Sensor installation position estimation device and method

Abstract

To provide a device and method for estimating a sensor arrangement position capable of automatically calculating positions and kinds of infrastructure sensors required in an operation design domain using the infrastructure sensors.SOLUTION: A sensor arrangement position estimation device includes: a map information conversion section for grid-dividing a three-dimensional map of an object area where a mobile body performs a limited area automatic operation; a blind angle estimation section for estimating blind angles between a first grid and a second grid on a movement route of the mobile body in the object area; a risk map generation section for generating a blind angle map on the object area from a plurality of blind angles estimated by the blind angle estimation section; and a sensor arrangement estimation section for estimating sensor arrangement positions from candidates of kinds and arrangement positions of the sensors scheduled to be arranged in the object area.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to a sensor installation position estimation device and method for estimating sensor installation positions when designing an operation design area for limited area automated driving. [Background technology]

[0002] Designing an operational design domain (ODD) is essential for limited-area autonomous driving. In particular, when designing an ODD that works in conjunction with infrastructure sensors, sensor placement that minimizes blind spots within the area is considered manually, so design costs increase as the area becomes more complex and expansive. Therefore, a method is needed to automatically calculate blind spots and sensor placement based on point cloud maps and moving objects.

[0003] The general method for setting the position and orientation of infrastructure sensors to be installed in ODDs is to manually identify dangerous areas within the target area where accidents are likely to occur and areas where moving objects cannot be recognized, and then determine the installation location. However, this method cannot accurately recognize blind spots of moving objects in wide and complex environments such as outdoors, or in cases where the ODD conditions are complex, and there is a possibility that areas within the ODD will not meet the conditions for autonomous driving.

[0004] Patent Document 1 describes a method for detecting blind spots at the design stage by inputting wall surface information and sensor pole position information and having the system automatically calculate the blind spots. Patent Document 1 is realized by a system comprising: a position information input unit that acquires position information of the wall edge and the pole on which the sensor is attached from an input device; a sensor blind spot detection unit that detects the presence or absence of blind spots between the wall surface and the sensor by having a central processing unit calculate the positional relationship between each wall surface and the invisible light beam based on the position information of the wall edge and the pole on which the sensor is attached acquired from the position information input unit; and a notification unit that notifies an external output device of the presence or absence of blind spots based on the results detected by the sensor blind spot detection unit. [Prior art documents] [Patent documents]

[0005] [Patent Document 1] Japanese Patent Application Laid-Open No. 2007-102258 Summary of the Invention [Problem to be solved by the invention]

[0006] Designing an operational design domain (ODD) is essential for limited-area autonomous driving. In particular, when designing an ODD that cooperates with infrastructure sensors, the sensor placement that minimizes blind spots within the area is manually considered, so design costs increase as the area becomes more complex and wide. In addition, since the size of the moving objects moving within the ODD and the sensors installed on them vary from one moving object to another, it is necessary to derive a sensor placement that can eliminate blind spots for all moving objects.

[0007] In this regard, the technology described in Patent Document 1 describes a technique for calculating the blind spots of the system based on wall surface information and the pole positions on which sensors are attached, but does not describe the optimal pole positions or sensor placement that minimizes blind spots.

[0008] In view of the above, the present invention aims to provide a sensor installation location estimation device and method that can automatically calculate the location and type of infrastructure sensors required in an operation design area where infrastructure sensors are used. [Means for solving the problem]

[0009] In view of the above, the present invention provides a "sensor installation position estimation device comprising: a map information conversion unit that divides into grids a three-dimensional map of a target area in which a mobile body performs limited-area autonomous driving; a blind spot estimation unit that estimates blind spots between a first grid and a second grid in the movement path of the mobile body in the target area; a risk map generation unit that generates a blind spot map for the target area from the multiple blind spots estimated by the blind spot estimation unit; and a sensor placement estimation unit that estimates the placement position of a sensor from the type of sensor to be installed within the target area and candidate installation positions."

