Mallat algorithm is used.

FIG. 9A is a diagram for explaining signal decomposition. A is a low frequency signal and D is a high frequency signal. FIG. 9B shows multistage signal decomposition using a filter bank. At each stage, the low frequency component and the high frequency component are separated. FIG. 9C shows the frequency bands that are separated at each decomposition stage when the original wavelet is separated three times. f _n is the center frequency of the mother wavelet. The wavelet frequency is usually calculated from a pseudo frequency corresponding to the scale. The relationship between frequency and scale is as follows.

Fa: pseudo frequency, Fn: center frequency of wavelet (0.66667 Hz in 'db2'), a: scale, Δ: sampling period (distance between scanning positions is 0.0045 m).
FIG. 9D shows decomposition of the reflection intensity (dry state) by the discrete wavelet analysis according to this embodiment. Here, the decomposition is performed five times. Signal D1 represents road roughness at fine details and signal D5 represents road roughness at low resolution. The spatial frequencies included in each high frequency signal are shown in Table 3 (described later).
In addition to using the wavelet-decomposed data for road surface roughness index calculation, it is also possible to integrate white line detection by this series of measurements by using it for white line edge detection.

In this embodiment, the road surface roughness index is obtained by taking the sum of the wavelength bands to be included as the road surface roughness. Therefore, data with n reflection intensity values can be obtained in each scan, whereas one data with road surface roughness can be obtained. By adding this road surface roughness index as a road surface feature amount as a variable after the multipoint reflection intensity value, the road surface state estimation accuracy is improved. Table 1 summarizes laser light absorption, total reflection, and road surface roughness indices for six road surface conditions. In Table 1, ◯ indicates strong, X indicates weak, and Δ indicates an intermediate value.

[B-3] Road surface recognition by naive Bayes estimator The naive Bayes estimator is used as supervised machine learning, and is effective when data according to the Bayes distribution is obtained for each class to be distinguished. It is a technique.

First, take the following steps to create teacher data.
(1) A reflection intensity value and a road surface roughness index obtained as a result of discrete wavelet decomposition are calculated from the obtained laser radar data.
(2) For each of the six classes to be classified, the average value μ _i and the standard deviation σ _i of the feature variables (f _i ) are calculated (i = 1 to N, N = n + m). At this time, n is the number of reflection intensities, m is the number of road surface roughness indices, and this time, n = 171 and m = 1, respectively. At this time, the multipoint reflection intensity value of 171 points expresses the scanning position and reflection intensity from the distal to the proximal as function values. The time when the white line is not included in the scanning range is 171 data, and when the white line is included in the scanning range, the processing removes the white line intensity at the scanning position corresponding to the white line from the estimated range. Since the number of feature variables can be changed in this way, a naive Bayes estimator is used as a machine learning method that can cope with the change. Furthermore, it aims at improving estimation accuracy by adding a road surface roughness index after the multipoint reflection intensity value.

Using the teacher data obtained as described above, the measurement road surfaces are classified by calculating equations (1) and (2) in real time.

[B-4] Road surface state estimation system Road surface state estimation is performed in real time using each scan data. The flow of data processing is shown in FIG.

(1) First, the position of the white line is recognized based on the road surface obtained from the laser radar and the white line reflection intensity value (see Patent Document 1 and Non-Patent Document 2).
(2) The white line portion is excluded from the road surface state estimation region based on the white line position information obtained by white line detection based on Patent Document 1 and Non-Patent Document 2. Specifically, from the number of data included in the current scanning range (171 points of data), if a white line is included, the number of data for the white line region (l) is subtracted. The number of multipoint reflection intensity values determined as a result is first set as the number of feature amounts defined at that time.
(3) A reflection intensity value (n−1, where l is the number of data on the white line) and a road surface roughness index (m) based on the reflection intensity value are calculated from the measurement data.
(4) Finally, the road surface is classified by the naive Bayes estimator.

[C] Road surface condition estimation test on Ube Industries private road [C-1] Construction of measurement environment Measurement tests were carried out on Ube Industries private road for about 6 months for sensor system evaluation in the Energy ITS Project (Non-patent Document 3) ). Among them, a road surface condition estimation test was conducted by laser radar. An overall configuration diagram of the test apparatus is shown in FIG. Information from a fail-safe ECU (FS-ECU), information from a white line recognition device using laser radar (active method), information from a road surface state estimation device using laser data, white line recognition device using image processing ( Information from the passive system), information from the GPS, information from the WEB camera, and information from the thermohygrometer are input to the data processing device. The data processing apparatus can be configured by a computer (including a hardware configuration such as an input unit, an output unit, a storage unit, and a calculation unit, and predetermined software). As described below, the test mainly serves as a data storage device that stores various data. In the test, as shown in FIG. 12, a laser radar road surface state estimation system and a data logger device for collecting these data were mounted on a trailer type truck, and measurements were taken during business travel. Laser radar used for white line recognition and road surface condition estimation is installed on the ceiling of the truck. In addition, a GPS receiver, a weather recording camera, and a temperature / humidity sensor are installed for environmental recording when performing long-term measurement, and these data are recorded in a data logger.

