Filter manufacturing method, filter manufacturing device, program, and filter device

Abstract

A filter manufacturing method includes: acquiring a plurality of feature quantities of a plurality of pedestrians serving as samples; a reference path acquisition unit that acquires a reference path, which is a path of the moving body generated by a function of the control device on the basis of the presence of the plurality of pedestrians; an individual path acquisition unit configured to acquire an individual path of the moving body, the individual path being a path of the moving body generated by a function of the control device on the basis of the presence of a pedestrian group obtained by removing one pedestrian from the plurality of pedestrians one by one; extracting an individual path having a smaller difference from the reference path than other individual paths; extracting pedestrians corresponding to the extracted individual paths; and generating a learned model by machine learning in such a manner that, when the extracted feature amount of the pedestrian is input into the learned model, information indicating that the pedestrian is to be excluded is output.

Description

Technical Field

[0001] This invention relates to a method for manufacturing filters, a filter manufacturing apparatus, a procedure, and a filter device. Background Technology

[0002] In recent years, for purposes such as transporting users' luggage, there has been a push towards the practical application of mobile bodies (referred to as robots, micro-mobile bodies, etc.) that move autonomously while following users. An invention related to a driving control device for micro-mobile bodies has been disclosed (Patent Document 1). Furthermore, research is also underway on mobile bodies that not only follow users but also guide them autonomously.

[0003] Prior art literature

[0004] Patent documents

[0005] Patent Document 1: Japanese Patent Publication No. 2020-529050 Summary of the Invention

[0006] The problem that the invention aims to solve

[0007] The aforementioned mobile objects sometimes move in areas where pedestrians are present (floors, sidewalks, open spaces, etc.). Unlike vehicles moving in lanes, when a mobile object moves among many pedestrians, a large number of pedestrians must be monitored to avoid contact, and the processing load on the control device may sometimes become too heavy. Therefore, it is preferable to limit the number of pedestrians to be monitored, but in previous technologies, the limitation of the number of pedestrians to be monitored has not been adequately studied.

[0008] The present invention was made in consideration of such circumstances, and one of its objectives is to provide a filter manufacturing method, filter manufacturing apparatus, program, and filter device that can support the control device in a manner that appropriately limits the number of pedestrians being monitored.

[0009] Solution for solving the problem

[0010] The filter manufacturing method, filter manufacturing apparatus, procedure, and filter device involved in this invention adopt the following structure.

[0011] (1): One aspect of the present invention relates to a filter manufacturing method and a filter manufacturing device manufacturing a filter manufacturing method, wherein the filter is a filter for a control device to limit the number of pedestrians monitored, the control device controls a moving body that moves autonomously in the area where pedestrians walk, the control device identifies objects including the pedestrians, and generates a path that the moving body should travel in the future based on the identification result, the filter manufacturing method comprising the following processing: obtaining multiple feature values ​​of multiple pedestrians as samples; obtaining a reference path, which is a path of the moving body generated by the function of the control device based on the presence of the multiple pedestrians; obtaining a separate path, which is a path of the moving body generated by the function of the control device based on the presence of the pedestrian group after removing one person from each of the multiple pedestrians; extracting a separate path whose difference from the reference path is smaller than other separate paths; extracting the pedestrians corresponding to the extracted separate paths; and generating the learned model by machine learning in a manner that outputs information of the pedestrians that are excluded objects when the feature values ​​of the extracted pedestrians are input into the learned model.

[0012] (2): In the above scheme (1), the plurality of features include the average of the minimum distances between a specified number of other pedestrians and the pedestrian in a specified observation period, counted in order of proximity to the pedestrian in the distance from ....

[0013] (3): In the above scheme (1), the plurality of feature quantities include the speed of the pedestrian.

[0014] (4): In the above (1) scheme, the control device generates a path for the moving body to move while guiding the guide, and the plurality of feature quantities include the distance from the straight line connecting the moving body and the guide to the pedestrian.

[0015] (5): In the above scheme (1), the plurality of features include the length of time the pedestrian is in the monitored area.

[0016] (6): In the above (1) scheme, the plurality of feature quantities include the orientation of the pedestrian when viewed from the moving body.

[0017] (7): In the above (1) scheme, the plurality of feature quantities include the orientation of the pedestrian in a coordinate system based on the moving body.

[0018] (8): Another aspect of the present invention relates to a filter manufacturing apparatus, which manufactures a filter, wherein,

[0019] The filter is used by a control device to narrow down the monitored pedestrians. The control device controls a moving body that moves autonomously in the area where pedestrians walk. The control device identifies objects including the pedestrians and generates a path that the moving body should travel in the future based on the identification results. The filter generation apparatus includes: a feature acquisition unit that acquires multiple feature values ​​of multiple pedestrians as samples; a reference path acquisition unit that acquires a reference path, which is the path of the moving body generated by the function of the control device based on the presence of the multiple pedestrians; a single path acquisition unit that acquires a single path, which is the path of the moving body generated by the function of the control device based on the presence of a group of pedestrians after removing one person from each of the multiple pedestrians; an extraction unit that extracts a single path whose difference from the reference path is smaller than other single paths, and extracts the pedestrians corresponding to the extracted single paths; and a learning unit that generates the learned model by machine learning in a way that outputs information indicating pedestrians that are excluded when the feature values ​​of the extracted pedestrians are input into the learned model.

