Method, program, autoencoder, and device for generating traffic scenarios.

Abstract

This provides a method for generating traffic scenarios that can simulate complex traffic environments. [Solution] The traffic scenario generation method is performed by a computer and includes a placement step S10 in which anchor boxes are placed at each of a plurality of different locations included in map information corresponding to the traffic scenario, and a representation step S20 in which the position of each of one or more objects, each of which is a moving body included in the traffic scenario, is represented using an offset relative to the anchor box closest to the object.

Description

[Technical Field]

[0001] The present invention relates to a method, program, autoencoder, and device for generating traffic scenarios. [Background technology]

[0002] Currently, research on autonomous driving of mobile objects such as vehicles is actively progressing. Simulation technologies are being developed to evaluate control methods in autonomous driving (see, for example, Patent Document 1). This allows for the safe evaluation of autonomous driving control methods without actually performing autonomous driving of vehicles on roads. When evaluating autonomous driving through simulation, traffic scenarios that mimic various surrounding environments of the autonomous vehicle are necessary. Here, a traffic scenario is a traffic environment represented by the position and size of objects at each time point. Since the proportion of accidents at intersections is high, it is necessary to perform simulations using traffic scenarios at intersections. Such traffic scenarios can be obtained, for example, by collecting data through actual driving. However, it is difficult to obtain a sufficient number and variety of traffic scenarios for simulation through actual driving. Therefore, the generation of traffic scenarios using generative models is being researched. For example, one technology used to generate traffic scenarios is the Variational Autoencoder (VAE), which is a model that compresses the dimensionality of the input and reconstructs it (see, for example, Patent Document 1). [Prior art documents] [Patent Documents]

[0003] [Patent Document 1] UK Patent Application Publication No. 2618419 [Overview of the Initiative] [Problems that the invention aims to solve]

[0004] However, conventional models cannot accurately reproduce traffic scenarios that simulate complex traffic environments such as intersections because they represent traffic scenarios only by the position of objects. Conventional models may generate inaccurate traffic scenarios, such as vehicles entering the wrong lane or vehicles colliding with obstacles.

[0005] This invention has been made in view of the above problems, and provides a method for generating traffic scenarios that can generate traffic scenarios that mimic complex traffic environments. [Means for solving the problem]

[0006] To achieve the above objective, a method for generating a traffic scenario according to one aspect of the present invention is a method for generating a traffic scenario performed by a computer, comprising: a placement step of placing anchor boxes at each of a plurality of different locations included in map information corresponding to the traffic scenario; and a representation step of representing the position of one or more objects, each of which is a moving body included in the traffic scenario, using an offset relative to the anchor box closest to the object.

[0007] Furthermore, in order to achieve the above objective, a program according to one aspect of the present invention is a program that causes a computer to execute the above-mentioned method for generating traffic scenarios.

[0008] Furthermore, in order to achieve the above objective, an autoencoder according to one aspect of the present invention is an autoencoder that is executed by a computer and reconstructs a traffic scenario, wherein an anchor box is placed at each of a plurality of different locations included in map information corresponding to the traffic scenario, and the position of each of one or more objects, each of which is a moving body included in the traffic scenario, is represented using an offset to the nearest anchor box to the object, and the autoencoder receives the positions of each of the one or more objects represented using the offset to the anchor box as input and outputs the reconstructed traffic scenario.

[0009] Furthermore, in order to achieve the above objective, a traffic scenario generation device according to one aspect of the present invention is a traffic scenario generation device that generates a traffic scenario, comprising: an arrangement unit that places anchor boxes at each of a plurality of mutually different locations included in map information corresponding to the traffic scenario; and an expression unit that represents the position of each of one or more objects, each of which is a moving body included in the traffic scenario, using an offset relative to the anchor box closest to the object.

