Estimation device, estimation program, estimation data generation method

Abstract

To provide an estimation device ensuring robust properties for estimation precision of a self-position.SOLUTION: A processor of an estimation device is configured so as to execute acquisition of a main estimation position Pd by dead reckoning based on inner field data acquired in an inner field, acquisition of a sub-estimation position Pm1 by marker matching based on marker observation data acquired by observation of a route marker in an outer field and marker position data of a route marker in an outer field, and generation of position estimation data Df of a self-position Pf by fusion for weighting the main estimation position Pd and the sub-estimation position Pm1 on the basis of a ratio of the respective estimation errors. Acquisition of the sub-estimation position Pm1 includes acquisition of the main estimation position Pd corrected by deviation between a marker prediction position predicted with the main estimation position Pd as a starting point from marker observation data and a marker installation position matched to a marker prediction position by marker position data as the sub-estimation position Pm1.SELECTED DRAWING: Figure 8

Description

Technical Field

[0004] , , , , ,

[0005] , , , ,

[0001] The present disclosure relates to an estimation technique for estimating the self-position of an autonomous driving device that autonomously drives in the outside world.

Background Art

[0002] The estimation technique disclosed in Patent Document 1 switches an estimation method for estimating the self-position of an autonomous driving device according to the estimation accuracy for each region simulated in advance. Specifically, in the first region, the self-position is estimated using point cloud data obtained by scanning the outside world from the autonomous driving device and map data of the outside world. On the other hand, in the second region, the self-position is estimated using the detection results from the autonomous driving device of a plurality of conductors installed in the outside world and the position information of the conductors in the outside world.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] However, in the estimation technique disclosed in Patent Document 1, the estimation accuracy for each estimation method in each region varies from moment to moment due to changes in external conditions such as the weather. Therefore, it is difficult to say that the true estimation accuracy can be ensured only by switching the estimation method, and it can be said that it lacks robustness.

[0005] An object of the present disclosure is to provide an estimation device that ensures robustness against the estimation accuracy of the self-position. Another object of the present disclosure is to provide an estimation program that ensures robustness against the estimation accuracy of the self-position. Yet another object of the present disclosure is to provide an estimation data generation method that ensures robustness against the estimation accuracy of the self-position.

Means for Solving the Problems

[0006] The following describes the technical means of solving the problem described in this disclosure. Note that the claims and the reference numerals in parentheses in this section indicate the correspondence with the specific means described in the embodiments detailed later, and do not limit the technical scope of this disclosure.

[0007] The first aspect of this disclosure is, An estimation device (1) having a processor (12) and estimating the self-position of an autonomous driving device (4) that autonomously drives in the external environment, The processor is Based on internal data (Di) acquired within the autonomous driving system, the main estimated position (Pd) is obtained as the self-position through dead reckoning, Based on marker observation data (Do) obtained from observations of multiple route markers (6) placed in the outside world by an autonomous driving device, and marker position data (Dmp) which defines the marker placement position (Pmp) of each route marker in the outside world, a sub-estimated position (Pm1) as the self-position is obtained by marker matching, The system is configured to generate self-position estimation data (Df) by fusion, which weights the main estimated position and sub-estimated position based on the ratio of their respective estimation errors. To obtain sub-estimated positions, This includes obtaining a sub-estimated position for the main estimated position, which is corrected by the deviation (δP) between the predicted marker position (Pp) predicted from the main estimated position based on marker observation data, and the marker placement position that matches the predicted marker position in the marker position data.

[0008] A second aspect of this disclosure is, An estimation program, which includes instructions to be executed by a processor (12) and is stored in a storage medium (10) for estimating the self-position of an autonomous driving device (4) that autonomously drives in the external environment, Based on internal data (Di) acquired within the autonomous driving system, the main estimated position (Pd) is obtained as the self-position through dead reckoning, Based on marker observation data (Do) obtained from observations of multiple route markers (6) placed in the outside world by an autonomous driving device, and marker position data (Dmp) which defines the marker placement position (Pmp) of each route marker in the outside world, a sub-estimated position (Pm1) as the self-position is obtained by marker matching, The command includes instructions to generate self-position estimation data (Df) by fusion, which weights the main estimated position and sub-estimated position based on the ratio of their respective estimation errors. To obtain sub-estimated positions, This includes obtaining a sub-estimated position for the main estimated position, which is corrected by the deviation (δP) between the predicted marker position (Pp) predicted from the main estimated position based on marker observation data, and the marker placement position that matches the predicted marker position in the marker position data.

[0009] A third aspect of this disclosure is: A method for generating estimated data, which is performed by a processor (12) to generate position estimation data (Df) obtained by estimating the self-position of an autonomous driving device (4) that autonomously drives in the external environment, Based on internal data (Di) acquired within the autonomous driving system, the main estimated position (Pd) is obtained as the self-position through dead reckoning, Based on marker observation data (Do) obtained from observations of multiple route markers (6) placed in the outside world by an autonomous driving device, and marker position data (Dmp) which defines the marker placement position (Pmp) of each route marker in the outside world, a sub-estimated position (Pm1) as the self-position is obtained by marker matching, This includes generating position estimation data by fusion, where the main estimated position and sub-estimated position are weighted based on the ratio of their respective estimation errors. To obtain sub-estimated positions, This includes obtaining a sub-estimated position for the main estimated position, which is corrected by the deviation (δP) between the predicted marker position (Pp) predicted from the main estimated position based on marker observation data, and the marker placement position that matches the predicted marker position in the marker position data.

[0010] According to these first to third embodiments, a main estimated position is obtained by dead reckoning based on internal data, and a sub-estimated position is obtained by marker matching based on marker observation data and marker position data. Here, in obtaining the sub-estimated position, attention is paid to the deviation between the marker prediction position predicted from the main estimated position based on marker observation data and the marker placement position that matches the marker prediction position in the marker position data. As a result, the estimation error of the sub-estimated position, which is obtained by correcting the main estimated position with the deviation between the marker prediction position and the marker placement position, can offset the influence of the estimation error of the main estimated position. Therefore, by weighting and fusing the main estimated position and the sub-estimated position based on the ratio of their respective estimation errors that can reflect changes in external conditions, it becomes possible to generate position estimation data that is robust to the accuracy of self-position estimation. [Brief explanation of the drawing]

[0011] [Figure 1] This block diagram shows the overall configuration of the estimation device according to the first embodiment. [Figure 2] This is a perspective view showing an autonomous driving device equipped with the estimation device according to the first embodiment, along with a route marker. [Figure 3] This is a plan view showing an autonomous driving device equipped with the estimation device according to the first embodiment, along with a route marker. [Figure 4] This is a block diagram showing the functional configuration of the estimation device according to the first embodiment. [Figure 5] This is a schematic diagram illustrating the estimation principle of the estimation device according to the first embodiment. [Figure 6] This is a schematic diagram illustrating the estimation principle of the estimation device according to the first embodiment. [Figure 7] It is a schematic diagram for explaining the estimation principle of the estimation device according to the first embodiment. [Figure 8] It is a flowchart showing the estimation flow according to the first embodiment. [Figure 9] It is a plan view showing an autonomous driving device on which an estimation device according to the second embodiment is mounted, together with route markers. [Figure 10] It is a schematic diagram for explaining the estimation principle of the estimation device according to the second embodiment.