[0020] (9): Another aspect of the present invention relates to a program executed by a filter-making apparatus for making a filter, wherein the filter is a filter for a control device to limit the number of pedestrians monitored, the control device controlling a moving body that moves autonomously in the area where pedestrians walk, the control device identifying objects including the pedestrians and generating a path that the moving body should travel in the future based on the identification results, the program being used to cause the filter-making apparatus to perform the following processing: obtaining multiple feature values ​​of multiple pedestrians as samples; obtaining a baseline path, the baseline path being the path of the moving body generated by the function of the control device based on the presence of the multiple pedestrians; obtaining a separate path, the separate path being the path of the moving body generated by the function of the control device based on the presence of a group of pedestrians after removing one person from the multiple pedestrians one by one; extracting a separate path whose difference from the baseline path is smaller than other separate paths; extracting the pedestrians corresponding to the extracted separate paths; and generating the learned model by machine learning in such a way that when the feature values ​​of the extracted pedestrians are input into the learned model, the model outputs information indicating that the pedestrians are excluded objects.

[0021] (10): Another aspect of the present invention relates to a filter device suitable for a control device that controls a moving body that moves autonomously in a pedestrian walking area, the control device identifying objects including the pedestrians and generating a path that the moving body should travel in the future based on the identification results, wherein the filter device uses a filter to limit the pedestrians that the control device is monitoring, the filter being generated by a filter making device performing the following process: obtaining multiple feature values ​​of multiple pedestrians as samples; obtaining a baseline path, the baseline path being the path of the moving body generated by the function of the control device based on the presence of the multiple pedestrians; obtaining a separate path, the separate path being the path of the moving body generated by the function of the control device based on the presence of a group of pedestrians after removing one person from the multiple pedestrians one by one; extracting a separate path whose difference from the baseline path is smaller than other separate paths; extracting the pedestrians corresponding to the extracted separate paths; and generating the learned model by machine learning in such a way that when the feature values ​​of the extracted pedestrians are input into the learned model, the model outputs information indicating that the pedestrians are excluded.

[0022] Invention Effects

[0023] According to the schemes (1) to (10), the control device can be supported in a way that appropriately limits the number of pedestrians to be monitored. Attached Figure Description

[0024] Figure 1 This is a diagram illustrating an example of the relationship between a moving body and a filter manufacturing device.

[0025] Figure 2 This is a structural diagram of the moving body.

[0026] Figure 3 This is a structural diagram of the control device.

[0027] Figure 4 This is a diagram representing an example of a path candidate.

[0028] Figure 5 This is a flowchart illustrating an example of the processing content of the prediction unit related to pedestrians.

[0029] Figure 6 This is a diagram illustrating an example of how to generate an ideal path.

[0030] Figure 7 This is a flowchart illustrating an example of the processing content of the prediction department related to the user.

[0031] Figure 8 This is a diagram illustrating an example of the characteristics of the third score.

[0032] Figure 9 This is a diagram used to illustrate the calculation process for the fourth score.

[0033] Figure 10 This is a diagram illustrating a scenario where a mobile device inhibits a user's approach to a pedestrian.

[0034] Figure 11 This is a diagram illustrating a scenario where a mobile device inhibits a user's approach to a pedestrian.

[0035] Figure 12 This is a structural diagram of the filter manufacturing device.

[0036] Figure 13 It is a diagram that schematically represents each characteristic quantity.

[0037] Figure 14 It is a graph used to illustrate the difference between a baseline path and individual paths.

[0038] Figure 15 This is a flowchart illustrating an example of the processing flow of a filter manufacturing apparatus.

[0039] Figure 16 It is a diagram that schematically represents the content of the model after it has been learned. Detailed Implementation

[0040] [summary]

[0041] Hereinafter, with reference to the accompanying drawings, embodiments of the filter manufacturing method, filter manufacturing apparatus, procedure, and filter apparatus of the present invention will be described. The filter according to the present invention is applicable to a control device for a moving body that controls a drive device for moving the moving body. The moving body is, for example, a moving body that autonomously moves while guiding a person in a pedestrian walking area. Alternatively, the moving body may move independently or follow a person at a certain distance. In the following description, the moving body is the moving body that guides a person. The pedestrian walking area refers to sidewalks, open spaces, building floors, etc., and may also include driveways. In the following description, no person is riding on the moving body, but a person may also be riding on the moving body. The person guiding the person is, for example, a pedestrian, but may also be a robot or an animal (hereinafter referred to as user U). The moving body moves slightly ahead of user U (who is an elderly person) while heading to a pre-given destination, thereby preventing other pedestrians who might obstruct user U's movement from getting too close to user U (i.e., moving in a way that clears a path for user U). User U is not limited to the elderly; it can also include people with difficulty walking, children, shoppers in supermarkets, patients moving within hospitals, and pets on walks. It should be noted that such actions may not be continuous but rather temporary. For example, when a mobile device is traveling alongside or chasing a user, and a specified situation (such as the presence of obstacles or traffic congestion) is detected in the user's direction of travel, the mobile device executes a prescribed algorithm to temporarily guide the user.

[0042] [Overall Structure]

[0043] Figure 1 This diagram illustrates an example of the relationship between the mobile body 1 and the filter manufacturing apparatus 200. The filter manufacturing apparatus 200, for example, includes a simulator 270 that simulates the functions of the control device 100 (described later) of the mobile body 1. Alternatively, the filter apparatus 200 may use the actual machine of the mobile body 1 instead of the simulator 270. The filter manufacturing apparatus 200 acquires sample data 310, which consists of data from multiple pedestrians, and manufactures a filter for use on the mobile body 1 based on this data. The manufactured filter (filtering program) is mounted on the mobile body 1. Hereinafter, the mobile body 1 will be described first, followed by the filter manufacturing apparatus 200.