[0010] These comprehensive or specific embodiments may be implemented as a system, method, integrated circuit, computer program, or recording medium such as a non-temporary computer-readable CD-ROM, or as any combination of a system, method, integrated circuit, computer program, and recording medium. [Effects of the Invention]

[0011] According to the present invention, it is possible to provide a method for generating traffic scenarios that can generate traffic scenarios that mimic complex traffic environments. [Brief explanation of the drawing]

[0012] [Figure 1] This is a block diagram showing an example of the functional configuration of a traffic scenario generation device according to an embodiment. [Figure 2] This is a plan view showing an example of an anchor box according to an embodiment. [Figure 3] This is a schematic plan view showing an object and an anchor box located in its vicinity. [Figure 4] This is a schematic plan view showing the offset of an object relative to the anchor box closest to the object. [Figure 5] This is a schematic plan view illustrating a method for representing a traffic scenario using an anchor box according to an embodiment. [Figure 6] This is a diagram showing an example of map information. [Figure 7] This figure shows the functional configuration of an autoencoder according to an embodiment. [Figure 8] It is a flowchart showing a method for generating a traffic scenario according to an embodiment. [Figure 9] It is a diagram showing a ground truth box indicating the correct position of an object in a dataset used for training an autoencoder and a reconstructed box indicating the position of the reconstructed object. [Figure 10] It is a diagram showing an overlapping region which is a region where the ground truth box and the reconstructed box overlap. [Figure 11] It is a diagram showing a combined region which is a region combining the ground truth box and the reconstructed box. [Figure 12] It is a schematic diagram explaining the final displacement error (FDE). [Figure 13] It is a diagram showing an example of the current position of an object reconstructed by a traffic scenario generation device according to an embodiment. [Figure 14] It is a diagram showing an example of the displacement of an object from the current time to 1.5 seconds later of an object reconstructed by a traffic scenario generation device according to an embodiment.

Embodiments for Carrying Out the Invention

[0013] [[ID=二十六]] Hereinafter, embodiments of the present invention will be specifically described with reference to the drawings.

[0014] Note that each of the embodiments described below shows comprehensive or specific examples. The numerical values, shapes, materials, components, arrangement positions and connection forms of the components, steps, order of steps, etc. shown in the following embodiments are merely examples and are not intended to limit the present invention. In addition, among the components in the following embodiments, components not described in the independent claims indicating the most general concept are described as optional components.

[0015] In addition, each figure is a schematic diagram and is not necessarily drawn precisely. Also, in each figure, the same reference numerals are assigned to the same constituent members.

[0016] (Embodiment) A method for generating traffic scenarios, an autoencoder, and a traffic scenario generation device according to an embodiment will be described.

[0017] [Traffic Scenario Generator] The traffic scenario generation device according to this embodiment will be described with reference to Figure 1. Figure 1 is a block diagram showing an example of the functional configuration of the traffic scenario generation device 10 according to this embodiment.

[0018] The traffic scenario generation device 10 is a device that generates traffic scenarios. The traffic scenario generation device 10 according to this embodiment outputs traffic scenarios using the traffic scenario generation method according to this embodiment, which will be described later. The traffic scenario generation device 10 is used, for example, to simulate the traffic environment in which a vehicle travels in a simulation of autonomous driving of a vehicle in a traffic environment.

[0019] As shown in Figure 1, the traffic scenario generation device 10 comprises a layout unit 20 and a display unit 30. In this embodiment, the traffic scenario generation device 10 further comprises a map information acquisition unit 40, a generation unit 50, an output unit 60, and a storage unit 70.

[0020] The placement unit 20 is a processing unit that places anchor boxes at each of several different locations included in the map information corresponding to the traffic scenario. An anchor box is a virtual frame placed at a location included in the map information. Here, the anchor box will be explained using Figure 2. Figure 2 is a plan view showing an example of an anchor box AB according to this embodiment. Figure 2 shows an example of anchor boxes AB placed at each location on the orthoimage. An orthoimage is an image of the ground surface, including roads, etc., viewed vertically from above, as shown in Figure 2.

[0021] In the example shown in Figure 2, anchor boxes AB are rectangular frames placed around each pixel. In this embodiment, two anchor boxes are placed around each pixel: one rectangular anchor box AB with its vertical direction being the longitudinal direction in Figure 2, and another anchor box AB rotated 90 degrees relative to the first anchor box. Note that Figure 2 only shows the two anchor boxes AB placed around a single pixel. The method of placing anchor boxes is not limited to this. For example, anchor boxes may be placed around only some of the pixels in the orthomosaic image in Figure 2. Also, the number of anchor boxes placed around each pixel does not have to be two; it could be one or three or more. Furthermore, the shape of the anchor boxes does not have to be rectangular.

