JP7843401B2 - Information processing device, map generation device, storage device, control method and program - Google Patents

Description

This invention relates to map data used for location estimation.

Conventionally, techniques have been known for estimating a vehicle's own position by matching shape data of surrounding objects measured using measuring devices such as laser scanners with map data that pre-stores the shapes of surrounding objects. For example, Patent Document 1 discloses an autonomous mobile system that determines whether detected objects in voxels, which are divided into spaces according to predetermined rules, are stationary or moving, and then matches map data with measurement data for voxels containing stationary objects. Non-Patent Document 1 discloses a technique for recognizing measurement point clouds constituting vegetation such as trees by analyzing measurement point clouds obtained from a lidar or similar device.

International Public Release WO2013/076829

Fabrice Monnier, Bruno Vallet, Bahman Soheilian, TREES DETECTION FROM LASER POINT CLOUDS ACQUIRED IN DENSE URBAN AREAS BY A MOBILE MAPPING SYSTEM, [online], 2012, ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences, [Retrieved July 14, 2018], Internet <URL: https://www.isprs-ann-photogramm-remote-sens-spatial-inf-sci.net/I-3/245/2012/isprsannals-I-3-245-2012.pdf>

In location estimation using map data represented by voxels, if location estimation is performed by referencing voxels that include vegetation that changes over time, there is a possibility that the location estimation results may be inaccurate.

This invention was made to solve the above-mentioned problems, and its main objective is to suitably achieve accurate position estimation.

The invention described in the claim is an information processing device having storage means for storing map data which includes, for each region that divides space, information indicating an object and information regarding a weight representing the reliability of the position estimation of the information indicating the object, wherein the weight for a region including vegetation is set to a lower value than the weight for a region of structures other than vegetation that does not include vegetation; and estimation means for estimating the position of a mobile body by weighting the results of comparing measurement information of an object measured by a measuring device mounted on the mobile body with the information indicating the object for each region.
Furthermore, the invention described in the claim comprises: recognition means for recognizing a region containing measurement points of vegetation from a region containing measurement points of an object measured by a measuring device, which is a region divided in space; and map generation means for generating map data that associates information indicating the object for each region with a weight representing the reliability of the position estimation of the information regarding the position, wherein the map generation means is a map generation device that makes the weight for the region containing vegetation smaller than the weight for the region of structures other than vegetation that does not contain vegetation.

Furthermore, the invention described in the claim includes a storage device for storing map data in which, for each region that divides space, information indicating an object and information regarding a weight representing the reliability of the position estimation of the information indicating the object are set to a lower value than the weight for a region of structures other than vegetation that does not include vegetation.
Furthermore, the invention described in the claims is a control method performed by a computer that refers to map data, wherein for each region that divides space, the control method includes information indicating an object and information regarding a weight representing the reliability of the position estimation of the information indicating the object, wherein the weight for a region including vegetation is set to a lower value than the weight for a region of structures other than vegetation that does not include vegetation, and the method includes an estimation step of estimating the position of the mobile body by weighting the results of comparing measurement information of an object measured by a measuring device mounted on the mobile body with the information indicating the object for each region.
Furthermore, the invention described in the claims is a program to be executed by a computer that references map data, which includes, for each region that divides space, information indicating an object and information regarding a weight representing the reliability of the position estimation of the information indicating the object, wherein the weight for a region including vegetation is set to a lower value than the weight for a region of structures other than vegetation that does not include vegetation, and the computer functions as an estimation means for estimating the position of the mobile body by weighting the results of comparing measurement information of an object measured by a measuring device mounted on the mobile body with the information indicating the object for each region.

This is a schematic configuration of the map update system. This shows the block configuration of the in-vehicle equipment and server device. An example of a general data structure for voxel data is shown. This is a flowchart showing the procedure for generating voxel data. This flowchart shows the procedure for vegetation determination. This is a flowchart showing the procedure for the position estimation process. This figure shows an overview of the position estimation process based on the first embodiment and the position estimation process based on the comparative example. An example of the general data structure of voxel data relating to the second embodiment is shown. This is a flowchart showing the procedure for generating voxel data according to the second embodiment. This flowchart shows the procedure for dynamic object detection processing according to the second embodiment.

According to a preferred embodiment of the present invention, the map data structure includes, for each spatially divided region, information indicating an object and information regarding the weights used when using the information indicating the object for position estimation, wherein the weight for regions including vegetation is set to a lower value than the weight for regions not including vegetation. This map data structure is referenced by an information processing device that estimates the position of a mobile body by weighting the results of matching measurement information of an object measured by a measuring device mounted on the mobile body with the information indicating the object for each region. A "spatially divided region" refers to a region obtained by dividing space according to a predetermined rule, such as a rectangular prism or cube of constant size. By referencing map data having this data structure, the information processing device can relatively lower the weighting of the matching results for regions including vegetation, thereby effectively suppressing a decrease in position estimation accuracy caused by vegetation movement, etc.

In one embodiment of the above data structure, each region further includes information for identifying whether or not it contains vegetation. This allows an information processing device referencing map data to suitably identify regions containing vegetation.