[0044] [Structure of the moving object]

[0045] Figure 2 This is a structural diagram of the mobile body 1. The mobile body 1 is equipped with, for example, an HMI (Human Machine Interface) 10, an object detection device 20, a drive unit 30, a sensor 40, and a control unit 100. These structures are supported or housed by the base 5.

[0046] The HMI10 provides various information to the user U and accepts input operations performed by the user U. The HMI10 includes various display devices, speakers, buzzers, touch panels, switches, buttons, near-field communication devices, etc. For example, the HMI10 accepts destination settings.

[0047] The object detection device 20 is a device that generates data for identifying other pedestrians and users U present in the vicinity of the moving body 1. The object detection device 20 may include, for example, a camera that captures images of the vicinity of the moving body 1. The object detection device 20 may also include sensor types such as radar devices, LIDAR (Light Detection and Ranging), and ultrasonic sensors, as well as object recognition devices that perform sensor fusion processing based on the outputs of these sensor types to determine objects.

[0048] The drive unit 30 is a mechanism for moving the moving body 1, including the base 5, in any direction. The drive unit 30 may include, for example, multiple wheels, a drive motor mounted on one or more wheels, and a steering device mounted on one or more wheels. There are no particular restrictions on the structure of the drive unit 30, and it can have any structure. In principle, the drive unit 30 moves the moving body 1 while simultaneously orienting the front surface of the base 5 toward the direction of travel of the moving body 1.

[0049] Sensor 40 is a sensor used to detect the behavior of the moving body 1. Sensor 40 includes, for example, a wheel speed sensor for detecting the speed of the wheels, an acceleration sensor for detecting the acceleration acting on the moving body 1, a yaw rate sensor mounted near the center of gravity in the horizontal direction of the base 5, a steering angle sensor for detecting the steering angle of the steering wheel (steering wheel), and an orientation sensor for detecting the orientation of the moving body 1 in the horizontal direction.

[0050] Figure 3This is a structural diagram of the control device 100. The control device 100 includes, for example, an identification unit 110, a filter processing unit 115, a path generation unit 120, a prediction unit 130, and a drive control unit 140. The path generation unit 120 includes a path candidate setting unit 122 and a score calculation unit 124. These components are implemented, for example, by executing a program (software) using a hardware processor such as a CPU (Central Processing Unit). Some or all of these components can be implemented by hardware (including circuitry) such as LSI (Large Scale Integration), ASIC (Application Specific Integrated Circuit), FPGA (Field-Programmable Gate Array), and GPU (Graphics Processing Unit), or through the coordinated use of software and hardware. The program can be pre-saved in a storage device such as an HDD (Hard Disk Drive) or flash memory (a storage device with a non-transitory storage medium), or it can be saved in a removable storage medium such as a DVD or CD-ROM (a non-transitory storage medium), and installed in the storage device by mounting the storage medium to a drive device.

[0051] The recognition unit 110 identifies objects, including other pedestrians (hereinafter referred to as pedestrian P) and users U, based on the data output by the object detection device 20. When the object detection device 20 is a camera, the recognition unit 110 identifies pedestrians by inputting camera images into a learned model used for pedestrian identification. The same applies to objects other than pedestrians. Furthermore, to distinguish between pedestrian P and user U, the recognition unit 110 may pre-store multiple images of user U captured by the camera as templates in a storage unit (not shown), and compare the templates with the camera images to determine user U. Additionally, the recognition unit 110 may utilize communication directionality to identify the location of user U by conducting short-range wireless communication between the terminal device held by user U and the HMI 10.

[0052] The filter processing unit 115 extracts one or more pedestrians P from a plurality of pedestrians P that are designated as monitoring targets (and as judgment materials for controlling the moving body 1). The filter processing unit 115 operates according to a filter (filtering program) manufactured by the filter manufacturing apparatus 200. The filter processing unit 115 is an example of a "filtering device". In the following description, the pedestrians P before being processed by the filter processing unit 115 are referred to as pedestrian candidates Pc, and the pedestrians P extracted by the filter processing unit 115 are referred to as pedestrians P. First, the filter processing unit 115 derives multiple feature values ​​for each pedestrian candidate Pc based on the recognition result of the recognition unit 110. The content of the feature values ​​is the same as that described in the description of the filter manufacturing apparatus 200 described later. Furthermore, by inputting the derived feature values ​​into the filter, it is determined whether to exclude the pedestrian candidate Pc from the monitoring targets. Pedestrian candidate Pcs that are not excluded by the filter processing unit 115 are processed as subsequent pedestrians P.