Embodiments for Carrying Out the Invention

[0012] Hereinafter, a plurality of embodiments of the present disclosure will be described based on the drawings. In addition, in each embodiment, the same reference numerals may be assigned to corresponding components, and redundant explanations may be omitted. Further, when only a part of the configuration is described in each embodiment, the configuration of other embodiments described previously can be applied to other parts of the said configuration. Furthermore, not only the combinations of configurations explicitly shown in the description of each embodiment, but also the configurations of a plurality of embodiments can be partially combined with each other as long as there is no problem with the combination.

[0013] (First Embodiment) The estimation device 1 according to the first embodiment shown in FIG. 1 estimates the self-position of an autonomous driving device 4 that autonomously travels in the external world. Here, in the present embodiment, a three-dimensional absolute coordinate system is defined assuming the vertical direction with index Z with respect to the two horizontal directions with indices X and Y. Further, for the autonomous driving device 4 that takes a posture necessary for autonomous driving in such a three-dimensional orthogonal coordinate system, a three-dimensional relative coordinate system is defined assuming the vertical direction with index x, the horizontal direction with index y, and the height direction with index z. However, it is assumed that the vertical direction (Z direction) of the three-dimensional absolute coordinate system and the height direction (z direction) of the three-dimensional relative coordinate system substantially coincide.

[0014] The autonomous driving device 4 may be a logistics vehicle or logistics robot that autonomously travels along roads inside and outside warehouses in a logistics facility, which is its external environment, to transport goods (example in Figure 2). The autonomous driving device 4 may be an autonomous vehicle capable of autonomously traveling on roads in a traffic environment, which is its external environment. The autonomous driving device 4 may be a serving robot that autonomously travels along roads inside a restaurant or hospital, which is its external environment, to deliver food and beverages. The autonomous driving device 4 may be a disaster relief robot that autonomously travels through a disaster area, which is its external environment, to transport supplies or collect information. The autonomous driving device 4 may, of course, be any other type of device. Furthermore, any type of autonomous driving device 4 may receive remote driving assistance or driving control from an external center.

[0015] As shown in Figures 2 and 3, multiple route markers 6 are installed in at least a portion of the travel path or road that the autonomous driving device 4 is intended to autonomously travel in the external environment. Each route marker 6 is mainly composed of a magnetized magnetic marker, such as one made of a ferrite magnet. In the first embodiment, each route marker 6 is placed at intervals on a travel trajectory Tr that is assumed to continuously provide a representative travel position, such as the lateral center position in a three-dimensional relative coordinate system, along the travel route on which the autonomous driving device 4 autonomously travels. Each route marker 6 is attached to or embedded in the road surface at its respective location, for example, with a protective sheet. All or some of the route markers 6 may be equipped with RFID (radio frequency identification).

[0016] As shown in Figure 1, the estimation device 1 is mounted on the autonomous driving device 4 together with the sensor unit 2 and the map unit 3. The sensor unit 2 consists of a marker sensor 20, an external sensor 22, and an internal sensor 24.

[0017] The marker sensor 20 shown in Figures 1, 2, and 4 acquires marker observation data Do, which can be used to estimate the motion of the autonomous driving device 4 in the external environment surrounding the autonomous driving device 4. The marker sensor 20 is mainly composed of a magnetic sensor or magnetic receiver that acquires marker observation data Do by observing route markers 6 present in the external environment from the autonomous driving device 4. The marker sensor 20 may also include an RFID reader. It is preferable that the mounting position and observation reference position of such a marker sensor 20 are pre-set for the autonomous driving device 4 so that route markers 6 that enter the observation range of a set distance can be observed.

[0018] The external sensor 22 shown in Figures 1, 2, and 4 acquires information other than marker observation data Do as external data De, which can be used for motion estimation of the autonomous driving device 4 in the external environment surrounding the autonomous driving device 4. The external sensor 22 may also acquire external data De by detecting objects present in the external environment from the autonomous driving device 4. This detection type external sensor 22 may be one or more types from among LiDAR (Light Detection and Ranging / Laser Imaging Detection and Ranging), cameras, radar, and sonar. The external sensor 22 may also acquire external data De by receiving wireless signals from wireless communication systems present in the external environment from the autonomous driving device 4. This receiving type external sensor 22 may be one or more types from among among GNSS (Global Navigation Satellite System) receivers and ITS (Intelligent Transport Systems) receivers. In the following explanation, we will use as an example a case in which a LiDAR system, which acquires a three-dimensional point cloud image as external data De by beam scanning the external environment and detecting reflected beams from targets, constitutes the external sensor 22.

[0019] The internal environment sensor 24 shown in Figures 1 and 4 acquires information usable for motion estimation of the autonomous driving device 4 in the internal environment, which is the internal environment of the autonomous driving device 4, as internal environment data Di. The internal environment sensor 24 may also acquire internal environment data Di by detecting specific kinetic physical quantities in the internal environment of the autonomous driving device 4. This detection type of internal environment sensor 24 may be one or more types, such as an inertial sensor, a velocity sensor, and a steering angle sensor. In the following, we will explain using as an example a case in which the internal environment sensor 24 consists of an inertial sensor that acquires the angular velocity and / or attitude angle of the autonomous driving device 4 as internal environment data Di, and a velocity sensor that acquires the velocity of the autonomous driving device 4 as internal environment data Di.

[0020] The map unit 3 shown in Figures 1 and 4 is composed of one or more types of non-transitory tangible storage media, such as semiconductor memory, magnetic media, and optical media, for non-temporarily storing or remembering map data Dm. The map unit 3 may also be a locator database used for advanced driver assistance or automatic driving control of the autonomous driving device 4. The map unit 3 may also be a navigation system database that guides the driving of the autonomous driving device 4. The map unit 3 may also be a planning unit database that plans the driving of the autonomous driving device 4. The map unit 3 may be composed of a combination of multiple types of these databases, etc.

[0021] Map unit 3 acquires and stores the latest map data Dm, for example, through communication with an external center. Map data Dm is a three-dimensional digital map including a map point cloud that maps the external world in which the autonomous driving device 4 is autonomously driving, and for example, a high-precision dynamic map is used. Such map data Dm may include road information that represents one or more types of information, such as the location, shape, size, and road surface condition of the road or driving path. Map data Dm may also include marking information that represents one or more types of information, such as the location, shape, size, and operating status of signs, lane markings, and traffic lights attached to the road or driving path. Map data Dm may also include structural information that represents one or more types of information, such as the location, shape, and size of buildings, structures, and plants facing the road or driving path.