In another embodiment of the above data structure, the information indicating the object is information regarding the mean and variance of the point cloud of the object's surface in each of the regions, and the information processing device calculates an evaluation value for the matching based on the measurement information of the object, the mean, the variance, and the weight. In this embodiment, the information processing device can calculate an evaluation value based on map data and measurement information and suitably estimate the position of the moving object.

In another embodiment of the data structure described above, the region containing vegetation is the region where, based on measurement points of objects within that region, it is determined that the object does not correspond to either a columnar body or a plane. In this embodiment, the region containing vegetation is preferably identified.

According to another embodiment of the present invention, the storage medium stores map data having the data structure described in any of the above.

According to yet another embodiment of the present invention, the storage device has storage means for storing map data in which, for each spatially divided region, information indicating an object and information regarding the weights used when using the information indicating the object for position estimation, the weights for regions including vegetation are set to lower values than the weights for regions not including vegetation.

In one embodiment of the above-described storage device, the storage device further comprises a transmission means for transmitting part or all of the map data to a vehicle or in-vehicle device. In this embodiment, the storage device functions as a distribution device for map data referenced in position estimation.

The following describes preferred embodiments of the present invention with reference to the drawings.

<First Example>
The first embodiment relates to position estimation based on voxel data.

(1) Overview of the map update system
Figure 1 shows a schematic configuration of a map update system according to the first embodiment. The map update system comprises an in-vehicle unit 1 that moves with the vehicle and a server device 2 that distributes map information. In Figure 1, only one set of in-vehicle unit 1 and vehicle communicating with the server device 2 is shown, but in reality, there are multiple sets of in-vehicle unit 1 and vehicles at different locations.

The in-vehicle unit 1 is electrically connected to external sensors such as a LiDAR (Light Detection and Ranging, or Laser Illuminated Detection and Ranging) and internal sensors such as a gyro sensor and a vehicle speed sensor. Based on the outputs of these sensors, it estimates the position of the vehicle on which the in-vehicle unit 1 is mounted (also called "vehicle position"). Based on the estimated vehicle position, the in-vehicle unit 1 performs automatic driving control of the vehicle so that it travels along a set route to a destination. The in-vehicle unit 1 stores a map database (DB: DataBase) 10 that includes voxel data. Voxel data is data that records position information of stationary structures, etc., for each region (also called "voxel") when a three-dimensional space is divided into multiple regions. The voxel data includes point cloud data of measured points of stationary structures within each voxel, represented according to a normal distribution. As described later, this data is used for scan matching using NDT (Normal Distributions Transform).

The in-vehicle unit 1 estimates the vehicle's position by performing scan matching based on NDT, using point cloud data obtained by converting measurement points on the object's surface output by the lidar into an absolute coordinate system, and voxel data corresponding to the voxel to which the point cloud data belongs. The in-vehicle unit 1 also transmits measurement data "D1," including the aforementioned point cloud data, to the server device 2. Furthermore, the in-vehicle unit 1 updates the map database 10 by receiving update data "D2" related to the map database 10 from the server device 2. The in-vehicle unit 1 is an example of an "information processing device" and a "position estimation device."

Server device 2 communicates data with in-vehicle units 1 that support multiple vehicles. Server device 2 stores a distribution map DB 23 for distribution to the in-vehicle units 1, and the distribution map DB 23 contains voxel data corresponding to each voxel. Server device 2 also stores a measurement point cloud DB 24 containing measurement data D1 received from the in-vehicle units 1. Server device 2 generates voxel data based on the measurement data D1 stored in the measurement point cloud DB 24 and updates the distribution map DB 23 based on the generated voxel data. Server device 2 also transmits update data D2, including the generated voxel data, to the in-vehicle units 1. Server device 2 is an example of a "storage device" and a "map generation device."

(2) Configuration of the in-vehicle unit
Figure 2(A) shows a block diagram representing the functional configuration of the in-vehicle unit 1. As shown in Figure 2(A), the in-vehicle unit 1 mainly consists of a communication unit 11, a storage unit 12, a sensor unit 13, an input unit 14, a control unit 15, and an output unit 16. The communication unit 11, storage unit 12, sensor unit 13, input unit 14, control unit 15, and output unit 16 are interconnected via a bus line.

The communication unit 11 transmits measurement data D1 generated by the control unit 15 to the server device 2 and receives update data D2 distributed from the server device 2, based on the control of the control unit 15. The communication unit 11 also transmits signals to the vehicle for controlling the vehicle and receives signals related to the vehicle's status from the vehicle.

The storage unit 12 stores programs executed by the control unit 15 and information necessary for the control unit 15 to perform predetermined processing. In this embodiment, the storage unit 12 stores a map database 10 containing voxel data.

The sensor unit 13 includes a lidar 30, a camera 31, a GPS receiver 32, a gyro sensor 33, and a speed sensor 34. The lidar 30 discretely measures the distance to an object in the external environment by emitting a pulsed laser within a predetermined angular range in the horizontal and vertical directions, and generates three-dimensional point cloud data indicating the position of the object. In this case, the lidar 30 has an irradiation unit that irradiates laser light while changing the irradiation direction, a light receiving unit that receives reflected (scattered) light from the irradiated laser light, and an output unit that outputs scan data based on the received signal output by the light receiving unit. The scan data is generated based on the irradiation direction corresponding to the laser light received by the light receiving unit and the response delay time of the laser light, which is determined based on the received signal described above. The lidar 30 is an example of a "measuring device".