[0053] [Route Candidate]

[0054] The path generation unit 120's path candidate setting unit 122 generates multiple path candidates Pmc(i) that will be candidates for the path Rm along which the mobile body 1 should move in the future. i is an identifier for the path candidate, taking a value from 1 to n (n is a natural number greater than 2). The path and path candidates are generated, for example, as a path obtained by connecting multiple (q) path points arranged at a specified distance. Path points represent the locations that the mobile body 1 should reach at specified intervals (every step). Figure 4 This is a diagram illustrating an example of a path candidate Rmc(i). Here, we assume n=5 and q=3. The path candidate setting unit 122 generates path candidates Rmc(1), Rmc(2), Rmc(4), and Rmc(5) by expanding to the left and right, for example, with path candidate Rmc(3) extending in a straight line along the front direction Dm of the moving body 1 as the center. The spacing between path points is constant, for example, and each path candidate has a length of q steps. The path candidates are not generated to reach the destination assigned to the moving body 1, but rather with a length of constant distance × q steps. For example, path candidate Rmc(2) is generated by turning left by a specified angle in each step. In addition, path candidate Rmc(1) is generated by turning left by a larger specified angle in each step. The opposite applies to path candidates Rmc(4) and Rmc(5). This rule is just one example; as long as the same tendency (gradual expansion) is represented, path candidates can be generated based on other rules. The path candidate setting unit 122 outputs the path candidate Rmc(i) to the prediction unit 130.

[0055] exist Figure 4In the example, the distance between path points is constant, but the distance can also be variable. A long distance means that the movement is faster within this interval, and a short distance means that the movement is slower within this interval. That is, the path Rm and the path candidate Rmc(i) can also include speed elements. For example, the path candidate setting unit 122 generates path candidates with different speeds based on path candidates generated by setting the distance between path points to constant.

[0056] [Prediction (Simulation on actual machine)]

[0057] When a path candidate Rmc(i) is given, the prediction unit 130 predicts the future positions of pedestrian P and user U for each path candidate Rmc(i). The prediction unit 130 predicts the future positions of pedestrian P and user U, for example, by the method described in any one of Non-Patent Documents 1 to 3, or by a combination thereof.

[0058] Figure 5 This is a flowchart illustrating an example of the processing performed by the prediction unit 130 related to pedestrian P. First, the prediction unit 130 performs endpoint prediction (S1) for each pedestrian P. For example, the prediction unit 130 observes the movement trajectory of pedestrian P in a time series and predicts the path ahead of the movement vector of pedestrian P in a state where there are no other pedestrians P, moving bodies 1, or users U in the vicinity as the endpoint (arrival location) of pedestrian P.

[0059] Next, the prediction unit 130 performs processing steps S2 to S6 for each path candidate Rmc(i). The prediction unit 130 sequentially inputs a path candidate Rmc(i) of a moving body 1 (S2), and generates the ideal path Ru#(i) for the case where user U follows moving body 1 by establishing a correspondence with the path candidate Rmc(i) (S3). The ideal path Ru#(i) is the path that user U is expected to move on, assuming that pedestrian P does not exist in the vicinity of moving body 1 and user U. Figure 6 This is a diagram illustrating an example of the method for generating the ideal path Ru#(i). Here, the generation of the ideal path Ru#(5) corresponding to the path candidate Rmc(5) is illustrated. For example, the prediction unit 130 generates the ideal path Ru#(5) for user U by moving the user U from step 0 (calculation time point) to step 1 towards the position Om-0 of the moving body 1 in step 0, moving the user U from step 1 to step 2 by moving the user U towards the position Om-1 of the moving body 1 in step 1, and moving the user U from step 2 to step 3 by moving the user U towards the position Om-2 of the moving body 1 in step 2, i.e., by following the moving body 1 one step slower. This generation method is just one example, and other generation methods may also be used.

[0060] return Figure 5 The prediction unit 130 performs the processing of steps S4 and S5 for each pedestrian P (k=1~m) with a step size of q. First, assuming that the moving body 1 and the user U move according to the path candidate Rmc(i) of the moving body 1 and the ideal path Ru#(i) of the user U respectively, the various forces acting on the pedestrian P(k) are calculated in the current step (S4). These forces are not actually acting but are considered as hypothetical forces generated due to the mental actions of the pedestrian P(k).

[0061] Hypothetical forces include, for example, (1) the influence F1 caused by the approach of other objects (in this case, moving body 1 and user U), (2) the turning force F2 which corrects the direction of travel to the destination after the change of the forward path mainly due to influence F1, and (3) the acceleration and deceleration force F3 which adjusts the speed to the desired speed after acceleration and deceleration mainly due to influence F1. F1 is a translational force acting in any direction along the two-dimensional plane, F2 is a rotational force centered on the pedestrian, and F3 is a translational force parallel to the pedestrian's movement vector. F1 to F3 can be simply represented as follows. In addition to these, integral elements, differential elements, etc., can be taken into consideration to calculate the hypothetical forces.

[0062] F1 = (weight) × (coefficient) × (relative velocity) / (distance from other objects)

[0063] F2 = (Moment of inertia) × (Coefficient) × (Angle difference between the nearest movement vector and the direction towards the destination)

[0064] F3 = weight × (coefficient) × (difference between desired speed and nearest speed)

[0065] Prediction unit 130 assumes that the hypothetical force continues to act within a one-step interval and predicts the position and velocity of pedestrian P(k) one step later (S5). This process, performed on each pedestrian for a number of steps (q steps), generates and outputs a predicted path Rp(k,i) for each pedestrian P(k) (S6). This predicted path Rp(k,i) is the result of predicting the future position of each pedestrian P(k) based on each path candidate Rmc(i).

[0066] Next, the forecasting department 130 followed... Figure 7 The flowchart is used to predict the future location of user U. Figure 7 This is a flowchart illustrating an example of the processing content of the prediction unit 130 related to user U.