As the laser radar, a scanning laser radar manufactured by Denso Corporation was used. The specifications of the laser radar are as follows. It should be noted that these specifications are only examples of laser radars that can be used in the present invention.

The number of scanning points on one scanning line in the maximum detection area (± 18 [deg]) of the laser radar is 451 points. In this embodiment, as shown in FIG. 12, 17.02-3.42 Data of 171 points corresponding to the irradiation angle of [deg] is acquired. The distance in the horizontal direction corresponding to 171 points is 0.7641 m. The reflected light corresponding to each of the 171 points (irradiation positions) is received, and the reflection intensity at 171 points is acquired.

The measurement vehicle travels on the road surface at 50-60 km / h, and the sampling frequency of the laser radar is 10 Hz. An example of the reflection intensity result obtained by measuring the road surface with the laser radar at this time is shown in FIG. The horizontal axis is the horizontal distance, and the vertical axis is the time. Although the reflection intensity is actually displayed in a color scale (luminance value), it is in gray scale due to limitations of the patent drawing. FIG. 13 is a reflection intensity map obtained by 100 scans acquired during a 10-second run on an aged asphalt / dry road surface. The portion with high density represents the reflection intensity from the white line portion.

The six conditions estimated in this test are “Aged asphalt / dry road surface”, “Aged asphalt / wet road surface”, “Aged asphalt / submersion road surface”, “New asphalt / dry road surface”, “New asphalt / wet road surface”. There are six types of "new asphalt / flooding surface".

(1) Reflection Intensity FIG. 14 shows the reflection intensities obtained as shown in FIG. 13 in which the horizontal axis represents the scan position and the vertical axis represents the reflection intensity. The reflection intensity on the road surface under the six conditions shows such a value with respect to the scan position. The reflection intensity of dry / new asphalt and wet / aged asphalt, and the reflection intensity of wet / new asphalt and submersion / new asphalt show similar values, but it can be confirmed that they are separated under other conditions.

(2) Road Surface Roughness Index Since it may be difficult to discriminate between road surface conditions only with the reflection intensity as shown in FIG. 14, a more accurate discrimination can be performed by adding a road surface roughness index. This time, there is one road surface roughness index, and it is added to the n reflection intensity values. This road surface roughness index is calculated in real time from raw data obtained by each scan. Roughness data in each scan is extracted by discrete wavelet decomposition, and a roughness index is calculated. At this time, road surface roughness wavelengths corresponding to D1 to D4 were adopted, and the sum of these was taken as a road surface roughness index. FIG. 15 shows the result calculated as the sum of D1 to D4. The road surface roughness index is a sum in the horizontal direction in FIG. 15 and corresponds to taking an area of variation in reflection intensity.

[C-2] Road surface state recognition results Road surface state recognition was performed using 1000 [sec] travel data and six road surface states were recognized. The road surface condition was acquired from the camera image. Average values and standard deviations in each class were calculated from 1000 training data for each class. Table 4 shows the result of road surface recognition by the naive Bayes estimator using only the multipoint reflection intensity value (171 reflection intensity values) as the road surface feature amount. The calculation formula used for discrimination is as follows.

Table 5 shows the results of road surface recognition by a naive Bayes estimator using a road surface roughness index as a parameter obtained by adding a road surface roughness index to multipoint reflection intensity values (171 reflection intensity values). The calculation formula used for discrimination is as follows.

In the above formula, RI _1-171 : each reflection intensity value, F: road surface roughness index, Class: road surface state, Z: constant.

When only the reflection intensity value was used, wet and aged asphalt had a low recognition rate of 27.26% and submerged and new asphalt were 21.86%, which combined the reflection intensity value and the road surface roughness index. In this case, the recognition rate was greatly improved, with 93.21% wet and aged asphalt and 56.43% flooded and new asphalt, confirming the effect of adding road surface roughness index as road surface feature amount.

As a result, it was confirmed that the discrimination between dry and wet, which was the object of this embodiment, was a recognition rate of 92% or more. Furthermore, when considering the road surface quality, it was confirmed that the recognition rate was over 92% for new asphalt and over 93% for aged asphalt. The recognition rate of submergence and new asphalt is low, but this is because the correct data of the road surface condition is determined from the weather camera data, so there is a deviation from the actual in the determination of the water film thickness when switching from wet to submersion This is thought to have occurred. In addition, since the data processing time at this time is 0.06 [sec / scanned data], it was confirmed that real-time processing can be performed without any problem.