The input unit 14 includes buttons, a touch panel, a remote controller, a voice input device, etc., for user operation. It receives inputs such as specifying a destination for route searching and specifying whether autonomous driving is on or off, and supplies the generated input signals to the control unit 15. The output unit 16 includes, for example, a display or speaker that outputs based on the control of the control unit 15.

The control unit 15 includes a CPU for executing programs and controls the entire in-vehicle unit 1. For example, the control unit 15 estimates the vehicle's position by performing scan matching based on NDT (Non-Definitive Telemetry) using point cloud data obtained by converting measurement points output by the lidar 30 to an absolute coordinate system, and voxel data corresponding to the voxels to which the point cloud data belongs. Furthermore, the control unit 15 updates the map DB 10 based on update data D2 received by the communication unit 11 from the server device 2.

Furthermore, the control unit 15 transmits the measurement data D1, generated based on the point cloud data output from the rider 30, to the server device 2 via the communication unit 11. In this case, for example, the control unit 15 converts the point cloud data of measurement points output from the rider 30 into an absolute coordinate system based on the estimated vehicle position and the position and orientation information of the rider 30 relative to the vehicle. The converted point cloud data is then included in the measurement data D1 and transmitted to the server device 2 via the communication unit 11. In another example, the control unit 15 includes the point cloud data (so-called raw data) output from the rider 30 and the data necessary to convert the point cloud data into an absolute coordinate system (such as the vehicle position mentioned above) in the measurement data D1 and transmits it to the server device 2 via the communication unit 11. In the latter example, the server device 2 generates point cloud data of measurement points represented in an absolute coordinate system based on the measurement data D1.

(3) Server device configuration
Figure 2(B) shows the schematic configuration of the server device 2. As shown in Figure 2(B), the server device 2 has a communication unit 21, a storage unit 22, and a control unit 25. The communication unit 21, the storage unit 22, and the control unit 25 are interconnected via a bus line.

The communication unit 21 communicates various data with the in-vehicle device 1 based on the control of the control unit 25. The storage unit 22 stores programs for controlling the operation of the server device 2 and holds information necessary for the operation of the server device 2. The storage unit 22 also stores the distribution map DB 23 and the measurement point cloud DB 24, which records point cloud data of measurement points of objects based on measurement data D1 transmitted from multiple in-vehicle devices 1.

The control unit 25, equipped with a CPU, ROM, RAM, etc. (not shown), performs various controls on each component within the server device 2. In this embodiment, the control unit 25 stores the measurement data D1 received by the communication unit 21 from the in-vehicle device 1 in the measurement point cloud DB 24, and generates voxel data based on the stored measurement data D1. In this case, the control unit 25 determines whether each voxel contains vegetation, and based on the determination result, generates a weighting value and a vegetation flag (described later) and includes them in the voxel data. Furthermore, the control unit 25 transmits update data D2 based on the generated voxel data to the in-vehicle device 1 via the communication unit 21. The control unit 25 is an example of a "computer" that executes programs.

(4) Scan matching based on NDT
Next, we will explain the scan matching based on NDT in this embodiment.

First, we will explain the voxel data used for scan matching based on NDT. Figure 3 shows an example of a general data structure for voxel data.

The voxel data includes parameter information for representing the point cloud within the voxel using a normal distribution. In this embodiment, as shown in Figure 3, it includes the voxel ID, voxel coordinates, mean vector, covariance matrix, weighting values, and vegetation flag. Here, "voxel coordinates" represent the absolute three-dimensional coordinates of a reference position, such as the center position of each voxel. Each voxel is a cube that divides space into a grid, and its shape and size are predetermined; therefore, the space of each voxel can be identified by its voxel coordinates. The voxel coordinates may also be used as the voxel ID.

The "mean vector" and "covariance matrix" represent the mean vector and covariance matrix that correspond to the parameters when representing the point cloud within the target voxel using a normal distribution. If the coordinates of any point "i" within any voxel "n" are defined as X _n (i) = [x _n (i), y _n (i), z _n (i)] ^T , and the number of points within voxel n is "N _n ", then the mean vector "μ _n " and covariance matrix "V _n " in voxel n are expressed by the following equations (1) and (2), respectively.

The "weighting value" is set to a value corresponding to the confidence level of the target voxel's voxel data (particularly the mean vector and covariance matrix), and represents the weight assigned to the target voxel in scan matching. In this embodiment, the weighting value for voxels determined to contain vegetation (also called "vegetation voxels Bv") is set lower than the weighting value for other voxels. The "vegetation flag" is flag information indicating whether or not the target voxel contains vegetation.

Next, we will explain scan matching using NDT with voxel data. In this embodiment, as will be described later, the in-vehicle device 1 calculates the value of the evaluation function (evaluation value) obtained by NDT scan matching by weighting it using the weighting values included in the voxel data. This allows the in-vehicle device 1 to effectively improve the position estimation accuracy based on NDT scan matching.