[0010] The present invention also provides a sensor installation position estimation method, characterized in that a computer device divides a three-dimensional map of a target area in which a mobile body will perform limited-area autonomous driving into grids, estimates blind spots between a first grid and a second grid in the movement path of the mobile body in the target area, generates a blind spot map for the target area from the estimated blind spots, and estimates the placement position of a sensor from candidate types and installation positions of sensors to be installed within the target area. [Effects of the Invention]

[0011] According to the present invention, it is possible to automatically calculate the positions and types of infrastructure sensors required in an operation design area in which infrastructure sensors are used. [Brief explanation of the drawings]

[0012] [Figure 1] 1 is a diagram showing an example of the configuration of a sensor installation position estimation device according to an embodiment of the present invention; [Figure 2A] FIG. 10 is a diagram showing an example of an operation design area. [Figure 2B] FIG. 2B is a diagram showing an example of the operation design area of ​​FIG. 2A represented as a three-dimensional map. [Figure 3] 4 is a flowchart showing an example of processing content of a sensor installation position estimation device according to an embodiment of the present invention. [Figure 4] FIG. 10 is a diagram showing an example of a 3D point cloud map divided into grid sizes. [Figure 5] 4 is a flowchart showing detailed processing of processing step S102 in FIG. 3. [Figure 6] FIG. 10 is a diagram showing an example of a three-dimensional grid displayed as a two-dimensional grid. [Figure 7A] FIG. 10 is a diagram showing an example of loaded route information. [Figure 7B] FIG. 10 is a diagram showing an example of loaded route information. [Figure 8] FIG. 2 is a diagram showing an example in which the routes of moving bodies assumed in the first embodiment are superimposed. [Figure 9] FIG. 10 is a diagram showing an example of an extracted common path. [Figure 10] FIG. 1 is a diagram showing the concept of extracting an area that can be recognized by a sensor mounted on a moving object. [Figure 11] A diagram showing the observation range. [Figure 12A] FIG. 10 is a diagram showing a case where a cylindrical object is placed at an observation point as point cloud information. [Figure 12B] FIG. 10 is a diagram showing an example of an image calculated based on coordinates. [Figure 13A] FIG. 10 is a diagram showing a case where a cylindrical object is placed at an observation point as point cloud information. [Figure 13B] FIG. 10 is a diagram showing an example of an image calculated based on coordinates. [Figure 14] FIG. 1 is a diagram showing observable and unobservable areas as seen from a moving object. [Figure 15] 4 is a flowchart showing detailed processing of processing step S103 in FIG. 3. [Figure 16] FIG. 16 is a diagram showing an example of a risk map created by the processing of FIG. 15. [Figure 17] 4 is a flowchart showing detailed processing of processing step S104 in FIG. 3. [Figure 18A] FIG. 10 is a diagram showing a case where a cylindrical object is placed at an observation point as point cloud information. [Figure 18B] FIG. 10 is a diagram showing an example of an image calculated based on coordinates. [Figure 18C] FIG. 10 is a diagram showing an example of an image calculated based on coordinates. [Figure 19] FIG. 1 is a diagram illustrating the concept of sensing multiplicity. [Figure 20] FIG. 10 is a diagram showing an example of a presentation of a sensor placement result. [Figure 21] FIG. 10 is a diagram showing an example of the configuration of a sensor installation position estimation device according to a second embodiment of the present invention. [Figure 22A] FIG. 7B is a diagram in which the routes of moving objects in a predetermined time period are added to the routes of FIGS. 7A and 7B. [Figure 22B] FIG. 7B is a diagram in which the routes of moving objects in a predetermined time period are added to the routes of FIGS. 7A and 7B. [Figure 23A] FIG. 22B is a diagram showing the risk map at the time of FIG. 22A. [Figure 23B] A diagram showing the risk map at the time of Figure 22B. [Figure 24A]FIG. 23B is a diagram showing an example of installing a mobile sensor in the case of FIG. 23A. [Figure 24B] FIG. 23C is a diagram showing an example of installing a mobile sensor in the case of FIG. 23B. [Figure 25] FIG. 10 is a diagram showing an example of the configuration of a sensor installation position estimation device according to a third embodiment of the present invention. [Figure 26] FIG. 10 is a diagram showing an example in which a sensor installation position 60 is added. [Figure 27A] FIG. 23B is a diagram showing an example of installing a mobile sensor in the case of FIG. 23A. [Figure 27B] FIG. 23C is a diagram showing an example of installing a mobile sensor in the case of FIG. 23B. [Figure 28] FIG. 10 is a diagram showing a case where candidates for sensor installation positions are input. [Figure 29] FIG. 10 is a diagram showing the results of the sensor installation positions. DETAILED DESCRIPTION OF THE INVENTION

[0013] Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings. [Example]

[0014] Fig. 1 is a diagram showing an example of the configuration of a sensor installation position estimation device 10 according to an embodiment of the present invention. The sensor installation position estimation device shown in Fig. 1 is realized using a computer device, and the processing results are displayed on an external terminal (not shown) equipped with a display function. In the following, a device equipped with a sensor installation position estimation function will be described as a "sensor installation position estimation device."