[0022] The estimation device 1 shown in Figures 1 and 4 is connected to the sensor unit 2 and the map unit 3 via one or more types of connections, such as a LAN (Local Area Network), wire harness, internal bus, and wireless communication line. The estimation device 1 may also be an ECU (Electronic Control Unit) dedicated to driving control, which performs advanced driving assistance or automatic driving control of the autonomous driving device 4. The estimation device 1 may also be a locator ECU used for advanced driving assistance or automatic driving control of the autonomous driving device 4. The estimation device 1 may also be a navigation device ECU that navigates the driving of the autonomous driving device 4. The estimation device 1 may also be a communication control device ECU that controls communication between the autonomous driving device 4 and the outside world.

[0023] As shown in Figure 1, the estimated device 1 is a dedicated computer comprising at least one memory 10 and one processor 12. The memory 10 is one or more types of non-transitory tangible storage medium, such as semiconductor memory, magnetic media, and optical media, which non-temporarily store or remember programs and data that can be read by the computer. The processor 12 includes one or more types as cores, such as a CPU (Central Processing Unit), GPU (Graphics Processing Unit), and RISC (Reduced Instruction Set Computer)-CPU.

[0024] The processor 12 executes multiple instructions contained in the estimation program stored in the memory 10. As a result, the estimation device 1 constructs multiple functional blocks for estimating the autonomous driving device 4's own position, as shown in Figure 4. In this way, the estimation device 1 constructs multiple functional blocks by having the processor 12 execute multiple instructions from the estimation program stored in the memory 10, which is used to estimate the autonomous driving device 4's own position.

[0025] The multiple functional blocks constructed by the estimation device 1 include a dead reckoning block 100, a marker matching block 110, a map matching block 120, and a fusion block 130. The dead reckoning block 100 estimates the self-position of the autonomous driving device 4 at the current estimation time t by dead reckoning based on internal data Di and a dynamics model, and obtains this estimation result as the main estimated position Pd (see Figure 5, which will be described in detail later). Here, the dynamics model is a model of the behavior of the autonomous driving device 4 based on dynamics, such as a two-wheeled vehicle model.

[0026] Therefore, the three-dimensional position coordinates Xd, Yd, and Zd that constitute the main estimated position Pd at the estimated time t are defined by equations 1, 2, and 3, respectively, using the three-dimensional position coordinates Xf, Yf, and Zf (see Figure 5) estimated by the fusion block 130 at the previous estimated time t-1, as will be described in detail later. Here, V in equations 1, 2, and 3 is input to the speed of the autonomous driving device 4 as internal data Di detected by the speed sensor among the internal sensors 24 at the estimated time t. δt in equations 1, 2, and 3 is input to the estimated interval (i.e., sampling period) between the estimated time t and the previous estimated time t-1. ωp in equation 3 is input to the pitch angle among the attitude angles of the autonomous driving device 4 as internal data Di detected by the inertial sensor among the internal sensors 24 at the estimated time t. Furthermore, in Figure 5 and Figure 10 described later, as well as in equations 1-3 and equations 4-34 described later, the parameters (variables) that need to be distinguished at each estimated time t and t-1 are expressed in the corresponding functional forms [t] and [t-1] for estimated times t and t-1, respectively.

number

number

number

[0027] Furthermore, θd in equations 1 and 2 is the yaw angle among the attitude angles at the currently estimated time t, and is defined by equation 4, using the yaw angle θf (see Figure 5) estimated by the fusion block 130 at the previous estimated time t-1, as will be described in detail later. Here, γ in equation 4 is the yaw rate of the autonomous driving device 4, which is input as internal data Di detected at the currently estimated time t by the inertial sensor among the internal sensors 24. δt in equation 4 is input as the estimated interval between the currently estimated time t and the previous estimated time t-1.

number

[0028] In conjunction with the estimation of the main estimated position Pd at the estimated time t, the dead reckoning block 100 obtains the predicted error in the estimation as the main estimation error ΔPd shown in Figure 4. At this time, the main estimation error ΔPd at the estimated time t is composed of the three-dimensional position errors ΔXd, ΔYd, and ΔZd, respectively, for the estimated three-dimensional position coordinates Xd, Yd, and Zd.

[0029] Therefore, the three-dimensional position errors ΔXd, ΔYd, and ΔZd at the current estimated time t are defined by equations 5, 6, and 7, respectively, using the three-dimensional position errors ΔXf, ΔYf, and ΔZf estimated by the fusion block 130 at the previous estimated time t-1, as will be described in detail later. Here, V in equations 5, 6, and 7 is input to the velocity at the current estimated time t, as measured by the velocity sensor among the internal sensors 24. ΔV in equations 5, 6, and 7 is input to the velocity error predicted as an error characteristic of the velocity sensor among the internal sensors 24. δt in equations 5, 6, and 7 is input to the estimated interval between the current estimated time t and the previous estimated time t-1. θd in equations 5 and 6 is input to the estimated yaw angle from equation 4 at the current estimated time t. Δωp in equation 7 is input to the pitch angle error predicted as an error characteristic of the inertial sensor among the internal sensors 24.

number

number

number

[0030] Furthermore, Δθf in equations 5 and 6 is the yaw angle error estimated by the fusion block 130 at the previous estimated time t-1, as will be described in detail later. The error Δθd of the yaw angle θd estimated at the current estimated time t is defined by equation 8, using the previous yaw angle error Δθf. Here, Δγ in equation 8 is the yaw rate error predicted as the error characteristic of the inertial sensor among the internal sensors 24. δt in equation 8 is the estimated interval between the current estimated time t and the previous estimated time t-1.

number

[0031] Based on the above, the dead reckoning block 100 estimates the three-dimensional position coordinates Xd, Yd, and Zd (numbers 1, 2, and 3) that give the main estimated position Pd, along with the yaw angle θd (number 4), with respect to the current estimated time t. Accordingly, the dead reckoning block 100 estimates the three-dimensional position errors ΔXd, ΔYd, and ΔZd (numbers 5, 6, and 7) that give the main estimated error ΔPd, along with the yaw angle error Δθd (number 8), with respect to the current estimated time t. The dead reckoning block 100 then stores these estimation results in memory 10, associated with the current estimated time t.

[0032] The marker matching block 110 shown in Figure 4 estimates the self-position of the autonomous driving device 4 at the estimated time t by marker matching based on marker observation data Do and marker position data Dmp, and obtains the estimation result as the first sub-estimated position Pm1. At this time, the marker observation data Do is the observation position po of the route marker 6 observed by the marker sensor 20 at the estimated time t, and is composed of three-dimensional relative coordinates xo, yo, zo with respect to the autonomous driving device 4, as shown in Figure 5.