Scan matching using NDT for a vehicle estimates the estimation parameter P = [t _x , t _y , t _z , t _ψ ] ^T , which consists of the amount of movement within the road plane (here referred to as xy coordinates) and the orientation of the vehicle. Here, "t _x " represents the amount of movement in the x direction, "t _y " represents the amount of movement in the y direction, "t _z " represents the amount of movement in the z direction, and "t _ψ " represents the yaw angle (the amount of change in angle in the yaw direction). Note that the pitch angle and roll angle are caused by road gradient and vibration, but are negligibly small.

When the point cloud data obtained from measurement points using LIDA2 is converted to an absolute coordinate system and then associated with the voxels to be matched, the coordinates of any point in the corresponding voxel n are given by X _L (i) = [x _n (i), y _n (i), z _n (i)] ^T. Then, using the estimation parameter P described above, when X _L (i) = [x _n (i), y _n (i), z _n (i)] ^T is transformed, the transformed coordinate "X' _n " is expressed by the following equation (3).

In this embodiment, the in-vehicle device 1 uses the coordinate-transformed point cloud and the mean vector _μn and covariance matrix _Vn included in the voxel data to calculate the evaluation function value "E _n " of voxel n, which is shown by the following equation (4).

Here, "w _n " represents the weighting value for voxel n. According to equation (4), the larger the weighting value w _n , the larger the evaluation function E _n will be.

Furthermore, after calculating the evaluation function value for each voxel (also called the "individual evaluation function value") shown by equation (4), the in-vehicle unit 1 calculates an overall evaluation function value "E" (also called the "overall evaluation function value") for all voxels subject to matching, shown by equation (5).

"M" indicates the number of voxels to be matched. The overall evaluation function value E is an example of an "evaluation value".

Subsequently, the in-vehicle unit 1 calculates the estimated parameters P that maximize the overall evaluation function value E using an arbitrary root-finding algorithm such as Newton's method. Then, the in-vehicle unit 1 applies the estimated parameters P to the vehicle's position (also called the "predicted vehicle position") "X ^- " predicted based on the output of the autonomous positioning device, etc., to estimate a vehicle position (also called the "estimated vehicle position") "X ^{^} " which is more accurate than the predicted vehicle position.

In this way, the in-vehicle unit 1 multiplies each voxel's data (mean vector, covariance matrix) by a weighting value that depends on whether or not it is a vegetation voxel Bv. As a result, the evaluation function E _n of vegetation voxels Bv becomes relatively smaller, and the position estimation accuracy by NDT matching is improved accordingly.

Note that the weighting value for vegetation voxels Bv may be 0. In this case, the estimation parameter P is estimated based on the evaluation function E _n of voxels other than vegetation voxels Bv, making it possible to perform position estimation that completely eliminates the influence of vegetation.

(5) Generating voxel data
Next, we will explain the process by which the server device 2 determines the vegetation voxel Bv when generating voxel data.

Figure 4 is a flowchart illustrating the voxel data generation process executed by server device 2. Server device 2 executes the processes shown in the flowchart in Figure 4 at predetermined timings.

First, the server device 2 recognizes the voxel to which each measurement point recorded in the measurement point cloud DB 24 belongs, and calculates the mean vector and covariance matrix of the measurement points for each voxel based on equations (1) and (2) described above (step S101).

Next, the server device 2 performs a vegetation determination process for each voxel containing the measurement point, determining whether or not it is a vegetated voxel Bv (step S102). A specific example of this vegetation determination process will be described later with reference to the flowchart in Figure 5.

Then, based on the determination result in step S102, the server device 2 sets a weighting value and a vegetation flag for each voxel (step S103). For example, the server device 2 pre-stores setting information indicating the weighting value and vegetation flag values to be assigned to vegetation voxels Bv and other voxels, respectively. Then, referring to the above setting information, the server device 2 assigns different weighting values and vegetation flags to the voxel determined to be vegetation voxel Bv and to the voxels other than vegetation voxel Bv in step S103. In this case, the weighting value for vegetation voxel Bv is set lower than the weighting values for the other voxels.

Furthermore, as a preprocessing step before the flowchart in Figure 4, the server device 2 may perform a process to exclude point clouds representing dynamic objects from the point clouds recorded in the measurement point cloud DB 24. For example, the server device 2 may identify point clouds that form specific dynamic objects such as pedestrians or vehicles based on processing such as pattern matching based on shape and size, and then delete the identified point clouds from the measurement point cloud DB 24.

Figure 5 is a flowchart showing an example of the vegetation determination process performed in step S102 of Figure 4. In the example shown in Figure 5, the server device 2 performs principal component analysis on each voxel of the measurement point cloud recorded in the measurement point cloud DB24 to determine voxels representing planar or columnar structures, and determines voxels that do not correspond to either planar or columnar structures as vegetation voxels Bv.