[0015] The sensor installation position estimation device 10 reads a three-dimensional map 20, moving body information 30, and sensor information 40 stored in a storage device connected to a computer device, and outputs a sensor arrangement 50.

[0016] As shown in Figure 1, the sensor installation position estimation device 10 has as its functions a map information conversion unit 11, a blind spot estimation unit 12, a risk map generation unit 13, and a sensor placement estimation unit 14. It inputs as input the information it uses a 3D map 20 of the relevant area, mobile object information 30 including the size, mounted sensors, and movement route of mobile objects traveling within the relevant area, and sensor information 40 to be installed within the relevant area, and outputs sensor placement 50, which are candidates for the type and position of sensors to be installed within the relevant area.

[0017] The sensor installation position estimation device 10 has, for example, a CPU, a GPU, a RAM, a ROM, etc., and can realize these functions by loading a predetermined program stored in the ROM into the RAM and executing it on the CPU. Note that some or all of the functions of the sensor installation position estimation device 10 may be realized using hardware such as an FPGA or an ASIC.

[0018] The 3D map 20 is, for example, a 3D point cloud map generated by measuring an area to be a traffic design area using LiDAR or the like. FIG. 2A shows an example of a traffic design area, and FIG. 2B shows an example of the traffic design area of ​​FIG. 2A represented as a 3D map. Multiple structures 100 exist within the traffic design area R of FIG. 2A. The 3D map of FIG. 2B, which includes the traffic design area R, is expressed in a coordinate system 103 with an arbitrary location as the origin. As a result of observing the traffic design area R using LiDAR or the like based on this coordinate system, information on the structures 100 is grasped as 101. The grasped information 101 on the structures 100 is expressed as a collection of points (point cloud) as shown, for example, in 102. Note that this map may also be generated using a 3D CAD drawing of the relevant area.

[0019] The mobile object information 30 is a group of information indicating the size of the mobile object traveling within the operation design area R, the route it will travel, and the type of sensors such as cameras and LiDAR that are installed, as well as their performance such as resolution and sensing range.

[0020] The sensor information 40 is a group of information indicating the type of sensor (e.g., camera or LiDAR), its resolution, sensing range, whether it can be used in dark areas, and price for candidate sensors to be installed within the operation design area R. Of this information, whether it can be used in dark areas is used to determine whether it can be used in dark places, and price is used to minimize the cost of installing the sensor.

[0021] The map information conversion unit 11, which is a processing function in the sensor installation position estimation device 10 in Fig. 1, divides the input 3D map into a predetermined grid. The divided grid information is input to the blind spot estimation unit 12. Furthermore, when a 3D CAD is input as a 3D map, the map information conversion unit 11 converts it into a 3D point cloud map. Conversion into a 3D map can be achieved by, for example, sampling the surface of each object recorded in the 3D CAD at equal intervals.

[0022] The blind spot estimation unit 12 first extracts locations where multiple moving objects pass by using the moving object information 30. Next, it estimates blind spot areas based on whether or not objects at a certain distance ahead can be recognized at each location, based on the 3D point cloud map output by the map information conversion unit 11 and the position information of the multiple moving objects passing by. The information on the estimated blind spot areas at each location and for each moving object is output to the risk map generation unit 13.

[0023] The risk map generation unit 13 extracts areas where blind spots may occur in the entire operation design area by superimposing the blind spot area estimation results at each position where multiple moving bodies pass, estimated by the blind spot estimation unit 12. The result of the risk map generation unit 13 is output to the sensor placement estimation unit 14.

[0024] The sensor placement estimation unit 14 calculates the placement of sensors that can minimize blind spots within the operation design area R generated by the risk map generation unit 13. At this time, candidate sensor installation locations are output, taking into consideration the upper cost limit of sensors that can be installed and the lighting conditions within the area. At this time, multiple candidate sensor installation locations are output.

[0025] The sensor placement 50 is a candidate for the sensor installation position estimated by the sensor placement estimation unit 14, and includes information on the sensor installation position within the operation design area as well as blind spots after the sensor is installed.

[0026] Next, the processing performed by the sensor installation position estimation device 10 will be described with reference to the flowchart in Fig. 3, using the configuration in Fig. 1 as Example 1. Here, it is assumed that a 3D point cloud map is input as the 3D map 20, two types of moving objects are input as moving object information 30, and one type each of a camera and LiDAR is input as sensor information. In addition, the 3D map has information on the brightness of lighting for each region, and this information is assumed to be carried over even when the map is divided into grids.