[0022] The placement unit 20 places anchor boxes based on map information and anchor box information stored in the storage unit 70, for example. Information regarding anchor boxes includes, for example, the size of the anchor box and the direction of its inclination. Here, the size of the anchor box may be a rectangular frame of a size similar to that of each of the one or more objects included in the traffic scenario. For example, an anchor box may be the smallest rectangular frame that surrounds an object in a top view. The placement unit 20 may also acquire map information and the like from sources other than the storage unit 70. For example, the placement unit 20 may acquire map information and the like that is input from outside the traffic scenario generation device 10.

[0023] The representation unit 30 is a processing unit that represents the position of one or more objects, each of which is a moving object included in a traffic scenario, using an offset relative to the anchor box closest to the object. By representing the position of each object in this way, the representation unit 30 can represent the relationships between multiple objects. In this embodiment, the representation unit 30 assigns a class, future information, etc., to each object. The representation unit 30 acquires, for example, the position information of one or more objects from the storage unit 70. The representation unit 30 may also acquire the position information of one or more objects from sources other than the storage unit 70. For example, the representation unit 30 may acquire the position information of one or more objects input from outside the traffic scenario generation device 10.

[0024] The method for representing the position of an object by the representation unit 30 according to this embodiment will be explained using Figures 3 to 5. Figure 3 is a schematic plan view showing an object and an anchor box located in its vicinity. Figure 4 is a schematic plan view showing the offset of an object with respect to the anchor box closest to the object. In Figure 4, the offset is indicated by an arrow. Figure 5 is a schematic plan view showing the method for representing a traffic scenario using the anchor boxes according to this embodiment.

[0025] As shown in Figure 3, there are multiple anchor boxes AB1 to AB3 positioned by the positioning unit 20 around the moving object V1. Note that the positioning of anchor boxes AB1 to AB3 is merely an example and is not limited to the positions shown in Figure 3. Of the multiple anchor boxes shown in Figure 3, the anchor box closest to object V1 is anchor box AB1. The distance between object V1 and the anchor boxes is defined, for example, by the distance between a point representing object V1 and the position where the anchor box is positioned. The point representing object V1 is defined, for example, by the center point or centroid in a top view of object V1. The position where the anchor box is positioned coincides, for example, with the center point or centroid of the anchor box. As shown in Figure 4, the position of object V1 is expressed using the offset of object V1 relative to anchor box AB1 (see the arrow shown in Figure 4). In this embodiment, the offset includes the left-right offset ΔX and the up-down offset ΔY in Figure 4 with respect to the coordinates where the anchor box AB1 is positioned, as well as the width offset ΔW, the length offset ΔL, and the inclination offset ΔYaw with respect to the anchor box AB1.

[0026] As shown in Figure 5, the representation unit 30 according to this embodiment represents not only the current position of object V1 (i.e., the position at time t=t0), but also the positions at times t=t1, t2, and t3. Times t1, t2, and t3 are, for example, 0.5 seconds, 1.0 seconds, and 1.5 seconds after t0, respectively. Thus, the representation method of the representation unit 30 when generating a traffic scenario from the present (time t0) to 1.5 seconds later will be described. The position of object V1 at time t=t1 is represented, for example, by an offset from the position of object V1 at time t=t0. Similarly, the position of object V1 at time t=t2 is represented by an offset from the position of object V1 at time t=t1, and the position of object V1 at time t=t3 is represented by an offset from the position of object V1 at time t=t2. Note that the representation method of the position of object V1 at times t=t1 to t3 is not limited to this. For example, the position of object V1 at times t=t1 to t3 may be expressed as an offset relative to the anchor box closest to object V1, similar to the position of object V1 at time t=t0.

[0027] The map information acquisition unit 40 acquires map information corresponding to a traffic scenario. The map information indicates a region that includes the location of one or more objects, each of which is a moving object included in the traffic scenario. The map information acquired by the map information acquisition unit 40 will be explained using Figure 6. Figure 6 is a diagram showing an example of map information. In this embodiment, as shown in Figure 6, the map information includes an orthoimage (in other words, an orthomap) corresponding to the traffic scenario. The map information acquisition unit 40 acquires map information from, for example, the storage unit 70. However, the map information acquisition unit 40 may acquire map information from sources other than the storage unit 70. For example, the map information acquisition unit 40 may acquire map information input from outside the traffic scenario generation device 10.