First, the server device 2 performs principal component analysis on each voxel of the measurement point cloud D24 (step S201). Specifically, the server device 2 calculates three sets of eigenvalues and eigenvectors corresponding to the first to third principal components from the voxel-specific covariance matrix calculated in step S101 of Figure 4. Here, the eigenvalues represent the variance (i.e., magnitude) of each principal component, and the eigenvectors represent the direction of each principal component. Therefore, the server device 2 considers that the largest eigenvalue and its corresponding eigenvector correspond to the first principal component, the second largest eigenvalue and its corresponding eigenvector correspond to the second principal component, and the remaining eigenvalues and eigenvectors correspond to the third principal component. Note that the first to third principal components are orthogonal to each other.

Then, based on the results of the principal component analysis in step S201, the server device 2 extracts voxels that form a plane (also called "planar voxels Bf") (step S202). For example, the server device 2 considers voxels whose magnitude of the third principal component (i.e., the variance shown by the eigenvalue of the third principal component) is less than or equal to a predetermined value to be planar voxels Bf. In addition, the server device 2 may also consider voxels whose eigenvectors corresponding to the third principal component have a direction that approximately coincides with the vertical or horizontal direction to be planar voxels Bf.

Furthermore, based on the results of the principal component analysis in step S201, the server device 2 extracts voxels that form columnar structures (also called "columnar voxels Bp") (step S203). For example, the server device 2 extracts voxels as columnar voxels Bp if the contribution rate of the first principal component is greater than or equal to a predetermined value, and the variance of the first principal component is close to a uniform distribution. In addition, the server device 2 may also consider voxels as columnar voxels Bp if the direction indicated by the eigenvector corresponding to the first principal component approximately coincides with the vertical or horizontal direction.

The server device 2 then determines that the remaining voxels containing measurement points that do not fall under either planar voxels Bf or columnar voxels Bp are vegetation voxels Bv (step S204). Here, since most static structures that are not vegetation consist of planar or columnar shapes, the applicant has obtained good recognition results for vegetation voxels Bv in experiments, etc., by considering voxels excluding those forming planar or columnar shapes as vegetation voxels Bv.

(6) Position estimation using voxel data
Next, we will explain the position estimation process using voxel data. Figure 6 is an example of a flowchart showing the procedure for position estimation using voxel data.

First, the in-vehicle unit 1 sets the initial value of the vehicle's position based on the output of the GPS receiver 32, etc. (step S301). Next, the in-vehicle unit 1 obtains the vehicle speed from the speed sensor 34 and the yaw angular velocity from the gyro sensor 33 (step S302). Then, based on the results obtained in step S302, the in-vehicle unit 1 calculates the vehicle's travel distance and the change in the vehicle's orientation (step S303).

Subsequently, the on-board unit 1 adds the travel distance and change in direction calculated in step S303 to the estimated vehicle position from one time point prior to calculate the predicted vehicle position (step S304). Then, based on the predicted vehicle position calculated in step S304, the on-board unit 1 refers to the map DB 10 and acquires voxel data for voxels surrounding the vehicle position (step S305). Furthermore, based on the predicted vehicle position calculated in step S304, the on-board unit 1 divides the point cloud data obtained from the lidar 30, converted to an absolute coordinate system, into individual voxels (step S306).

Then, the in-vehicle device 1 calculates NDT scan matching using the mean, covariance matrix, and weighting values included in the voxel data acquired in step S305 (step S307). Specifically, for each voxel to which a measurement point was assigned in step S306, the in-vehicle device 1 calculates the individual evaluation function value _E shown in equation (4) using the mean, covariance matrix, and weighting values included in the corresponding voxel data. Then, based on the individual evaluation function value _E for each voxel, the in-vehicle device 1 calculates the overall evaluation function value E shown in equation (5). In this case, the overall evaluation function value E becomes a nonlinear equation that includes the variables of each element of the estimated parameter P.

Then, the in-vehicle device 1 determines the estimated parameter P that maximizes the overall evaluation function value using numerical methods for solving nonlinear equations, such as Newton's method (step S308). Subsequently, the in-vehicle device 1 calculates the estimated vehicle position by applying the estimated parameter P calculated in step S308 to the predicted vehicle position calculated in step S304 (step S309).

Here, the weighting value used to calculate the individual evaluation function value E _n in step S307 is set lower for vegetation voxels Bv than for other voxels. As a result, the onboard unit 1 relatively reduces the degree of influence on position estimation for vegetation voxels Bv that represent vegetation, thereby suitably improving the position estimation accuracy.

Figure 7 shows an overview of the processing of position estimation based on the flowchart in Figure 6 using weighted values (also called "main position estimation") and position estimation that refers to voxel data generated based on a measurement point cloud with vegetation measurement points removed instead of using weighted values (also called "position estimation based on comparative example"). Here, we consider NDT scan matching targeting voxel B1, which contains parts of vegetation and structures respectively, and voxel B2, which is adjacent to voxel B1, contains part of a structure, and does not contain vegetation.