[0027] In the sensor installation position estimation device 10, first, in processing step S101, the 3D point cloud map is divided into a predetermined grid size. As shown in FIG. 4, an example of division into grids is performed based on predetermined W (width), D (depth), and H (height), and the division into grids is performed as shown in 105, with each grid having a unit area. Note that although the grid 105 is displayed in three dimensions, for ease of explanation, the following description will be given using a 2D grid 106, as shown in FIG. 6, in which the 3D point cloud map divided into grids is viewed from directly above. Note that in the 2D grid 106 in FIG. 6, the portion with a background color indicates the area where the structure 101 exists. The 3D point cloud map divided into grids is output in processing step S102.

[0028] In processing step S102, blind spots within the operation design area R are estimated. The processing in processing step S102 will be explained using the flowchart shown in Fig. 5. In processing step S102, first, in processing step S201, the route of a moving object traveling within the operation design area R is expressed as an undirected graph using nodes and edges, and the size of the moving object and sensor information mounted on the moving object are read.

[0029] Examples of the loaded route information are shown in Figures 7A and 7B. Figures 7A and 7B show the routes that each moving object moves in the two-dimensional area represented by the two-dimensional grid 106 shown in Figure 6, with edges represented by lines such as 107 and nodes represented by circles such as 108. The moving object moves along the route indicated by the edge 107 represented by a line, and checks for blind spots using the positions of the nodes 108 represented by circles.

[0030] Next, in processing step S202, the route information of each mobile body read in processing step S201 is superimposed. Here, the route of a single mobile body is as shown in FIG. 7A or FIG. 7B, but by superimposing the routes of multiple mobile bodies or previously experienced mobile body routes, it is possible to form a route for the entire operation design domain R. An example of superimposition is shown in FIG. 8. The superimposed mobile body route information will include the movement routes of all mobile bodies, as shown in 109. Here, only edge information is shown in 109, but the route information also includes node information.

[0031] Next, in processing step S203, common route information along which the moving bodies travel is extracted. The extraction of common routes is performed for each grid. The extraction is performed based on whether the routes of each moving body are within the same grid when viewed from directly above as shown in FIG. 6. For this reason, in this embodiment, only the XY directions are taken into consideration, and differences in the Z direction are not taken into consideration. In this embodiment, the common routes of FIGS. 7A and 7B are extracted, and therefore the route information (nodes, edges) shown in 110 of FIG. 9 is extracted.

[0032] Next, in processing step S204, blind spot observation points are set based on the route information. The method for setting blind spot observation points will be explained using the area 111 surrounded by the dashed line in Figure 9 as an example. The blind spot observation point is set at the midpoint 116 (dotted line) of the intersection position 115 between the coordinates 112 of each node and the grid boundary at the end of a perpendicular line 114 extended in the direction in which the edge 113 extends.

[0033] The results of setting these blind spot observation points for each node in each grid are shown enlarged in Figure 9. In the enlarged example in Figure 9, since two nodes 112 are included in each grid, a total of four midpoints 116 are generated for each perpendicular line 114, and a total of 12 midpoints 116 shown by dotted lines are generated for the three grids.

[0034] Next, as shown in processing steps S205 to S207 in Figure 5, processing steps S208 to S211 are repeated for each target grid at each node, for the number of moving objects. Here, L3: Number of Target Grids in processing step S207 indicates the number of destination grids that a moving object must be able to recognize when moving along a route, and is determined from the acceleration / deceleration performance of the moving object and the upper limit ODD speed set as an ODD condition. In this example, the description will be given on the assumption that a moving object must be able to recognize up to three grids ahead.

[0035] In processing step S208, an area that can be recognized by the sensor mounted on the moving object is extracted based on the position of the moving object. The area extraction method will be explained using area 116 in FIG. 10. Here, in the enlarged view of area 116 in FIG. 10, it is assumed that a moving object exists at position and orientation 117 and that recognition is required up to three grid points ahead. In this case, it is necessary to be able to observe objects installed at observation points set up to node 118. Therefore, it is necessary to be able to observe objects installed at observation points 119 to 121 in addition to the observation point near node 118. In this example, the horizontal observation angle of the LiDAR is set to 120 degrees. FIG. 11 is a diagram showing the observation range under the above conditions. When a moving object exists at 117 in FIG. 11, the observation range is the area indicated by the dashed line 123.