[0033] Therefore, the estimated time t that is the target of the marker matching process in the marker matching block 110 is the timing of acquisition of the input marker observation data Do. Here, the timing of acquisition of marker observation data Do may be set for each observation period of the route marker 6 by the marker sensor 20 in the area where the route marker 6 is installed. Alternatively, the timing of acquisition of marker observation data Do may be set to the timing when the route marker 6 is actually observed by the marker sensor 20 in the area where the route marker 6 is installed.

[0034] Therefore, the marker matching block 110 predicts the marker predicted position Pp shown in Figure 5, starting from the main estimated position Pd at the same time t by the dead reckoning block 100, using the marker observation data Do at the current estimated time t. For this reason, the current estimated time t, at which the dead reckoning process is executed in the dead reckoning block 100, which inputs the main estimated position Pd to the marker matching block 110, is effectively synchronized with the timing of acquisition of the marker observation data Do.

[0035] The predicted three-dimensional position coordinates Xp, Yp, and Zp for the marker prediction position Pp are obtained by correcting the three-dimensional position coordinates Xd, Yd, and Zd of the main estimated position Pd using coordinate transformation values ​​between the three-dimensional relative coordinates xo, yo, and zo of the marker observation position po and the mounting position and observation reference position of the marker sensor 20. Therefore, the prediction error for the three-dimensional position coordinates Xp, Yp, and Zp of the marker prediction position Pp includes the main estimation error ΔPd for the three-dimensional position coordinates Xd, Yd, and Zd of the main estimated position Pd, and the observation error for the three-dimensional relative coordinates xo, yo, and zo of the marker observation data Do.

[0036] For this marker observation data Do, the marker position data Dmp in Figure 4 is input as the marker placement position Pmp in Figure 5, which is the setting position of each route marker 6 in the outside world, and includes the three-dimensional position coordinates Xmp, Ymp, Zmp of each marker 6. At this time, the marker position data Dmp may be read from the map unit 3 as part of the map data Dm (in the case of Figures 1 and 4). The marker position data Dmp may be stored in the map unit 3 or memory 10 in accordance with the map data Dm described above, and may be read from that storage location.

[0037] The marker matching block 110 then recognizes the marker placement position Pmp that matches the predicted marker position Pp at the estimated time t, as shown in Figure 5, in the marker position data Dmp, and the deviation δP between the predicted marker position Pp and the marker position Pp. By correcting the main estimated position Pd at the estimated time t using the recognized deviation δP, the marker matching block 110 obtains the first sub-estimated position Pm1 at the same time t.

[0038] At this time, the three-dimensional position coordinates Xm1, Ym1, and Zm1 in Figures 4 and 5 that constitute the first sub-estimated position Pm1 are defined by equations 9, 10, and 11, respectively. That is, each three-dimensional position coordinate Xm1, Ym1, and Zm1 is obtained by subtracting the deviation between the three-dimensional position coordinates Xp, Yp, and Zp of the marker prediction position Pp and the three-dimensional position coordinates Xmp, Ymp, and Zmp of the marker installation position Pmp, respectively, as the deviation δP in Figure 5, from the three-dimensional position coordinates Xd, Yd, and Zd of the main estimated position Pd. At the same time, the yaw angle θm1 in Figures 4 and 5, which is the attitude angle at the current estimated time t, is obtained by correcting the estimated value θf at the previous estimated time t-1 based on the deviation between the coordinates Yp and Ymp of each position Pp and Pmp, and the distance traveled by the autonomous driving device 4 during the estimated interval between times t and t-1.

number

number

number

[0039] In conjunction with the estimation of the first sub-estimated position Pm1 at the estimated time t, the marker matching block 110 obtains the predicted error in the estimation as the first sub-estimate error ΔPm1 shown in Figure 4. At this time, the first sub-estimate error ΔPm1 at the estimated time t is composed of the three-dimensional position errors ΔXm1, ΔYm1, and ΔZm1, respectively, for the estimated three-dimensional position coordinates Xm1, Ym1, and Zm1.

[0040] Here, from the relationship between the second and third terms on the right-hand side of equations 9, 10, and 11, each three-dimensional position error ΔXm1, ΔYm1, and ΔZm1 is influenced by the prediction error related to the three-dimensional position coordinates Xp, Yp, and Zp of the marker prediction position Pp, and the map error related to the three-dimensional position coordinates Xmp, Ymp, and Zmp of the marker installation position Pmp, respectively. Of these errors, in particular, the prediction error related to the three-dimensional position coordinates Xp, Yp, and Zp in the second term on the right-hand side includes the main estimation error ΔPd related to the three-dimensional position coordinates Xd, Yd, and Zd of the main estimation position Pd, and the observation error related to the three-dimensional relative coordinates xo, yo, and zo of the marker observation data Do, as described above.

[0041] Therefore, the main estimation error ΔPd related to the three-dimensional position coordinates Xd, Yd, and Zd is substantially offset from the influence on each three-dimensional position error ΔXm1, ΔYm1, and ΔZm1 due to the relationship between the first and second terms on the right-hand side of equations 9, 10, and 11. Thus, each three-dimensional position error ΔXm1, ΔYm1, and ΔZm1 is left with observation errors related to the three-dimensional relative coordinates xo, yo, and zo, and map errors related to the three-dimensional position coordinates Xmp, Ymp, and Zmp, respectively. Accordingly, each three-dimensional position error ΔXm1, ΔYm1, and ΔZm1 are predetermined to different or identical fixed values ​​based on the observation error predicted as the error characteristics of the marker sensor 20 relative to the route marker 6 and the measurement error of the marker position data Dmp relative to the route marker 6. At this time, the error Δθm1 predicted for the estimation of the yaw angle θm1 at the current estimation time t is also predetermined to a similar fixed value.

[0042] Based on the above, the marker matching block 110 estimates the three-dimensional position coordinates Xm1, Ym1, and Zm1 of numbers 9, 10, and 11 that give the first sub-estimated position Pm1, along with the yaw angle θm1, with respect to the current estimated time t. Accordingly, the marker matching block 110 estimates the three-dimensional position errors ΔXm1, ΔYm1, and ΔZm1 that give the first sub-estimated error ΔPm1, along with the yaw angle error Δθm1, with respect to the current estimated time t. The marker matching block 110 then stores these estimation results in memory 10 in association with the current estimated time t.