Voxels B1 and B2, indicated by arrow A1 labeled "Weighting," show a schematic representation of the voxel data used in this position estimation. Here, hatched area 50 shows the distribution of the vegetation measurement point cloud used to generate the voxel data, and hatched areas 51 and 52 show the distribution of the structure measurement point cloud used to generate the voxel data. Furthermore, dashed frame 60 shows the variance of the point cloud within voxel B1 as indicated by the covariance matrix recorded as voxel data, and dashed frame 61 shows the variance of the point cloud within voxel B2 as indicated by the covariance matrix recorded as voxel data. Additionally, voxels B1 and B2, indicated by arrow A2 labeled "Point Cloud Removal," show a schematic representation of the voxel data used in position estimation based on the comparative example. Here, hatched areas 54 and 55 show the distribution of the structure measurement point cloud used to generate the voxel data. Furthermore, the dashed box 63 shows the variance of the point cloud within voxel B1 as indicated by the covariance matrix recorded as voxel data, and the dashed box 64 shows the variance of the point cloud within voxel B2 as indicated by the covariance matrix recorded as voxel data.

In the comparative example, the position estimation method generates voxel data after removing the measurement point cloud representing vegetation. Therefore, the dashed frame 63 representing the dispersion of the point cloud within voxel B1 is located at a position centered on the structure indicated by the hatched region 54. In contrast, in this position estimation method, the dashed frame 60 representing the dispersion of the point cloud within voxel B1 spans both the vegetation indicated by the hatched region 50 and the structure indicated by the hatched region 51. Furthermore, in the comparative example, voxels B1 and B2 are treated with the same weight. In contrast, in this position estimation method, voxel B1 is considered a vegetation voxel Bv, and the weighting value of voxel B1 is set to a smaller value than the weighting value of voxel B2, which is not a vegetation voxel Bv.

Here, when the in-vehicle device 1 acquires the measurement point cloud 40, represented by the "x" marks within voxels B1 and B2, using the lidar 30, in this position estimation method, since the weighting value of voxel B1 is lower than that of voxel B2, the voxel data in voxel B2 is preferentially matched with the measurement point cloud 40. On the other hand, in the position estimation method based on the comparative example, since no weighting value is set for each voxel, the matching of voxel data and the measurement point cloud 40 is performed with equal weighting across the entire system (in this case, voxels B1 and B2).

The matching result when performing this position estimation is indicated by arrow A3, and the matching result when performing position estimation based on the comparative example is indicated by arrow A4. Here, in the position estimation based on the comparative example, the measurement point cloud of vegetation within voxel B1 is excluded during voxel data generation. Therefore, the matching between the measurement point cloud 40 within voxel B1, which includes vegetation as a measurement target, and the voxel data becomes inaccurate. Furthermore, because no weighting is applied in the position estimation based on the comparative example, the overall matching accuracy is reduced due to the inconsistency between the voxel data in voxel B1 and the measurement point cloud 40.

On the other hand, in this position estimation method, since the measurement point cloud of vegetation within voxel B1 is not excluded during the generation of voxel data, matching between the measurement point cloud 40 within voxel B1, which includes vegetation as the measurement target, and the voxel data is suitably performed. However, because the vegetation within voxel B1 fluctuates due to wind and other factors, the matching accuracy in voxel B1 is lower than that in voxel B2. Therefore, if NDT matching is performed without weighting each voxel, the overall matching accuracy will decrease due to the influence of the matching accuracy of voxel B1. In contrast, in this position estimation method, the weighting value of voxel B1 is set relatively low, allowing for matching while suitably reducing the influence of vegetation fluctuations.

As described above, the data structure of the map data in this embodiment includes, for each voxel, a mean vector and covariance matrix relating to the position of the object surface, and weighting values for using these in position estimation. The weighting value for vegetation voxels Bv is set lower than the weighting value for voxels without vegetation. This map data with this data structure is suitably referenced by the on-board unit 1, which estimates the vehicle's position by weighting the results of comparing the measurement information of objects measured by a measuring device such as a lidar 30 mounted on the vehicle with the voxel data, for each voxel.

Furthermore, the server device 2 in this embodiment performs the following processes: recognizing vegetation voxels Bv containing vegetation measurement points from voxels containing measurement points of objects measured by a measuring device such as a lidar 30; and generating map data that associates the mean vector and covariance matrix for each voxel with weighting values used for position estimation. Here, the server device 2 sets the weighting value for vegetation voxels Bv to be smaller than the weighting value for voxels without vegetation.

Furthermore, the in-vehicle device 1 in this embodiment stores voxel data including the mean vector and covariance matrix for each voxel that divides space, and weighting values for determining the weights used when using these for position estimation. It performs the following processes: acquiring measurement information about objects measured by a measuring device such as a lidar 30; comparing the measurement information with the voxel data for each voxel; and estimating the vehicle's position by weighting the results of this comparison based on the weighting values. In this case, the weighting value for vegetation voxels Bv is smaller than the weighting value for voxels without vegetation.

<Second Example>
In the second embodiment, the in-vehicle device 1 differs from the first embodiment in that it performs dynamic object detection based on voxel data instead of, or in addition to, position estimation based on voxel data. The configurations of the map update system, in-vehicle device 1, and server device 2 in the second embodiment are the same as those shown in Figures 1 and 2, so their explanation is omitted.

Figure 8 shows an example of a schematic data structure for voxel data in the second embodiment.