[0028] The generation unit 50 is a processing unit that generates traffic scenarios. The generation unit 50 generates traffic scenarios using an autoencoder. The generation unit 50 inputs the position of one or more objects, represented using an offset relative to an anchor box, to an autoencoder that reconstructs the traffic scenario. In this embodiment, the generation unit 50 inputs map information indicating a region containing the position of one or more objects to the autoencoder.

[0029] The autoencoder used by the generation unit 50 receives the position of each of one or more objects represented using an offset relative to an anchor box as input and outputs a reconstructed traffic scenario. The autoencoder used by the generation unit 50 in reconstructing the traffic scenario is a trained autoencoder. The autoencoder according to this embodiment will be explained with reference to Figure 7. Figure 7 is a diagram showing the functional configuration of the autoencoder 50M according to this embodiment.

[0030] The autoencoder 50M is operated by a computer to reconstruct a traffic scenario. The autoencoder 50M takes the position of one or more objects, expressed using an offset relative to an anchor box, as input and outputs the reconstructed traffic scenario. As shown in Figure 7, the autoencoder 50M has a first encoder 51, a second encoder 52, a latent space 53, and a decoder 54.

[0031] The first encoder 51 is an encoder that receives a traffic scenario, including the positions of one or more objects, and convolves the traffic scenario. In this embodiment, the first encoder 51 receives information on one or more objects at the current time and a difference vector from the current time to 1.5 seconds later. The first encoder 51 extracts the features of the traffic scenario and compresses its dimensions.

[0032] The second encoder 52 is an encoder that receives map information corresponding to a traffic scenario and convolves the map information. In this embodiment, the second encoder 52 extracts feature quantities from the map information, including orthomosaic images, and compresses its dimensions.

[0033] The latent space 53 is a low-dimensional feature vector formed by combining the latent space compressed by the first encoder 51 and the feature vector extracted by the second encoder.

[0034] Decoder 54 is a decoder that reconstructs a traffic scenario using the latent space 53 as input.

[0035] In this embodiment, the autoencoder 50M is a variational autoencoder. That is, in the autoencoder 50M, the compressed information acquired by the first encoder 51 is not used directly as the latent space, but rather a normal distribution with the compressed information as a parameter is used as the latent space. In the autoencoder 50M, the distribution used as the latent space may be a distribution other than a normal distribution. In this case, in order to enable learning by backpropagation, the latent space (Z) is expressed using the following differentiable equation (1).

[0036]

number

[0037] Here, the parameters μ and σ represent compressed information, and ε represents a random number based on a standard normal distribution (i.e., a normal distribution with mean 0 and standard deviation 1).

[0038] In the autoencoder 50M according to this embodiment, the loss function that shows the error between the input traffic scenario and the reconstructed traffic scenario has a regularization term and a reconstruction term. In the regularization term, the Kullback-Leibler divergence (D) is calculated so that the latent space approaches a standard normal distribution. KLThe following loss function is calculated. In the reconstruction term, if you use MSE (Mean Squared Error), which is usually used in variational autoencoders, the proportion of areas where no objects exist on the orthomosaic image is very high, so it is not possible to calculate the loss targeting only objects. Therefore, in this embodiment, the following loss function used in PointPillars, a 3D point cloud-based object detection model, is used.

[0039]

number

[0040] Here, q φ (z|x) is the approximate distribution that the autoencoder 50M learns, and p θ (z) is the prior distribution of the latent variable z. In this embodiment, the loss (L) is the loss for the object position and size at the current time. box ) and the loss (L) for objects from the current time to 1.5 seconds later. future The calculation is performed using smoothL1 as shown below.

[0041]

number

[0042] Also, the loss on the object's class (L cls The following is calculated using the Focal Loss.

[0043]

number

[0044] Here, p represents the class probability.

[0045] Also, there is a loss regarding the orientation of the object at the current time (L dir This is calculated using Cross Entropy.

[0046] β shown in each of the above expressions box β cls β dir β future β KL β, α, and γ are parameters. In the present embodiment, each parameter is set as follows, for example, to perform learning of the autoencoder 50M.

[0047]

Number

[0048] The output unit 60 outputs the traffic scenario reconstructed by the generation unit 50. The output traffic scenario is used, for example, in the simulation of autonomous driving.

[0049] The storage unit 70 stores information used in the traffic scenario generation device 10. In the present embodiment, the storage unit 70 stores information corresponding to the traffic scenario and the like.