[0043] The map matching block 120 shown in Figure 4 estimates the self-position of the autonomous driving device 4 at the estimated time t by map matching based on external data De and map data Dm, and acquires this estimation result as the second sub-estimated position Pm2. At this time, the external data De is input to a three-dimensional point cloud image Des, which includes the scanned point cloud Cs in a three-dimensional relative coordinate system, scanned by the LiDAR of the external sensor 22 at the estimated time t as shown in Figures 6 and 7.

[0044] Therefore, the estimated time t at which the map matching process is executed in the map matching block 120 is the acquisition timing of the three-dimensional point cloud image Des, which is input as external data De. Here, the acquisition timing of the three-dimensional point cloud image Des is set for each scanning cycle of the external environment by the LiDAR sensor 22.

[0045] For this three-dimensional point cloud image Des, the map data Dm shown in Figures 6 and 7 is a three-dimensional digital map Dmm containing a map point cloud Cm that maps objects in the external environment of the autonomous driving device 4. At this time, the map point cloud Cm, which maps external objects in the three-dimensional absolute coordinate system of the three-dimensional digital map Dmm, is transformed into a three-dimensional relative coordinate system. The coordinate-transformed map point cloud Cm is then allocated to multiple three-dimensional voxels Vm that divide the external environment in the three-dimensional relative coordinate system. Note that in Figures 6 and 7, the three-dimensional voxels Vm are schematically illustrated in two dimensions, vertical (x direction) and horizontal (y direction), with the height direction (z direction) omitted.

[0046] Upon receiving input involving such coordinate transformations and allocations, the map matching block 120 recognizes the map point cloud Cm that matches the scanned point cloud Cs of the three-dimensional point cloud image Des, as shown in Figures 6 and 7, along with the corresponding scanned point cloud Cs in the map matching. The map matching block 120 then converts the position coordinates of the map point cloud Cm, matched with the scanned point cloud Cs, back to the three-dimensional absolute coordinate system, thereby obtaining the three-dimensional position coordinates Xm2, Ym2, and Zm2 that constitute the second sub-estimated position Pm2 at the estimated time t, as shown in Figure 4. At this time, the yaw angle θm2, shown in Figure 4, among the attitude angles at the estimated time t, is also estimated by map matching.

[0047] In conjunction with the estimation of the second sub-estimated position Pm2 at the estimated time t, the map matching block 120 obtains the predicted error in the estimation as the second sub-estimate error ΔPm2 shown in Figure 4. At this time, the second sub-estimate error ΔPm2 at the estimated time t is composed of the three-dimensional position errors ΔXm2, ΔYm2, and ΔZm2, respectively, for the estimated three-dimensional position coordinates Xm2, Ym2, and Zm2.

[0048] Therefore, the three-dimensional position errors ΔXm2, ΔYm2, and ΔZm2 at the estimated time t are defined by equations 12, 13, and 14, using the necessary parameters from the point cloud quantity-dependent likelihoods Lxs, Lys, and Lzs, and the noise quantity-dependent likelihood Ln, respectively, in the three-dimensional relative coordinate system. Here, the functions fx, fy, and fz in equations 12, 13, and 14 are functions that decrease the corresponding three-dimensional position errors ΔXm2, ΔYm2, and ΔZm2 as the corresponding likelihoods Lxs, Lys, Lzs, and Ln increase, respectively.

number

number

number

[0049] Furthermore, the map matching block 120 obtains the predicted error Δθm2 in the estimation of the yaw angle θm2 at the estimated time t. In this case, the yaw angle error Δθm2 is defined by equation 15, which uses the point cloud quantity-dependent likelihood Lys and the noise quantity-dependent likelihood Ln. Here, the function fθ in equation 15 is a function that decreases the yaw angle error Δθm2 as the likelihoods Lys and Ln increase.

number

[0050] In numbers 12, 13, 14, and 15, the point cloud quantity-dependent likelihoods Lxs, Lys, and Lzs are input by accumulating (summing) the reciprocals of the variation amounts in each direction in the three-dimensional relative coordinate system for the map point cloud Cm that is matched with the scanned point cloud Cs, over all voxels Vm, including the map point cloud Cm that is matched with the scanned point cloud Cs, as shown with dot hunting in Figure 6. On the other hand, in numbers 12, 13, 14, and 15, the noise quantity-dependent likelihood Ln is input by inputting the reciprocal of the total number of voxels Vm, including the scanned point cloud Cs that are unmatched with the map point cloud Cm, as shown with dot hunting in Figure 7.

[0051] Based on the above, the map matching block 120 estimates the three-dimensional position coordinates Xm2, Ym2, and Zm2 that give the second sub-estimated position Pm2, along with the yaw angle θm2, with respect to the current estimated time t. Accordingly, the map matching block 120 estimates the three-dimensional position errors ΔXm2, ΔYm2, and ΔZm2 that give the second sub-estimated error ΔPm2, along with the yaw angle error Δθm2, with respect to the current estimated time t. The map matching block 120 then stores these estimation results in memory 10 in association with the current estimated time t.

[0052] The fusion block 130 shown in Figure 4 estimates its own position Pf by employing a fusion method that depends on the acquisition timing of either the marker observation data Do or the external data De, which is a three-dimensional point cloud image Des, at the estimated time t. At the estimated time t, which is the acquisition timing of the marker observation data Do, the fusion block 130 fuses the main estimated position Pd and the first sub-estimated position Pm1 by filtering them through a Kalman filter. As a result, position estimation data Df is generated such that it includes the self-position Pf, which is fused by weighting the main estimated position Pd and the first sub-estimated position Pm1 based on the ratio of their respective estimation errors ΔPd and ΔPm1.

[0053] In this case, the three-dimensional position coordinates Xf, Yf, and Zf of the self-position Pf are defined by equations 16, 17, and 18, respectively, using the three-dimensional position coordinates Xd, Yd, and Zd of the main estimated position Pd and the three-dimensional position coordinates Xm1, Ym1, and Zm1 of the first sub-estimated position Pm1. Here, the Kalman gains Kx, Ky, and Kz in equations 16, 17, and 18 are variably set based on the corresponding ratios of the three-dimensional position errors ΔXd, ΔYd, and ΔZd of the main estimated position Pd and the three-dimensional position errors ΔXm1, ΔYm1, and ΔZm1 of the first sub-estimated position Pm1. In particular, equations 16, 17, and 18 are set so that the Kalman gains Kx, Ky, and Kz increase as the relative ratio of the three-dimensional position errors ΔXd, ΔYd, and ΔZd to the three-dimensional position errors ΔXm1, ΔYm1, and ΔZm1 becomes smaller.