As described above, this embodiment aims to develop road surface state estimation technology using laser radar that is used for white line recognition in automatic driving and platooning, and uses laser radar reflection intensity and road surface roughness index to We developed an algorithm to determine road dry / wet by using a naive Bayes estimator for learning. As a feature of this method, first, the road surface state estimation method using the conventional laser radar uses only the reflection intensity, whereas the multipoint reflection intensity value included in the measurement range is used to determine the positional relationship and the laser. For example, the mechanical properties of light reflection can be considered. Second, in order to improve estimation accuracy, the feature amount is increased using the road surface roughness component as an index, and wavelet decomposition is used to calculate the swell component related to the road surface roughness in order to classify the road surface roughness index by the spatial frequency of the road surface. The road surface state is estimated by using a naive Bayes estimator as a machine learning method for the feature quantity prepared in this way. During traveling, the number of feature quantities of the multipoint reflection intensity value used for road surface state estimation varies depending on the presence or absence of a white line. A naive Bayes estimator was selected as a machine learning method that can also be applied to estimation objects whose number of features can change. The road surface information required for platooning is Dry / Wet information of asphalt, but using the results obtained by actual vehicle measurements, the conditions of the road surface are not only dry / wet but also dry / wet / flooded and asphalt old and new Road surface discrimination was performed on the expanded six road surfaces. As a result, a recognition rate of 92% or higher was obtained for the dry / wet discrimination required in this project. It was also confirmed that the quality of asphalt can be distinguished from old and new.
In the above-described example, an example in which the road surface state discriminator is configured from a naive Bayes estimator has been described. However, the road surface state discriminator may be configured from a neural network. As shown in FIG. 16, the neural network includes 171 input nodes (171 multipoint reflection intensity values), output nodes that label three classes (road surface conditions), and 10 hidden nodes. I have.

It will be understood by those skilled in the art that the above-described embodiment is merely an example illustrating a configuration embodying the present invention, and various changes and improvements can be made without departing from the technical idea of the present invention.

Claims

A light projecting unit that emits laser light so as to scan the road surface;
A light receiving unit for obtaining a reflection intensity value corresponding to each irradiation point on the road surface;
A road surface state discriminator for determining the road surface state using the reflection intensity value,
A set of reflection intensity values acquired corresponding to each irradiation point on the road surface (referred to as a “multi-point reflection intensity value set”) is the irradiation position and reflection in the perspective direction based on the light projecting part depending on the road surface condition. Regulates the relationship with intensity values,
The road surface state discriminator is generated by machine learning using a plurality of multipoint reflection intensity value sets respectively acquired for a plurality of road surfaces having different known road surface states as learning data,
The road surface state discriminator receives a multipoint reflection intensity value set acquired for a road surface whose road surface state is unknown, and determines and outputs the road surface state.
Road surface state estimation device.
The road surface state estimating device is mounted on a vehicle traveling on the road surface,
The road surface state estimation device according to claim 1, wherein the laser light is irradiated so as to scan in a direction across the road surface.
The laser light irradiation range includes a lane marking extending in the traveling direction of the road surface,
The multipoint reflection intensity value set acquired for a road surface whose road surface state is unknown includes a reflection intensity value corresponding to the lane marking,
The road surface state discriminator performs road surface state determination using a reflection intensity value set obtained by removing a reflection intensity value corresponding to the lane marking from the acquired multipoint reflection intensity value set.
The road surface state estimation apparatus according to claim 2.
Analyzing means for obtaining a road surface roughness index depending on the road surface state by analyzing the multipoint reflection intensity value set,
The learning data includes a plurality of road surface roughness indices respectively acquired for a plurality of road surfaces having different known road surface conditions,
The road surface state discriminator receives a multipoint reflection intensity value set and road surface roughness index acquired for a road surface whose road surface state is unknown, and determines and outputs the road surface state.
The road surface state estimating device according to any one of claims 1 to 3.
The road surface state estimating apparatus according to claim 4, wherein the analysis means is discrete wavelet analysis.
The road surface state estimation device according to any one of claims 1 to 5, wherein the road surface state discriminator comprises a naive Bayes estimator.
The road surface state estimation device according to any one of claims 1 to 5, wherein the road surface state discriminator is constituted by a neural network.
The road surface state estimating device according to any one of claims 1 to 7, wherein the road surface state discriminator discriminates at least a dry state and a wet state.