[0036] Assuming that a LiDAR is mounted on a moving object, we will explain how to extract point clouds observable by the LiDAR. In this article, in addition to the position of a 3D point expressed by X, Y, and Z, the 3D point also has the index of the grid to which it belongs, and binary information to determine whether it is a point cloud of a map or a point cloud of an object installed on the observation point.

[0037] First, the points contained in the 3D point cloud map are transformed into a coordinate system centered on the moving object using the following equation (1). In equation (1), p is the position of an arbitrary point, R is a rotation matrix expressed as a 3x3 orthogonal matrix, t is a translation vector, and p' is the point position after transformation. Here, R and t are the position and orientation of the moving object on the 3D point cloud map.

[0038]

number

[0039] Next, the point cloud is inverted using equation (2): where pi is a point in the 3D point cloud map with the position and orientation of the moving object as the origin, pia is the distance from the origin of each point, and r is a value less than or equal to 0.

[0040]

number

[0041] Next, the convex hull structure of the inverted point cloud is found, and the point cloud that forms the vertices of the convex hull is extracted. There are existing algorithms for extracting the convex hull, such as the incremental method and the divide-and-conquer method, so please refer to those methods for details. The 3D point cloud extracted here covers the entire periphery of the moving object, so only points where the vertical angle radv calculated using equation (3) is between 45° and -45°, and the horizontal angle radh calculated using equation (4) is between 60° and -60° are considered to be observable 3D points. In this case, if a LiDAR capable of 360° sensing is used, equation (4) is not necessary.

[0042]

number

[0043]

number

[0044] Next, an extraction method for point clouds observable when a camera is mounted on a moving object will be described. Even when extracting point clouds observable by a camera, the conversion of point cloud coordinates in the (1) and (2) equations performed with LiDAR is carried out to extract the point clouds that become the vertices of the convex hull. Then, the positions on the image are calculated by the following (5) and (6) equations.

[0045]

Number

[0046]

Number

[0047] Next, it is determined whether to project onto the image depending on whether the xi and yi calculated by the (5) and (6) equations are within the preset image size. For example, when the width of the image is 1920 and the height is 1080, it is determined that it is projected into the image when 0 < xi < 1920 and 0 < yi < 1080 are satisfied. Also, at this time, if the point cloud belonging to the grid set with dim illumination is projected onto the image, even if it is within the image, it will not be shown in the image and is set as an invalid value. This time, the (5) and (6) equations were used for explanation assuming a pinhole camera, but another model such as a fish-eye camera may also be used.

[0048] Fig. 12A is a diagram showing the case where a cylindrical object is arranged as point cloud information at the observation point, and Fig. 12B is a diagram showing an example of an image calculated based on the coordinates, and there are objects in the observation range. For example, when a cylindrical object is arranged as point cloud information at the observation point 124 in Fig. 12A, an image as shown at 125 on the left in Fig. 12B is obtained based on the coordinates calculated by the (5) and (6) equations. Similarly, when a cylindrical object is arranged as group information at the observation point 126 in Fig. 12A, an image as shown at 127 on the right in Fig. 12B is obtained.

[0049] As another example, Fig. 13A shows a case where a cylindrical object is placed at an observation point as point cloud information, and Fig. 13B shows example images calculated based on the coordinates, showing a case where an object is present in the observation range and a case where an object is not present. When a cylindrical object is placed at observation point 128 in Fig. 13A as point cloud information, an image is obtained in which a portion of the cylindrical object is visible, as shown at 129 in the left diagram of Fig. 13B. Furthermore, when a cylindrical object is placed at observation point 130 in Fig. 13A as point cloud information, an image is obtained in which the cylinder is not visible, as shown in the right diagram of Fig. 13B.

[0050] Next, in processing step S209, if LiDAR is installed, the number of points that satisfy equations (3) and (4) is counted, and if a camera is installed, the number of points whose coordinates calculated by equations (5) and (6) fall within the image is counted. If the counting result shows that the number of points having the attribute of the point cloud of an object installed at the observation point is equal to or greater than a preset threshold, the area where the observation point is located is determined to be an observable area in processing step S210, and if it is less than the threshold, the area where the observation point is located is determined to be an unobservable area in processing step S201.

[0051] 12A and 12B are determined to be observable regions that satisfy the conditions, while the examples shown in Figures 13A and 13B are determined to be unobservable regions that do not satisfy the conditions. Additionally, while the observability of objects within the target region was determined based on the number of cylindrical objects placed on the observation points that were observed, it is also possible to run an algorithm that recognizes cylindrical objects on an image generated based on the coordinates calculated by equations (5) and (6), and determine the region as observable if the algorithm recognizes a cylindrical object.