The voxel data shown in Figure 8 includes a voxel ID, voxel coordinates, a stationary obstacle flag, a vegetation flag, a mean vector, a covariance matrix, and weighting values. Here, the stationary obstacle flag indicates whether the target voxel is a voxel where stationary obstacles mainly exist and where dynamic object detection is not required (also called a "stationary obstacle voxel"). A stationary obstacle voxel indicates a region where the entirety or majority of the target voxel is composed of stationary obstacles, and therefore, detection of dynamic objects that should be detected for driving purposes is not required (there is no room for dynamic objects). Accordingly, when the in-vehicle unit 1 detects dynamic objects based on the output of the lidar 30, etc., it identifies stationary obstacle voxels by referring to the stationary obstacle flag and performs dynamic object detection on voxels other than stationary obstacle voxels. By limiting the detection area for dynamic objects in this way, the load on the dynamic object detection process is suitably reduced while improving the accuracy of dynamic object detection. The stationary obstacle flag is an example of "stationary object data".

Furthermore, in this embodiment, when generating a stationary obstacle flag, the server device 2 does not consider voxels determined to be vegetation voxels Bv as stationary obstacle voxels. This ensures that dynamic objects appearing near vegetation are reliably detected. The server device 2 is an example of a "stationary object data generation device."

Figure 9 is a flowchart showing the procedure for generating voxel data according to the second embodiment. The server device 2 executes the process shown in the flowchart in Figure 9 at predetermined timings.

First, the server device 2 refers to the measurement point cloud DB 24, which records the measurement data D1 transmitted from each in-vehicle unit 1, and calculates the mean vector and covariance matrix of the measurement points for each voxel (step S401). Next, the server device 2 performs a vegetation determination process for each voxel containing a measurement point, determining whether or not it is a vegetation voxel Bv (step S402). This vegetation determination process is performed, for example, based on the flowchart in Figure 5 described above.

Next, the server device 2 generates a stationary obstacle flag based on the vegetation determination result in step S402 (step S403). At this time, the server device 2 generates a stationary obstacle flag indicating that voxels determined to be vegetation voxels Bv are not stationary obstacle voxels. This effectively suppresses the oversight of dynamic objects caused by misidentifying areas where dynamic objects could exist as stationary obstacle voxels.

Furthermore, as a preprocessing step before the flowchart in Figure 9, the server device 2 may perform a process to exclude measurement point clouds representing dynamic objects from the measurement point cloud DB 24. For example, the server device 2 identifies measurement point clouds that form specific dynamic objects such as pedestrians or vehicles based on processing such as pattern matching based on shape and size, and deletes the identified measurement point clouds from the measurement point cloud DB 24. The server device 2 may also set a stationary obstacle flag to indicate that voxels containing measurement points representing dynamic objects to be excluded are not stationary obstacle voxels, so that they are included in the detection area for dynamic objects.

Figure 10 is a flowchart showing the procedure for detecting a dynamic object according to the second embodiment. The server device 2 repeatedly executes the process shown in the flowchart in Figure 10, for example, while the vehicle is in motion.

First, the in-vehicle unit 1 acquires a measurement point cloud based on the output of the lidar 30, measuring the surface of objects around the vehicle (step S501). Next, the in-vehicle unit 1 acquires voxel data of voxels present around the vehicle's position by referring to the map DB 10 based on the vehicle's position (step S502). The aforementioned vehicle position is, for example, a predicted vehicle position obtained by adding the travel distance and change in direction calculated based on the output of the sensor unit 13 to the estimated vehicle position one time step prior, similar to the position estimation process of the first embodiment shown in Figure 6. Then, the in-vehicle unit 1 identifies stationary obstacle voxels from the voxels containing the measurement points acquired in step S501 by referring to the stationary obstacle flag included in the voxel data of each voxel (step S503). Then, the in-vehicle unit 1 performs dynamic object detection processing, excluding stationary obstacle voxels from the detection target area (step S504). In other words, the in-vehicle device 1 considers the stationary obstacle voxels as areas where no dynamic objects appear, skipping the dynamic object detection process, and performs the dynamic object detection process for the remaining voxels.

As described above, the server device 2 according to the second embodiment performs the following processes: detecting vegetation voxels Bv containing vegetation measurement points from voxels containing measurement points related to objects measured by a measuring device such as the lidar 30; and generating a stationary obstacle flag indicating that a stationary obstacle exists in a region of voxels from which at least vegetation voxels Bv have been excluded from the voxels containing measurement points.

<Different example>
The following describes preferred modifications of the first and second embodiments. These modifications may be applied in combination to each of the embodiments described above.

(Variation 1)
Voxel data is not limited to a data structure that includes a mean vector and a covariance matrix, as shown in Figures 3 and 8. For example, voxel data may include point cloud data obtained by converting measurement data of the measurement and maintenance vehicle used to calculate the mean vector and covariance matrix into an absolute coordinate system. In this case, the in-vehicle device 1 is not limited to scan matching by NDT, but may also perform self-position estimation by applying other scan matching methods such as ICP (Iterative Closest Point).

(Variation 2)
In the first embodiment, the voxel data shown in Figure 3 had both weighting values and vegetation flags, but it may also have only one of them.