[0067] The prediction unit 130 performs processing steps S11 to S14 for each path candidate Rmc(i). The prediction unit 130 inputs the predicted paths of each pedestrian corresponding to the path candidate Rmc(i) (S11). Next, the prediction unit 130 performs processing steps S12 and S13 of step q. The prediction unit 130 assumes that the moving body 1 and each pedestrian P(k) move according to the path candidate Rmc(i) and the predicted paths Rp(k,i) of each pedestrian, and calculates the various forces acting on the user U in the current step (S12). The method for calculating these various forces can be compared with... Figure 5 The processing in S4 is the same. The prediction unit 130 assumes that the hypothetical force continues to act within a one-step interval and predicts the position and velocity of user U after one step (S13). This processing of q steps is performed, thereby generating and outputting the predicted path Ru(i) of user U for each path candidate Rmc(i) (S14). The predicted path Ru(i) is the result of predicting the future position of user U for each path candidate Rmc(i).

[0068] [Evaluation of Path Candidates]

[0069] When the scoring calculation unit 124 obtains the predicted path Ru(i), ideal path Ru#(i), and predicted path Rp(k,i) corresponding to i=1~n respectively from the prediction unit 130 as the prediction result, it calculates the comprehensive score SC(i) of each candidate path Rmc(i). The comprehensive score SC(i) is represented by equation (1). α and β are coefficients, SC1(i) ​​is the first score, and SC2(i) is the second score. Regarding the comprehensive score SC(i), the smaller the value, the better the evaluation. The coefficients α, β, and ζ and η (described later) are positive values.

[0070] SC(i)=α·SC1(i)+β·SC2(i)…(1)

[0071] The first score SC1(i) ​​is a score based on the positional relationship between mobile device 1 and the destination location assigned to mobile device 1. For example, the first score SC1(i) ​​is calculated as follows: the longer the distance X1 between the terminal point (the point furthest from mobile device 1) of the path candidate Rmc(i) and the destination location assigned to mobile device 1, the larger the first score SC1(i); the shorter the distance X1, the smaller the first score SC1(i). The first score SC1(i) ​​can be obtained directly using the distance X1, or by calculating the logarithm, exponent, etc., of the distance X1, or it can be the output value of any function with the distance as the input value.

[0072] The second score SC2(i) is based on the positional relationship between user U and pedestrian P from the current point in time to a future point in time (the point in time when they reach the end of the path). Regarding the second score SC2(i), the more the interval between user U and pedestrian P is maintained, the more certain the value. For example, the second score SC2(i) is calculated based on equation (2). SC3(i) is the third score, and SC4(i) is the fourth score.

[0073] SC2(i)=ζ·SC3(i)+η·SC4(i)…(2)

[0074] For example, the third score SC3(i) is calculated as follows: The distance between the path points of the user U's anticipated path Ru(i) and the path points of all pedestrians P's predicted paths Rp(k,i) is calculated for each step. The shorter the distance X2 (i.e., the distance when user U and any pedestrian P are closest during step q) is, the larger the third score SC3(i); the longer the distance X2 is, the smaller the third score SC3(i). However, when the distance X2 is sufficiently long (the interval between user U and pedestrian P is sufficiently maintained), the third score SC3(i) can be a score representing a constant value. For example, when the distance X2 is above a boundary value of approximately 2 [m], the third score SC3(i) can be a score representing zero. Figure 8 This is a graph illustrating an example of the characteristics of the third score SC3(i). To prevent user U from approaching pedestrian P, the third score SC3(i) is preferably calculated in a manner that increases rapidly as the distance X2 decreases. For example, an obstacle function, an exponential function, etc., are preferred, but the calculation is not limited to these, as long as they represent the same tendency.

[0075] The fourth score SC4(i) is calculated in such a way that the greater the deviation between the predicted path Ru(i) and the ideal path Ru#(i), the larger it becomes. For example, the score calculation unit 124 squares the distance between the path points of the predicted path Ru(i) and the ideal path Ru#(i) at each step, and calculates the square of the total value as the fourth score SC4(i). Figure 9 This is a diagram illustrating the calculation process for the fourth score SC4(i). In the diagram, r1 is the distance between the path points of the predicted path Ru(i) and the ideal path Ru#(i) in their respective first steps, r2 is the distance between the path points of the predicted path Ru(i) and the ideal path Ru#(i) in their respective second steps, and r3 is the distance between the path points of the predicted path Ru(i) and the ideal path Ru#(i) in their respective third steps. The fourth score SC4(i) is calculated, for example, by √(r1... 2 +r2 2 +r3 2 And find it.

[0076] After calculating the comprehensive score SC(i) of each path candidate Rmc(i), the score calculation unit 124 selects the path candidate Rmc(i) with the smallest comprehensive score SC(i) and outputs it to the drive control unit 140 as the path Rm of the moving body 1. The drive control unit 140 controls the drive device 30 installed on the moving body 1 to make the moving body 1 move along the path Rm.

[0077] By processing in this way, mobile body 1 is controlled to perform the following processes in parallel: guiding user U to the given destination; and preventing other pedestrians P from getting too close to user U due to its own presence. For example, this is because, in a scenario where pedestrian P is approaching user U, the overall score of the path candidate Rmc(i) where mobile body 1 enters between pedestrian P and user U to restrain pedestrian P's approach should become higher.