[0052] By performing the processing of processing steps S208 to S211 in Fig. 5, the observable area (safe area) and unobservable area (blind area) seen from the moving body at position 116 in Fig. 10 described in this embodiment become as shown in Fig. 14. In Fig. 14, 130 is the blind area and 131 is the safe area.

[0053] The above processing steps S208 to S211 are repeated for all moving bodies and all nodes extracted as common areas, and the safe area and blind area of ​​each moving body as seen from each node are estimated. Information on the estimated safe area and blind area of ​​each moving body as seen from each node is output to processing step S103.

[0054] Next, the risk map estimation shown in processing step S103 of Fig. 2 will be described using the flowchart of Fig. 15. In the risk map estimation of processing step S103, first, in processing step S301, the grid of the risk map is initialized as a visible area (safe area).

[0055] Next, in processing steps S302 and S303, processing steps S304 to S307 are repeated for the number of moving bodies and nodes. In processing step S304, safe / blind spot information for each grid seen from each node, estimated by each node, is extracted, and in processing step S306, this is compared with the risk map initialized in processing step S301. In processing step S306, if the corresponding grid information on the risk map is a safe area and the grid information seen from the node is a blind spot area, the corresponding area on the risk map is classified as blind spot information.

[0056] As a result of this processing, any grid on the risk map that has been determined to be a blind spot area at least once becomes a blind spot area. When processing step S103 is performed on the area in which the sensor placement is estimated in this embodiment, the risk map shown in Figure 16 is obtained. At this time, to reiterate, the diagonal lines in 133 represent blind spots, and the horizontal lines in 134 represent safe areas. In the example shown, there are three blind spot areas, 133A, 133B, and 133C.

[0057] Next, detailed processing of the sensor arrangement estimation processing step S104 in Fig. 3 will be described with reference to the flowchart in Fig. 17. In processing step S401, which is the first step of the detailed processing of the sensor arrangement estimation processing step S104, the sensor installation position is estimated.

[0058] The method for estimating the sensor installation position will be described with reference to Fig. 18A, which shows a case where a cylindrical object is placed at an observation point as point cloud information, and Fig. 18B, which shows an example image calculated based on coordinates. At this time, candidates for the sensor installation position are calculated for a one-grid area 135 (an area consisting of nine grids) surrounding a blind spot area 133. In the example of Fig. 18A, the candidates are a one-grid area 135A surrounding a blind spot area 133A, a one-grid area 135B surrounding a blind spot area 133B, and a one-grid area 135C surrounding a blind spot area 133C.

[0059] First, for the area of ​​one surrounding grid (135A, 135B, 135C in FIG. 18A), attention is paid to the boundary between the area where no object exists within the grid and the area where an object exists. In the upper part of FIG. 18B, the white part on the left is the area where no object exists within the grid, and the part with a background color on the right is the area where an object exists. Positions 136 on the boundary between these areas, as shown by 137 to 139 in FIG. 18B, are calculated.

[0060] Next, in the lower part of Fig. 18B, the area where an object exists within the grid (140 in Fig. 18B) is estimated from the point cloud information included in the 3D map, and 137 to 139 are moved in the shortest way to the position closest to 140. This operation can calculate the position of 141 by, for example, extracting an observable point cloud from 138, estimating plane parameters using RANSAC or the like, and finding perpendicular lines from this plane and positions 137 to 139.

[0061] At this time, any value may be given to the height information as the position. The positions calculated here are registered as candidate sensor installation positions. In this embodiment, the position of a circle such as that shown by 142 in FIG. 18A is calculated as a candidate sensor installation position.

[0062] Next, in processing steps S402 to S404, the processing of processing step S405 is repeated. Here, the number of sensors is the number of candidate sensor installation positions estimated in processing step S401, the sensor types are the number of sensor types stored in the sensor information 40, and the orientation candidates are the results of changing the orientation by, for example, 10 degrees around the Z axis with respect to a certain installation position.

[0063] In processing step S405, the blind spot area observable from the sensor is estimated based on the input installation position, sensor type, and posture. The blind spot area observable is estimated by repeating processing steps S207 to S211 in Figure 5, and is determined on a grid-by-grid basis. The position of the observation point set within the grid is determined using the position set in processing step S102.

[0064] Next, in processing step S406, combinatorial optimization is performed to estimate a combination of sensor installation positions that will eliminate all blind spots within the target area, using the information on observable blind spots determined for each sensor installation position, sensor type, and posture. In this example, combinations that satisfy the following constraints are calculated using a greedy method as a set cover problem. Note that there are known methods for solving combinatorial optimization problems, so these can be referenced for details.