number

number

number

[0054] Along with the estimation of the self-position Pf at the estimated time t, which is the timing for acquiring the marker observation data Do, the three-dimensional position errors ΔXf, ΔYf, and ΔZf, which are the estimation error ΔPf, are also estimated and added to the position estimation data Df, as shown in Figure 4. Thus, each of the three-dimensional position errors ΔXf, ΔYf, and ΔZf is defined by equations 19, 20, and 21, using the three-dimensional position errors ΔXd, ΔYd, and ΔZd of the main estimation error ΔPd and the three-dimensional position errors ΔXm1, ΔYm1, and ΔZm1 of the first sub-estimation error ΔPm1. Here, the Kalman gains Kx, Ky, and Kz in equations 19, 20, and 21 are common with equations 16, 17, and 18, respectively.

number

number

number

[0055] Furthermore, at the estimated time t, which is the timing for acquiring marker observation data Do, the yaw angle θf and its error Δθf, shown in Figure 4, among the attitude angles of the autonomous driving device 4, are also estimated by fusion similar to that of the self-position Pf. At this time, the yaw angle θf is defined by equation 22, using yaw angles θd and θm1. Along with this, the yaw angle error Δθf is defined by equation 23, using yaw angle errors Δθd and Δθm1. Here, the Kalman gain Kθ, which is common to equations 22 and 23, is variably set based on the ratio of yaw angle errors Δθd and Δθm1. In particular, in equations 22 and 23, the Kalman gain Kθ is set to increase as the relative ratio of yaw angle error Δθd to yaw angle error Δθm1 becomes smaller.

number

number

[0056] Meanwhile, at the estimated time t, which is the timing for acquiring the three-dimensional point cloud image Des, the fusion block 130 fuses the main estimated position Pd and the second sub-estimated position Pm2 by filtering them through a Kalman filter. As a result, position estimation data Df is generated such that the main estimated position Pd and the second sub-estimated position Pm2 are fused together, weighted according to the ratio of their respective estimation errors ΔPd and ΔPm2, and this fusion includes the self-position Pf.

[0057] In this case, the three-dimensional position coordinates Xf, Yf, and Zf of the self-position Pf are defined by equations 24, 25, and 26, respectively, using the three-dimensional position coordinates Xd, Yd, and Zd of the main estimated position Pd and the three-dimensional position coordinates Xm2, Ym2, and Zm2 of the second sub-estimated position Pm2. Here, the Kalman gains kx, ky, and kz in equations 24, 25, and 26 are variably set based on the corresponding ratios of the three-dimensional position errors ΔXd, ΔYd, and ΔZd of the main estimated position Pd and the three-dimensional position errors ΔXm2, ΔYm2, and ΔZm2 of the second sub-estimated position Pm2. In particular, in equations 24, 25, and 26, the Kalman gains kx, ky, and kz are set to increase as the relative ratio of the three-dimensional position errors ΔXd, ΔYd, and ΔZd to the three-dimensional position errors ΔXm2, ΔYm2, and ΔZm2 becomes smaller.

number

number

number

[0058] Along with the estimation of the self-position Pf at the estimated time t, which is the timing for acquiring the three-dimensional point cloud image Des, the three-dimensional position errors ΔXf, ΔYf, and ΔZf, which are the estimation error ΔPf, are also estimated and added to the position estimation data Df, as shown in Figure 4. Thus, each three-dimensional position error ΔXf, ΔYf, and ΔZf is defined by equations 27, 28, and 29, using the three-dimensional position errors ΔXd, ΔYd, and ΔZd of the main estimation error ΔPd and the three-dimensional position errors ΔXm2, ΔYm2, and ΔZm2 of the second sub-estimation error ΔPm2. Here, the Kalman gains kx, ky, and kz in equations 27, 28, and 29 are common with equations 24, 25, and 26, respectively.

number

number

number

[0059] Furthermore, at the estimated time t, which is the timing for acquiring the three-dimensional point cloud image Des, the yaw angle θf and its error Δθf, shown in Figure 4, among the attitude angles of the autonomous driving device 4, are estimated by fusion similar to that of the self-position Pf. At this time, the yaw angle θf is defined by equation 30, using the yaw angles θd and θm2. Along with this, the yaw angle error Δθf is defined by equation 31, using the yaw angle errors Δθd and Δθm2. Here, the Kalman gain kθ, which is common to equations 30 and 31, is variably set based on the ratio of the yaw angle errors Δθd and Δθm2. In particular, in equations 30 and 31, the Kalman gain kθ is set to increase as the relative ratio of the yaw angle error Δθd to the yaw angle error Δθm2 becomes smaller.

number

number

[0060] Regardless of the acquisition timing described above, the position estimation data Df, which estimates the self-position Pf at the estimated time t by fusion, is output from the fusion block 130 as shown in Figure 4. As a result, the three-dimensional position coordinates Xf, Yf, Zf output as the self-position Pf are stored in memory 10 in association with the estimated time t, and then fed back to the dead reckoning process at the next processing time by the dead reckoning block 100. Furthermore, the three-dimensional position coordinates Xf, Yf, Zf output as the self-position Pf may be used, for example, for advanced driving assistance or automatic driving control of the autonomous driving device 4.

[0061] Based on the combined efforts of blocks 100, 110, 120, and 130 described above, the following describes the estimation data generation method, which generates position estimation data Df obtained by estimating the self-position Pf of the autonomous driving device 4, as shown in Figure 8 (hereinafter referred to as the estimation flow). This estimation flow is executed repeatedly at each estimation interval while the autonomous driving device 4 is running. In this estimation flow, "S" refers to multiple steps executed by multiple instructions included in the estimation program.

[0062] In S10, the dead reckoning block 100 determines whether the current estimated time t is the timing for acquiring marker observation data Do. If the result is positive, the estimation flow proceeds to S20. In S20, the dead reckoning block 100 acquires the main estimated position Pd of the autonomous driving device 4 at the current estimated time t through dead reckoning based on the internal data Di and the dynamics model.

[0063] In S30, following S20, the marker matching block 110 obtains the first sub-estimated position Pm1 of the autonomous driving device 4 at the estimated time t by marker matching based on the marker observation data Do and the marker position data Dmp. At this time, the first sub-estimated position Pm1 is the main estimated position Pd corrected by the deviation δP between the marker prediction position Pp predicted from the main estimated position Pd starting from the marker observation data Do and the marker installation position Pmp that matches the prediction position Pp in the marker position data Dmp.

[0064] In S40, following S30, the fusion block 130 generates position estimation data Df for its own position Pf by fusing the main estimated position Pd and the first sub-estimated position Pm1 at the current estimated time t, weighted based on the ratio of their respective estimation errors ΔPd and ΔPm1. The generated position estimation data Df is output to the memory 10 and stored, and is then fed back to the dead reckoning block 100 for dead reckoning at the next estimated time. Upon completion of the execution of S40, the current execution of the estimation flow ends.