For example, if the voxel data has a data structure that does not include weighting values, the in-vehicle unit 1 sets the weighting values used in step S307 of the position estimation process shown in Figure 6 to predetermined values corresponding to the vegetation flag values. In this case, the in-vehicle unit 1 sets the weighting value set when the vegetation flag indicates that it is a vegetation voxel Bv to a smaller value than the weighting value set when the vegetation flag indicates that it is not a vegetation voxel Bv. This allows the in-vehicle unit 1 to relatively reduce the degree of influence on position estimation for vegetation voxels Bv, similar to the case where the voxel data has a data structure with weighting values, thereby suitably improving position estimation accuracy. In this case, the vegetation flag is an example of "information for determining weights."

(Variation 3)
The vehicle may have built-in functionality equivalent to that of the in-vehicle unit 1. In this case, the vehicle's electronic control unit (ECU) executes a program stored in the vehicle's memory, thereby performing processing equivalent to that of the control unit 15 of the in-vehicle unit 1.

(Variation 4)
In the second embodiment, instead of excluding vegetation voxels Bv (i.e., voxels containing vegetation) from stationary obstacle voxels, the in-vehicle unit 1 may exclude voxels containing only vegetation (i.e., voxels containing vegetation but not other stationary structures) from stationary obstacle voxels. In this case, the in-vehicle unit 1 considers voxels containing both vegetation and stationary structures as areas where no dynamic objects appear and skips the dynamic object detection process.

(Variation 5)
The in-vehicle unit 1 may perform vegetation determination processing instead of the server device 2. In this case, for example, in the first embodiment, the in-vehicle unit 1 determines vegetation voxels Bv by performing the processing in step S101 and the vegetation determination processing in step S102 of Figure 4 on the point cloud data obtained by converting the measurement points output by the lidar 30 into an absolute coordinate system. The in-vehicle unit 1 then includes the information of the determined vegetation voxels Bv in the measurement data D1 and transmits it to the server device 2. In this case, the server device 2 generates and updates weighted values for voxel data and vegetation flags based on the measurement data D1 received from the in-vehicle unit 1. Similarly in the second embodiment, the in-vehicle unit 1 determines vegetation voxels Bv by performing the processing in step S401 and the vegetation determination processing in step S402 of Figure 9 on the point cloud data obtained by converting the measurement points output by the lidar 30 into an absolute coordinate system. The vehicle-mounted unit 1 then includes the information of the determined vegetation voxel Bv in the measurement data D1 and transmits it to the server device 2. In this case, the server device 2 generates a stationary obstacle flag based on the measurement data D1 received from the vehicle-mounted unit 1.

1. In-vehicle unit 2. Server unit 10. Map database
23. Distribution Map Database
24. Measurement Point Cloud Database
11, 21 Communication Unit 12, 22 Storage Unit 15, 25 Control Unit 13 Sensor Unit 14 Input Unit 16 Output Unit

Claims (7)

In each of the regions that divide the space,
Information that describes an object,
This includes information about weights representing the confidence level in the position estimation of information indicating the object,
A storage means for storing map data in which the weight for areas including vegetation is set to a lower value than the weight for areas of structures other than vegetation that do not include vegetation,
An estimation means for estimating the position of the moving body by comparing measurement information of an object measured by a measuring device mounted on the moving body with information indicating the object in each region, and weighting the results for each region.
An information processing device having The information relating to the object is information relating to the mean and variance of the point cloud on the surface of the object in each of the regions.
The information processing device according to claim 1, wherein the information processing device calculates an evaluation value for the matching based on the measurement information of the object, the mean, the variance, and the weight. The information processing device according to claim 1 or 2, wherein the region including the vegetation is a region where, based on measurement points of an object within that region, the object is determined not to be either a columnar body or a plane. A recognition means that recognizes a region containing measurement points of vegetation from a region containing measurement points of an object measured by a measuring device, which is a region that has been demarcated in space.
The system includes a map generation means that generates map data relating information indicating the object for each region with a weight representing the confidence level in position estimation of the information regarding that location,
The map generation means is a map generation device that makes the weight for the area including vegetation smaller than the weight for the area of structures other than vegetation that does not include vegetation. In each of the regions that divide the space,
Information that describes an object,
This includes information about weights representing the confidence level in the position estimation of information indicating the object,
A storage device having storage means for storing map data in which the weight for areas including vegetation is set to a lower value than the weight for areas of structures other than vegetation that do not include vegetation. In each of the regions that divide the space,
Information that describes an object,
This includes information about weights representing the confidence level in the position estimation of information indicating the object,
A control method performed by a computer referencing map data, wherein the weight for an area including vegetation is set to a lower value than the weight for an area of non-vegetation structures that does not include vegetation,
A control method comprising an estimation step of estimating the position of a moving body by comparing measurement information of an object measured by a measuring device mounted on the moving body with information indicating the object in each region, and then weighting the results for each region. In each of the regions that divide the space,
Information that describes an object,
This includes information about weights representing the confidence level in the position estimation of information indicating the object,
A program to be executed by a computer that references map data, wherein the weight for areas including vegetation is set to a lower value than the weight for areas of structures other than vegetation that do not include vegetation.
A program that causes the computer to function as an estimation means for estimating the position of the moving body by comparing measurement information of an object measured by a measuring device mounted on the moving body with information indicating the object in each of the regions, and then weighting the results for each region.