[0078] Figure 10 and Figure 11 This is a diagram illustrating a scenario where a moving body 1 suppresses the approach of user U and pedestrian P. Figure 10 and Figure 11 In the diagram, the left figure shows the movement path of pedestrian P when the coefficient β=0 and only the path Rm to the destination is selected; the right figure shows the movement path of pedestrian P when the coefficient β is sufficiently greater than 0. Figure 10 In the example shown in the right figure, the moving body 1 oriented towards the right-bulging path Rm, thereby (under the influence of influence F1) causing pedestrian P to take a left-leaning path, thus inhibiting pedestrian P's approach to user U. Figure 11 In the example shown in the right figure, the moving body 1 moves at a low speed, thereby (under the influence of F1) causing pedestrian P to take a path to the right, thus inhibiting pedestrian P from approaching user U. In this way, the control device 100 can generate the path of the moving body 1 in a manner that allows for appropriate behavior towards user U.

[0079] The coefficients α, β, ζ, and η can be fixed values ​​or can be adjusted arbitrarily.

[0080] In the above description, the control device 100 is mounted on the mobile body 1, but it is not limited to this. It can also be a structure that is set in a place far away from the mobile body 1, obtains the output data of the object detection device 20 through communication, and sends the drive instruction signal to the drive device 30, that is, remotely controls the mobile body 1.

[0081] [Filter Making]

[0082] Figure 12This is a structural diagram of a filter manufacturing apparatus 200. The filter manufacturing apparatus 200 includes, for example, a feature acquisition unit 210, a reference path acquisition unit 220, a single path acquisition unit 230, an extraction unit 240, a learning unit 250, a simulator 270, and a storage unit 300. Components other than the storage unit 300 are implemented by executing programs (software) using hardware processors such as CPUs. Some or all of these components can be implemented by hardware (including circuitry) such as LSIs, ASICs, FPGAs, and GPUs, or through the coordinated use of software and hardware. The program can be pre-saved in storage devices such as HDDs and flash memory (storage devices with non-transitory storage media), or it can be saved in removable storage media such as DVDs and CD-ROMs (non-transitory storage media), and installed in the storage device by mounting the storage media to a drive device.

[0083] Storage unit 300 can be RAM (Random Access Memory), flash memory, HDD, etc. Sample data 310, model setting information 320, and other information are stored in storage unit 300. Sample data 310 can be, for example, time-series data such as the relative position, velocity, and acceleration of pedestrians as observed from moving body 1, and the position of user U as observed from moving body 1. Alternatively, sample data 310 can be time-series data in an absolute coordinate system. Sample data 310 can be data obtained when moving body 1 in a real environment, or it can be hypothetical data created based on probability calculations. In the following description, similar to the description of filter processing unit 115, the pedestrian P included in sample data 310 will be referred to as pedestrian candidate Pc. Model setting information 320 specifies the type, number of layers, and node connection information of the machine learning model that forms the basis of the learned model that functions as the filtering program.

[0084] The feature acquisition unit 210 derives multiple feature quantities based on the sample data. These multiple feature quantities may include, for example, the following. Among the feature quantities, there may also be feature quantities derived through simulation performed by the simulator 270. Figure 13 This diagram schematically represents the following characteristic quantities. For ease of layout on paper, pedestrian candidate Pcs have been appropriately divided according to the characteristic quantities.

[0085] (1) The average of the minimum distances Dmin between the other pedestrian candidates Pc and the pedestrian candidate Pc during the specified observation period, counted in order of proximity from the pedestrian candidate Pc to the farthest point.

[0086] (2) The speed VPc of the pedestrian candidate Pc

[0087] (3) The distance D1 from the straight line connecting user U and moving body 1 to pedestrian candidate Pc.

[0088] (4) The length of time T1 during which the pedestrian candidate Pc is in the monitored area A1.

[0089] (5) The orientation (direction) θ of the pedestrian candidate Pc as observed from the moving body 1

[0090] (6) The orientation (movement vector or body orientation) φ of the pedestrian candidate Pc in the coordinate system of the moving body 1 (e.g., a coordinate system with the central axis of the base body 5 as one coordinate axis, represented by the X and Y axes in the figure).

[0091] (7) The swing amount Viv obtained by evaluating the future trajectory of the pedestrian candidate Pc based on the prescribed benchmark.

[0092] The reference path acquisition unit 220 acquires a reference path, which is the path of the moving body 1 generated by the function of the control device 100 based on the presence of multiple pedestrian candidate Pcs. For example, the reference path acquisition unit 220 acquires the reference path by assigning sample data 310 to the simulator 270. The simulator 270 has the same path generation function as the aforementioned control device 100 (e.g., a virtual machine or physical processor).

[0093] The individual path acquisition unit 230 acquires an individual path for each pedestrian candidate Pc. This individual path refers to the path of the moving body 1 generated by the control device 100 based on the presence of the remaining pedestrian candidate Pcs excluding the selected ones, selected sequentially from the sample data 310. For example, the reference path acquisition unit 220 acquires individual paths by feeding the simulator 270 data after removing pedestrian candidate Pcs from the sample data 310 one by one.

[0094] The extraction unit 240 extracts individual paths whose difference from the reference path is smaller compared to other individual paths. For example, the extraction unit 240 extracts (NM) pedestrian candidate Pcs from the number of N pedestrian candidate Pcs, counting them in ascending order of the difference from the reference path compared to other individual paths, with M remaining pedestrian candidate Pcs. "Difference from the reference path" refers, for example, to the value of RMSD (root mean square deviation) calculated regarding the difference in position at each step when the reference path and individual paths are divided into a finite number of steps. Figure 14 This is a graph used to illustrate the difference between a baseline path and individual paths. In the graph, RP is the baseline path, IP is the individual path, and ΔY is the difference in position over several steps.