[0050] The traffic scenario generation device 10 according to the present embodiment has the above configuration, and thus can generate a traffic scenario.

[0051] [Traffic Scenario Generation Method] The traffic scenario generation method according to the present embodiment will be described using FIG. 8. FIG. 8 is a flowchart showing the traffic scenario generation method according to the present embodiment.

[0052] The traffic scenario generation method according to the present embodiment is executed by a computer. Hereinafter, an example of the traffic scenario generation method executed in the above-described traffic scenario generation device 10 will be described.

[0053] As shown in FIG. 8, in the traffic scenario generation method according to the present embodiment, first, the arrangement unit 20 of the traffic scenario generation device 10 arranges an anchor box at each of a plurality of different positions included in the map information corresponding to the traffic scenario (arrangement step S10).

[0054] Next, the representation unit 30 of the traffic scenario generation device 10 represents the position of one or more objects, each of which is a moving body included in the traffic scenario, using an offset relative to the anchor box closest to the object (representation step S20).

[0055] Next, the generation unit 50 of the traffic scenario generation device 10 inputs the position of one or more objects, represented using an offset relative to the anchor box, to the autoencoder 50M that reconstructs the traffic scenario (input step S30). In this embodiment, the autoencoder 50M is a variational autoencoder. In this embodiment, the generation unit 50 also inputs map information indicating a region containing the positions of one or more objects to the autoencoder 50M. The map information may include orthomosaic images.

[0056] Next, the generation unit 50 generates a reconstructed traffic scenario using the autoencoder 50M (generation step S40).

[0057] Next, the output unit 60 of the traffic scenario generation device 10 outputs the reconstructed traffic scenario (output step S50).

[0058] The traffic scenario generation method according to this embodiment can generate traffic scenarios as described above.

[0059] [Experimental results and effects] This document describes the experimental results and effects of the traffic scenario generation device 10, the traffic scenario generation method, and the autoencoder 50M according to this embodiment.

[0060] In the experiment, 346 traffic scenarios from multiple intersections were used as the dataset. These traffic scenarios were obtained through data collection in real-world environments. To generate traffic scenarios by focusing on objects on the road, only cars (passenger cars), buses, trucks, and truck trailers were used as moving objects. Furthermore, orthomosaic images were used as map information to learn about the surrounding environment. Therefore, object information representing the traffic scenarios is converted to the orthomosaic coordinate system before use.

[0061] Using the dataset described above, we trained an autoencoder 50M, and then reconstructed the training dataset using the trained autoencoder 50M. We then evaluated the traffic scenarios reconstructed in this way. Below, we describe the evaluation results for the object reconstruction accuracy at the current time and the object reconstruction accuracy 1.5 seconds after the current time.

[0062] The reconstruction accuracy of objects at the current time was evaluated using IoU (Intersection over Union). The calculation of IoU is explained using Figures 9 to 11. Figure 9 shows the ground truth box B1, which shows the correct position of an object in the dataset used to train the 50M autoencoder, and the reconstruction box B1r, which shows the position of the reconstructed object. Figure 10 shows the overlapping region R1, which is the region where the ground truth box B1 and the reconstruction box B1r overlap. In Figure 10, the area hatched with diagonal lines indicates the overlapping region R1. The overlapping region R1 is the region included in both the ground truth box B1 and the reconstruction box B1r. Figure 11 shows the combined region R2, which is the region formed by combining the ground truth box B1 and the reconstruction box B1r. In Figure 11, the area hatched with diagonal lines indicates the combined region R2. The combined region R2 is the region included in at least one of the ground truth box B1 and the reconstruction box B1r.

[0063] IoU is expressed as the ratio of the area of ​​the overlapping region R1 shown in Figure 10 to the area of ​​the joined region R2 shown in Figure 11. In other words, IoU increases as the discrepancy between the correct box B1 and the reconstructed box B1r decreases, and approaches 1.