[0065] On the other hand, if a negative determination is made in S10, the estimation flow proceeds to S50. In S50, the map matching block 120 determines whether the current estimated time t is the timing for acquiring the three-dimensional point cloud image Des from the external data De. If a negative determination is made as a result, the current execution of the estimation flow ends. If a positive determination is made, the estimation flow proceeds to S60. In S60, the map matching block 120 obtains the second sub-estimated position Pm2 of the autonomous driving device 4 at the current estimated time t by map matching based on the external data De and the map data Dm.

[0066] In S70, following S60, the fusion block 130 generates position estimation data Df for its own position Pf by fusing the main estimated position Pd and the second sub-estimated position Pm2 at the current estimated time t, weighted based on the ratio of their respective estimation errors ΔPd and ΔPm2. The generated position estimation data Df is also output to the memory 10 and stored, and is fed back to the dead reckoning block 100 for dead reckoning at the next estimated time. Upon completion of the execution of S70, the current execution of the estimation flow ends.

[0067] (Effects and Benefits) The effects and advantages of the first embodiment described above will be explained below.

[0068] According to the first embodiment, a main estimated position Pd is obtained by dead reckoning based on internal data Di, and a first sub-estimated position Pm1 is obtained by marker matching based on marker observation data Do and marker position data Dmp. Here, in obtaining the first sub-estimated position Pm1, the deviation δP between the marker prediction position Pp predicted from the marker observation data Do starting from the main estimated position Pd and the marker installation position Pmp that matches the marker prediction position Pp in the marker position data Dmp is considered. As a result, the estimation error ΔPm1 of the first sub-estimated position Pm1, which is obtained by correcting the main estimated position Pd with the deviation δP between the marker prediction position Pp and the marker installation position Pmp, can be used to offset the influence of the estimation error ΔPd of the main estimated position Pd. Therefore, the main estimated position Pd and the first sub-estimated position Pm1 are weighted and fused based on the ratio of their respective estimation errors ΔPd and ΔPm1, which can reflect changes in external conditions. This makes it possible to generate position estimation data Df that is robust to the estimation accuracy of the self-position Pf.

[0069] According to the first embodiment, at the time of acquiring marker observation data Do, position estimation data Df is generated by fusion of the main estimated position Pd obtained by dead reckoning and the first sub-estimated position Pm1 obtained by marker matching based on the marker observation data Do, weighted according to the ratio of their respective estimation errors. On the other hand, at the time of acquiring the three-dimensional point cloud image Des as external data De, position estimation data Df is generated by fusion of the main estimated position Pd obtained by dead reckoning and the second sub-estimated position Pm2 obtained by map matching based on the acquired image Des, weighted according to the ratio of their respective estimation errors. Thus, at the time of acquiring the three-dimensional point cloud image Des, the first sub-estimated position Pm1, which has been corrected for estimation errors, is acquired by dead reckoning based on the past self-position Pf estimated from the second sub-estimated position Pm2 using the acquired image Des, and at the time of acquiring marker observation data Do, the latest self-position Pf can be estimated with high accuracy. Therefore, it becomes possible to improve the reliability of generating position estimation data Df that is robust to the estimation accuracy of the self-position Pf.

[0070] (Second embodiment) The second embodiment is a modification of the first embodiment.

[0071] As shown in Figure 9, in the second embodiment, pairs of route markers 6 are placed symmetrically on the lateral aspect of the assumed travel trajectory Tr of the autonomous driving device 4 in a three-dimensional relative coordinate system, and these pairs are installed at intervals along the trajectory Tr. In the second embodiment, the autonomous driving device 4 controls the travel direction Td on the trajectory Tr in a direction orthogonal to the virtual straight line L connecting the pair of route markers 6 (hereinafter referred to as pair markers 6 in the description of the second embodiment) as shown in Figure 9. At this time, the actual yaw angle of the autonomous driving device 4 is feedback controlled based on the yaw angle θf estimated by the estimation device 1 (see the first embodiment), thereby enabling the travel direction Td to be aligned with the direction orthogonal to the virtual straight line L.

[0072] Therefore, in the marker matching block 110 of the second embodiment, the three-dimensional position coordinates Xm1, Ym1, and Zm1 of the first sub-estimated position Pm1 at the estimated time t shown in Figure 10 are defined by numbers 32, 33, and 34, respectively, where the suffixes of the pair marker 6 are i=1 and i=2. That is, each three-dimensional position coordinate Xm1, Ym1, and Zm1 is obtained by subtracting the average deviation δP between the three-dimensional position coordinates Xp_i, Yp_i, and Zp_i of the marker prediction position Pp_i of each pair marker 6 and the three-dimensional position coordinates Xmp_i, Ymp_i, and Zmp_i of the marker placement position Pmp_i of each pair marker 6 from the three-dimensional position coordinates Xd, Yd, and Zd of the main estimated position Pd, as shown in Figure 10. At the same time, the yaw angle θm1 at the estimated time t is obtained as an attitude angle, representing the direction of travel Td of the autonomous driving device 4 on the trajectory Tr, i.e., the direction perpendicular to the virtual straight line L.

number

number

number

[0073] In this second embodiment as well, the three-dimensional position errors ΔXm1, ΔYm1, and ΔZm1 (see the first embodiment), which constitute the first sub-estimated error ΔPm1 at the estimated time t, are substantially offset by the influence of the main estimation error ΔPd related to the three-dimensional position coordinates Xd, Yd, and Zd, respectively, due to the relationship between numbers 32, 33, and 34. Therefore, each of the three-dimensional position errors ΔXm1, ΔYm1, and ΔZm1 retains the observation error related to the three-dimensional relative coordinates xo, yo, zo of the marker prediction position Pp, and the map error related to the three-dimensional position coordinates Xmp, Ymp, and Zmp of the marker installation position Pmp, respectively, just as in the first embodiment. Thus, each of the three-dimensional position errors ΔXm1, ΔYm1, and ΔZm1 is predetermined to be different or the same fixed value, based on the observation error predicted as the error characteristics of the marker sensor 20 for the pair marker 6 and the measurement error of the marker position data Dmp for the pair marker 6. At this time, the error Δθm1 predicted for the estimation of the yaw angle θm1 at the estimated time t (see the first embodiment) is also set to a similar fixed value in advance.

[0074] The second embodiment described above also achieves the same effects as the first embodiment. Moreover, according to the second embodiment, the driving direction Td of the autonomous driving device 4 is controlled to be orthogonal to the virtual straight line L connecting the pair markers 6, thereby suppressing the influence of the yaw angle estimation accuracy of the autonomous driving device 4 on the estimation error ΔPm1 of the first sub-estimated position Pm1. Therefore, it becomes possible to improve the reliability of generating position estimation data Df that is robust to the estimation accuracy of the self-position Pf.

[0075] (Other embodiments) Although several embodiments have been described above, this disclosure is not intended to be limited to those embodiments, and can be applied to various embodiments and combinations without departing from the spirit of this disclosure.