[0095] Figure 15This is a flowchart illustrating an example of the processing flow of the filter manufacturing apparatus 200 described above. First, the feature acquisition unit 210 derives multiple feature quantities from the sample data 310 for each of the multiple pedestrian candidate Pcs (S20). Next, the reference path acquisition unit 220 uses the simulator 270 to generate a reference path (S21).

[0096] Next, the individual path acquisition unit 230 and the extraction unit 240 perform the processing steps S22 and S23 for each pedestrian candidate Pc(u) included in the sample data 310. The parameter u is the identifier of the pedestrian candidate Pc, u=1~N. First, the individual path acquisition unit 230 uses the simulator 270 to generate an individual path IP(u) based on the existence of the remaining pedestrian candidate Pcs excluding the selected pedestrian candidate Pc(u) (S22). Next, the extraction unit 240 calculates the difference ΔP(u) between the individual path IP(u) and the reference path RP (S23).

[0097] Furthermore, the extraction unit 240 sequentially extracts pedestrian candidates Pc from the side with the smaller difference ΔP and sets them as excluded pedestrian candidates (S24).

[0098] The learning unit 250 generates a fully learned model that functions as a filter through machine learning. The learning unit 250 sets the features of the pedestrian candidate Pc (hereinafter referred to as pedestrian candidate Pc(p)) extracted by the extraction unit 240 as positive examples of learning data, and sets the features of the unextracted pedestrian candidate Pc (hereinafter referred to as pedestrian candidate Pc(n)) as negative examples of learning data. Then, the learning unit 250 generates a fully learned model through machine learning in a manner that outputs information indicating a pedestrian who is excluded (e.g., the probability of being excluded, or a flag indicating that someone is excluded) when the features of pedestrian candidate Pc(p) are input, and outputs information indicating a pedestrian who is not excluded when the features of pedestrian candidate Pc(n) are input.

[0099] The learned model is, for example, a random forest. Figure 16 This diagram schematically illustrates an example of a learned model. Alternatively, the learned model could be a Deep Neural Network (DNN) or similar form. Furthermore, the learned model is not limited to outputting information indicating whether a pedestrian candidate Pc is an excluded pedestrian when given the feature values ​​of one pedestrian candidate Pc; it can also be a model that outputs information indicating which pedestrian candidate Pc is an excluded pedestrian when given the feature values ​​of multiple pedestrian candidate Pcs. In the latter case, the learned model can also be created by adjusting the criteria (threshold, etc.) for setting each pedestrian candidate Pc as an excluded pedestrian based on the number of pedestrian candidate Pcs given the input feature values.

[0100] [Constriction of characteristic quantities]

[0101] In the above description, the feature quantity is set to be fixed, but the feature quantity can also be a feature quantity that is prepared in advance and narrowed down from a larger number of feature quantities. For example, in addition to the feature quantities described above, the following can also be prepared in advance: (8) the distance between the pedestrian candidate Pc and the moving body 1, (9) the distance between the pedestrian candidate Pc and the user U, (10) the distance between the pedestrian candidate Pc and the destination, (11) the angle between the orientation of the pedestrian candidate Pc and the straight line connecting the user U and the moving body 1, (12) the amount of swing of the orientation of the pedestrian candidate Pc, (13) the distance between the future position of the pedestrian candidate Pc and the moving body 1, etc. In this case, the filter making device 200 can also make a learned model by eliminating feature quantities one by one, just like narrowing down the pedestrian candidate Pc, and sequentially eliminating the feature quantities corresponding to the learned model whose pedestrian candidate Pc does not change significantly between the other learned models.

[0102] According to the implementation described above, the control device can be supported in a way that appropriately limits the number of pedestrians to be monitored.

[0103] The implementation methods described above can be performed as follows.

[0104] A filter manufacturing apparatus for manufacturing filters, wherein,

[0105] The filter is used by the control device to narrow down the monitored pedestrians. The control device controls a moving body that moves autonomously within the area where pedestrians are walking. The control device identifies objects, including the pedestrians, and generates a path that the moving body should travel based on the identification results.

[0106] The filter manufacturing apparatus includes:

[0107] One or more storage media for storing computer-readable instructions; and

[0108] The processor connected to the one or more storage media,

[0109] The processor performs the following processing by executing computer-readable instructions:

[0110] Obtain multiple feature values ​​of multiple pedestrians who become samples;

[0111] A baseline path is obtained, which is the path of the moving body generated by the function of the control device based on the presence of the multiple pedestrians;

[0112] A separate path is obtained, which is the path of the moving body generated by the function of the control device based on the existence of the pedestrian group after removing one person one by one from the plurality of pedestrians;

[0113] Extract individual paths whose difference from the baseline path is smaller compared to other individual paths;

[0114] Extract pedestrians corresponding to the extracted individual paths; and

[0115] The learned model is generated through machine learning by using the extracted pedestrian features as input to the learned model and outputting information about pedestrians who are excluded from the learning process.

[0116] The above describes specific embodiments of the present invention, but the present invention is not limited to such embodiments in any way. Various modifications and substitutions can be made without departing from the spirit of the present invention.