[0065] In this case, it is advisable to consider the following constraints: for example, blind spots must be fully observed with a number of sensors equal to or greater than the sensing multiplicity, the total cost of the installed sensors must be within a preset price range, and multiple sensors must not be installed in the same location.

[0066] Here, sensing multiplicity refers to the minimum number 144 of sensors that must be observing the blind spot area represented by the diagonal line 143 in Fig. 19. In this case, this number was set to the number of moving objects traveling in the corresponding grid minus 1. This is because there is a possibility that a new blind spot will be created by a moving object traveling in the corresponding area, which may result in a collision between moving objects.

[0067] At this time, especially when optimizing sensor placement in a simple area like this, multiple sensor placement locations are calculated. At this time, the user is presented with a set of installation locations and risk maps for multiple configurations, such as a configuration with the smallest total sensor cost, a configuration with the smallest number of sensors, or a configuration with the highest sensing multiplicity.

[0068] An example of presentation is shown in Fig. 20. The information presented to the user is assumed to be the installation position and orientation 145 of the sensor, the state 146 of the area through which multiple moving objects pass, and the number 147 of sensors observing each area.

[0069] By carrying out the above series of processes, the installation positions of the infrastructure sensors are estimated. [Example]

[0070] 21 is a diagram illustrating a configuration example of a sensor installation position estimation device according to Example 2 of the present invention. Example 2 illustrates a case in which a mobile sensor identification unit 15 is added to the configuration of FIG. 1, and the sensor information 40 includes a mobile sensor that is equipped with a sensor and moves to a predetermined position to monitor the surroundings.

[0071] This embodiment is suitable for application to cases where the routes and number of moving objects moving within an area change over time. For example, in addition to the moving objects shown in Figures 7A and 7B, a moving object moving along route 148 in Figure 22A is added between 09:00 and 12:00, and a moving object moving along route 149 in Figure 22B is added between 13:00 and 17:00.

[0072] In this case, the risk map from 09:00 to 12:00 will be the risk map shown in 150 of Figure 23A, and from 13:00 to 17:00 will be the risk map shown in 151 of Figure 23B. In these risk maps, 133A, 133B, 133C, 133D, 133E, 133F, 133G, and 133H represent blind spots. If only sensors with fixed positions can be used for this risk map, as shown in Example 1, for example, it is necessary to install sensors in a configuration that allows observation of all blind spots (133A, 133B, 133C, 133D, 133E, 133F, 133G, and 133H).

[0073] On the other hand, when mobile sensors are available, two types of placements are determined in advance as shown in Figures 24A and 24B, and the mobile sensor 152 in Figure 24A is autonomously moved to position 153 in Figure 24B depending on the time, thereby enabling sensor placement according to changes in blind spots to be achieved with a small number of sensors without human intervention.

[0074] In Example 2, a risk map is created from the number and routes of moving objects for each time period, blind spots that change over time (133E, 133F and 133G, 133H in Figures 24A and 24B) are extracted, and it is determined whether the sensor observing the relevant area can be a mobile sensor.

[0075] For this reason, in the second embodiment, the blind spot estimation processing step S102, the risk map generation processing step S103, and the sensor placement estimation processing step S104 are repeated twice, that is, the number of times the configuration changes over time.

[0076] The mobile sensor is identified by the mobile sensor identification unit 15, which is newly added in Fig. 21. The mobile sensor identification unit 15 first extracts the changing blind spot area. The changing blind spot area is extracted by comparing the state of each grid of the risk map in the time direction.

[0077] Next, the sensors observing the blind spot areas that change over time are extracted, in this embodiment, the sensor 152 observing 133E and 133F and the sensor 153 observing 133G and 133H. The sensors observing the changing blind spot areas are extracted based on information about which areas each sensor was observing when the sensor arrangement was estimated.

[0078] Next, a pair is created between sensors 152 and 153 that are observing the changing blind spot area, and the blind spot area that can be observed is estimated when sensor 153 is placed at the position of 152, and when sensor 152 is placed at the position of 153. If the blind spot area that needs to be observed can be observed at both positions, the paired sensor is replaced with a mobile sensor.

[0079] The above processing makes it possible to design a sensor layout that takes mobile sensors into consideration. [Example]

[0080] In the third embodiment, a configuration will be described in which a sensor installation position 60 is added to the configuration in Fig. 1. The sensor installation position 60 is information used when the sensor placement estimation unit 14 estimates the sensor installation position, and describes the position and orientation at which the sensor can be installed.