[0076] In the modified example, the dedicated computer constituting the estimation device 1 may have at least one of the digital circuit and the analog circuit as a processor. Here, the digital circuit is at least one of the following, for example, ASIC (Application Specific Integrated Circuit), FPGA (Field Programmable Gate Array), SOC (System on a Chip), PGA (Programmable Gate Array), and CPLD (Complex Programmable Logic Device). Furthermore, such a digital circuit may have a memory that stores a program.

[0077] In a modified example, the map matching block 120 may be omitted, and a fusion block 130 may be provided in which fusion processing is not performed at the timing of acquiring the three-dimensional point cloud image Des from the external data De. In this case, steps S50 to S70 may be omitted in the estimation flow, and the current execution of the estimation flow may be terminated in accordance with the negative determination in S10.

[0078] In the modified marker matching block 110, the second sub-estimated position Pm2 may be input to the acquisition timing of the three-dimensional point cloud image Des as external data De, instead of the main estimated position Pd, and used to predict the marker predicted position Pp. In the modified version, instead of the three-dimensional absolute coordinate system and the three-dimensional relative coordinate system, a two-dimensional absolute coordinate system and a two-dimensional relative coordinate system may be defined, in which the vertical direction (Z direction) and height direction (z direction) are not assumed, respectively. In this case, "three-dimensional" should be read as "two-dimensional" when each embodiment is implemented.

[0079] In addition to the embodiments described so far, the above-described embodiments and modifications may be implemented in the form of a processing circuit (e.g., a processing ECU, etc.) or a semiconductor device (e.g., a semiconductor chip, etc.) as a control device configured to be mounted on an autonomous vehicle 4 and having at least one processor 12 and one memory 10. [Explanation of Symbols]

[0080] 1: Estimation device, 4: Autonomous driving device, 6: Route marker, 10: Memory, 12: Processor, Cm: Map point cloud, Cs: Scanning point cloud, De: External data, Df: Position estimation data, Di: Internal data, Dm: Map data, Dmp: Marker position data, Do: Marker observation data, L: Virtual line, Pd: Main estimated position, Pm1: First sub-estimated position, Pm2: Second sub-estimated position, Pmp: Marker placement position, Pp: Marker prediction position, Td: Driving direction, Vm: Voxel, δP: Deviation

Claims

1. An estimation device (1) having a processor (12) and estimating the self-position of an autonomous driving device (4) that autonomously drives in the external environment, The aforementioned processor, The main estimated position (Pd) as the self-position is obtained by dead reckoning based on internal data (Di) acquired within the internal environment of the autonomous driving device, The sub-estimated position (Pm1) as the self-position is obtained by marker matching based on marker observation data (Do) acquired by the autonomous driving device observing multiple route markers (6) installed in the external environment, and marker position data (Dmp) that defines the marker installation position (Pmp), which is the set position of each route marker in the external environment. The system is configured to generate position estimation data (Df) for the self-position by fusion, which weights the main estimated position and the sub-estimated position based on the ratio of their respective estimation errors. The acquisition of the aforementioned sub-estimated position is as follows: An estimation device that includes obtaining the main estimated position as the sub-estimated position by correcting the main estimated position using the deviation (δP) between the marker prediction position (Pp) predicted from the marker observation data starting from the main estimated position, and the marker installation position that matches the marker prediction position in the marker position data.

2. The aforementioned processor, To obtain the aforementioned main estimated position, Obtaining the first sub-estimated position, which is the aforementioned sub-estimated position, The second sub-estimated position (Pm2) is obtained as the self-position by map matching based on external data (De) acquired by scanning the external environment from the autonomous driving device and map data (Dm) that maps the external environment. At the timing when the marker observation data is acquired, the position estimation data is generated by fusion, which weights the main estimated position and the first sub-estimated position based on the ratio of their respective estimation errors, and this data is fed back to the dead reckoning. The estimation device according to claim 1, configured to generate position estimation data by fusion, weighting the main estimated position and the second sub-estimated position based on the ratio of their respective estimation errors, at the timing when the external data is acquired, and to feed this data back to the dead reckoning.

3. The aforementioned external data is, The system includes a scan point cloud (Cs) obtained by scanning the external environment, The data in the aforementioned map is The estimation device according to claim 2, comprising a map point cloud (Cm) obtained by mapping the external environment to a plurality of voxels (Vm) obtained by dividing the external environment into three dimensions.

4. The autonomous driving device, The estimation device according to claim 1 or 2, which controls the driving direction (Td) of the autonomous driving device in a direction orthogonal to a virtual straight line (L) connecting the pair of route markers.

5. The generation of the aforementioned position estimation data is, The estimation device according to claim 1 or 2, further comprising outputting the generated position estimation data.

6. An estimation program, which includes instructions to be executed by a processor (12) and is stored in a storage medium (10) for estimating the self-position of an autonomous driving device (4) that autonomously drives in the external environment, The main estimated position (Pd) as the self-position is obtained by dead reckoning based on internal data (Di) acquired within the internal environment of the autonomous driving device, The sub-estimated position (Pm1) as the self-position is obtained by marker matching based on marker observation data (Do) acquired by the autonomous driving device observing multiple route markers (6) installed in the external environment, and marker position data (Dmp) that defines the marker installation position (Pmp), which is the set position of each route marker in the external environment. The instruction includes the following: generating position estimation data (Df) of the self-position by fusion, which weights the main estimated position and the sub-estimated position based on the ratio of their respective estimation errors; The acquisition of the aforementioned sub-estimated position is as follows: An estimation program that includes obtaining the main estimated position as the sub-estimated position by correcting it using the deviation (δP) between the marker prediction position (Pp) predicted from the marker observation data starting from the main estimated position and the marker installation position that matches the marker prediction position in the marker position data.

7. A method for generating estimated data, which is performed by a processor (12) in order to generate position estimation data (Df) obtained by estimating the self-position of an autonomous driving device (4) that autonomously drives in the external environment, The main estimated position (Pd) as the self-position is obtained by dead reckoning based on internal data (Di) acquired within the internal environment of the autonomous driving device, The sub-estimated position (Pm1) as the self-position is obtained by marker matching based on marker observation data (Do) acquired by the autonomous driving device observing multiple route markers (6) installed in the external environment, and marker position data (Dmp) that defines the marker installation position (Pmp), which is the set position of each route marker in the external environment. This includes generating the position estimation data by fusion, which weights the main estimated position and the sub-estimated position based on the ratio of their respective estimation errors. The acquisition of the aforementioned sub-estimated position is as follows: An estimation data generation method that includes obtaining the main estimated position as the sub-estimated position by correcting the main estimated position using the deviation (δP) between the marker prediction position (Pp) predicted from the marker observation data starting from the main estimated position, and the marker installation position that matches the marker prediction position in the marker position data.

Images