[0117] Explanation of reference numerals in the attached figures

[0118] 1. Moving body

[0119] 20 Object detection devices

[0120] 30 Drive unit

[0121] 100 Control device

[0122] 110 Identification Department

[0123] 115 Filter Processing Department

[0124] 120 Path Generation Department

[0125] 122 Path Candidate Setting Department

[0126] 124 Score Calculation Department

[0127] 130 Forecasting Department

[0128] 140 Drive Control Unit

[0129] 200 Filter Manufacturing Device

[0130] 210 Feature Quantity Acquisition Section

[0131] 220 Baseline Path Acquisition Unit

[0132] 230 Individual Path Acquisition Section

[0133] 240 Extraction Section

[0134] 250 Study Department

[0135] 300 Storage Unit

[0136] 310 Sample Data

[0137] 320 Model settings information.

Claims

1. A method for manufacturing a filter, comprising a filter manufacturing apparatus for manufacturing a filter, wherein, The filter is used by the control device to narrow down the monitored pedestrians. The control device controls a moving body that moves autonomously within the area where pedestrians are walking. The control device identifies objects including the pedestrians and generates a path that the moving body should take in the future based on the identification results. The filter manufacturing method includes the following steps: Obtain multiple feature values ​​of multiple pedestrians who become samples; A baseline path is obtained, which is the path of the moving body generated by the function of the control device based on the presence of the multiple pedestrians; A separate path is obtained, which is the path of the moving body generated by the function of the control device based on the existence of the pedestrian group after removing one person one by one from the plurality of pedestrians; Extract individual paths whose difference from the baseline path is smaller compared to other individual paths; Extract pedestrians corresponding to the extracted individual paths; as well as The learned model is generated by machine learning in a manner that outputs information about pedestrians who are excluded from the learning process when the extracted pedestrian features are input into the learned model.

2. The filter manufacturing method according to claim 1, wherein, The plurality of features include the average of the minimum distances between a specified number of other pedestrians and the pedestrian in the observation period, counted in ascending order of distance from the pedestrian in the observation.

3. The filter manufacturing method according to claim 1, wherein, The plurality of features include the speed of the pedestrian.

4. The filter manufacturing method according to claim 1, wherein, The control device generates a path for the moving body to guide the person being guided. The plurality of features include the distance from the straight line connecting the moving body to the guide object to the pedestrian.

5. The filter manufacturing method according to claim 1, wherein, The plurality of features include the length of time the pedestrian is in the monitored area.

6. The filter manufacturing method according to claim 1, wherein, The plurality of features include the pedestrian's orientation as observed from the moving body.

7. The filter manufacturing method according to claim 1, wherein, The plurality of feature quantities include the orientation of the pedestrian in a coordinate system based on the moving body.

8. A filter manufacturing apparatus for manufacturing filters, wherein, The filter is used by the control device to narrow down the monitored pedestrians. The control device controls a moving body that moves autonomously within the area where pedestrians are walking. The control device identifies objects including the pedestrians and generates a path that the moving body should take in the future based on the identification results. The filter manufacturing apparatus includes: The feature acquisition unit acquires multiple feature values ​​of multiple pedestrians that become samples; A reference path acquisition unit acquires a reference path, which is the path of the moving body generated by the function of the control device based on the presence of the multiple pedestrians. A separate path acquisition unit acquires a separate path, which is the path of the moving body generated by the function of the control device based on the existence of the pedestrian group after removing one person from each of the plurality of pedestrians; The extraction unit extracts individual paths whose difference from the baseline path is smaller than that of other individual paths, and extracts the pedestrians corresponding to the extracted individual paths; as well as The learning unit generates the learned model by means of machine learning, which outputs information about pedestrians that are excluded from the learning model when the extracted features of the pedestrians are input into the learned model.

9. A program executed by a filter-making apparatus for making a filter, wherein, The filter is used by the control device to narrow down the monitored pedestrians. The control device controls a moving body that moves autonomously within the area where pedestrians are walking. The control device identifies objects including the pedestrians and generates a path that the moving body should take in the future based on the identification results. The program is used to cause the filter manufacturing apparatus to perform the following processing: Obtain multiple feature values ​​of multiple pedestrians who become samples; A baseline path is obtained, which is the path of the moving body generated by the function of the control device based on the presence of the multiple pedestrians; A separate path is obtained, which is the path of the moving body generated by the function of the control device based on the existence of the pedestrian group after removing one person one by one from the plurality of pedestrians; Extract individual paths whose difference from the baseline path is smaller compared to other individual paths; Extract pedestrians corresponding to the extracted individual paths; as well as The learned model is generated by machine learning in a manner that outputs information about pedestrians who are excluded from the learning process when the extracted pedestrian features are input into the learned model.

10. A filtering device adapted to a control device that controls a moving body autonomously moving in a pedestrian walking area, the control device identifying objects including the pedestrians and generating a future path for the moving body based on the identification result, wherein... The filter device uses a filter to limit the number of pedestrians that the control device is monitoring. The filter is generated by a filter manufacturing device through a process that includes: Obtain multiple feature values ​​of multiple pedestrians who become samples; A baseline path is obtained, which is the path of the moving body generated by the function of the control device based on the presence of the multiple pedestrians; A separate path is obtained, which is the path of the moving body generated by the function of the control device based on the existence of the pedestrian group after removing one person one by one from the plurality of pedestrians; Extract individual paths whose difference from the baseline path is smaller compared to other individual paths; Extract pedestrians corresponding to the extracted individual paths; as well as The learned model is generated by machine learning in a manner that outputs information about pedestrians who are excluded from the learning process when the extracted pedestrian features are input into the learned model.

Images