[0081] When the sensor installation possible location 60 is input, the processing of processing step S401 in FIG. 17 is omitted, and the sensor installation candidate location listed in the sensor installation possible location 60 is used instead of the sensor installation candidate location estimated by processing step S401.

[0082] For example, suppose seven candidate sensor installation locations (154 to 160) are given for a risk map as shown in FIG. 26. In this case, the optimal combination is estimated from the given candidate sensor installation locations, and the sensor placements shown in FIGS. 27A and 27B are obtained. FIG. 27A is a diagram showing an example of mobile sensor installation in the case of FIG. 23A, and FIG. 27B is a diagram showing an example of mobile sensor installation in the case of FIG. 23B. In this case, 133A and 133B are observed from 155 in FIG. 27A, 133B and 133E are observed from 156, and 133C and 133D are observed from 157. In addition, in FIG. 27B, 133A and 133B are observed from 155, 133D and 133E are observed from 158, and 133B and 133C are observed from 159.

[0083] 27A and 27B show the case where all blind spots are observed. In this case, the estimated sensor installation positions and the risk map for when sensors are installed in those locations are output.

[0084] Next, when a candidate sensor installation position as shown in Fig. 28 is input, the result shown in 160 in Fig. 29 is obtained as the sensor installation position. At this time, since there is no sensor that can observe the area 161 among the candidate sensor installation positions, the user is informed that the input candidate sensor installation position cannot observe all blind spot areas, and the blind spot areas that cannot be observed are displayed.

[0085] Although the embodiments of the present invention have been described above, the present invention is not limited to the above-described embodiments, and various modifications can be made without departing from the scope of the claims. For example, the above-described embodiments provide detailed descriptions of the present invention, and it is not necessary to include all of the described configurations. Furthermore, configurations of other embodiments can be added to the configuration. Furthermore, some of the configurations can be added, deleted, or replaced. Furthermore, while the configuration of the sensor location estimation function 10 shown in FIG. 1 has been described assuming that all of the functions are located on the same computer, the processing described herein may be performed by communication with only some of the functions, for example, with the sensor location estimation unit 14 located on the cloud and the remaining functions located on a computer. [Explanation of symbols]

[0086] 10: Sensor installation position estimation function, 11: Map information conversion unit, 12: Blind spot estimation unit, 13: Risk map estimation unit, 14: Sensor placement estimation unit, 20: 3D map, 30: Moving object information, 40: Sensor information, 50: Sensor information

Claims

1. a map information conversion unit that divides a three-dimensional map of a target area in which a mobile object performs limited-area autonomous driving into grids; a blind spot estimation unit that estimates a blind spot of a grid within a moving path of a moving object in the target area; a risk map generation unit that generates a blind spot map for the target area from the plurality of blind spots estimated by the blind spot estimation unit; a sensor placement estimation unit that estimates placement positions of sensors based on the types and placement positions of sensors to be installed within the target area; the blind spot estimation unit classifies, based on the position of the moving object within the grid, a destination grid that the moving object must be able to recognize when moving on a route into a safety area that can be observed by a sensor mounted on the moving object and a blind spot area that cannot be observed; The sensor installation position estimation device is characterized in that the sensor installation position estimation unit calculates the sensor installation position based on preset sensor installation position candidates.

2. The sensor installation position estimation device according to claim 1, A three-dimensional map of a target area in which a mobile body will perform limited-area autonomous driving, the route and sensor information of the mobile body in the target area, and candidate types and installation locations of sensors to be installed within the area are input, a risk map generation unit that generates a risk map using the estimated blind spots; and a sensor placement estimation unit that estimates sensor placement positions from candidate sensor types and placement positions to be installed within a target area.

3. A sensor installation position estimation method implemented using a computer device, comprising: The computer device divides a three-dimensional map of a target area in which the mobile body will perform limited-area autonomous driving into grids; Estimating a blind spot of a grid within a moving path of a moving object in the target area; generating a blind spot map for the target area from the estimated blind spots; Estimating the placement position of the sensor from the type and candidate installation positions of the sensor to be installed within the target area; Based on the position of the mobile object within the grid, the grid of the destination of the mobile object that must be recognized when the mobile object moves on the route is classified into a safe area that can be observed by a sensor mounted on the mobile object and a blind area that cannot be observed; A sensor installation position estimation method characterized by calculating a sensor installation position based on preset sensor installation position candidates.

Images