[0064] The reconstruction accuracy of the object 1.5 seconds after the current time was evaluated using the final displacement error (FDE). FDE is explained using Figure 12. Figure 12 is a schematic diagram illustrating FDE. The solid arrows in Figure 12 indicate the correct displacement of object V1. The two black circles on the solid arrows indicate the correct displacement of object V1 0.5 seconds and 1.0 seconds after the current time. The dashed arrows in Figure 12 indicate the displacement of the reconstructed object V1. The two black circles on the dashed arrows indicate the displacement of the reconstructed object V1 0.5 seconds and 1.0 seconds after the current time. As shown by the dotted arrows in Figure 12, the FDE of object V1 1.5 seconds after the current time is expressed as the difference between the correct displacement of object V1 at 1.5 seconds and the displacement of the reconstructed object V1 at 1.5 seconds. Therefore, the FDE decreases as the displacement of the reconstructed object approaches the correct displacement, and approaches 0.

[0065] The results of this experiment will be explained using Figures 13 and 14. Figure 13 is a diagram showing an example of the current position of an object reconstructed by the traffic scenario generation device 10 according to this embodiment. In Figure 13, the correct position of the object is shown by a rectangular frame, and the reconstructed position of the object is shown by a hatched area. Eleven objects are shown in Figure 13. Figure 14 is a diagram showing an example of the displacement of an object reconstructed by the traffic scenario generation device 10 according to this embodiment from the current time to 1.5 seconds later. In Figure 14, the position of the reconstructed object at the current time is shown by a hatched area, and the displacement of the reconstructed object at 0.5 seconds, 1.0 seconds, and 1.5 seconds later from the current time is shown by marks including dashed frames. Also in Figure 14, the correct displacement of the object at 0.5 seconds, 1.0 seconds, and 1.5 seconds later from the current time is shown by marks including solid frames.

[0066] As shown in Figure 13, the current positions of the objects are reproduced with good accuracy. The average IoU is 0.78, indicating that the current positions of most objects are correctly reproduced.

[0067] As shown in Figure 14, the displacement of the objects is reproduced to a position that is generally close to the correct displacement. The average FDE is 0.628m, and considering the width of the road (approximately 2.75m or more and 3.5m or less) and the width of the objects (approximately 1.7m or more and 2.5m or less), it can be said that most of the objects are contained within the road.

[0068] As described above, in the traffic scenario generation device 10, traffic scenario generation method, and autoencoder 50M according to this embodiment, the position of each of one or more objects is represented using an offset relative to the anchor box closest to that object. This makes it possible to learn focusing only on one or more objects, and thus enables the accurate generation of traffic scenarios that simulate complex traffic environments that take into account the relationships between multiple objects.

[0069] Furthermore, in this embodiment, map information indicating a region containing the respective positions of one or more objects is input to the second encoder 52 separately from the individual position information of one or more objects, and convolution is performed to extract only the feature quantities of the map information separately from the object information. Therefore, even in complex traffic scenarios, the feature quantities of the map information can be extracted with high accuracy.

[0070] Furthermore, in this embodiment, the autoencoder 50M is a variational autoencoder. This allows the autoencoder 50M to learn the latent features of the dataset, making it possible to generate new traffic scenarios different from the dataset used for training.

[0071] Furthermore, in this embodiment, the map information includes orthomosaic images. This allows, for example, orthomosaic images created from point cloud information obtained by LiDAR (Light Detection and Ranging) used in autonomous driving to be used as map information.

[0072] (Other variations, etc.) The above describes a method for generating traffic scenarios according to one aspect of the present invention, based on embodiments. However, the present invention is not limited to these embodiments. Various modifications to the embodiments that a person skilled in the art could conceive of are also included within the scope of the present invention, as long as they do not depart from the spirit of the present invention.

[0073] For example, in the above embodiment, only cars, buses, trucks, and truck trailers were dealt with as objects, but other moving objects may also be dealt with. For example, motorcycles, bicycles, etc., may be dealt with as objects.

[0074] Furthermore, the following forms may also be included within the scope of one or more aspects of this disclosure.

[0075] (1) At least some of the components included in the traffic scenario generation device 10 described above may be a computer system or a network-connected server consisting of a microprocessor, ROM, RAM, hard disk unit, etc. A computer program is stored in the RAM or hard disk unit. The microprocessor achieves its function by operating in accordance with the computer program. Here, the computer program is composed of a combination of multiple instruction codes that indicate commands to the computer in order to achieve a predetermined function.

[0076] (2) At least some of the components included in the traffic scenario generation device 10 described above may be made up of a single system LSI (Large Scale Integration). The system LSI is a multi-functional LSI manufactured by integrating multiple components onto a single chip, and specifically, it is a computer system that includes a microprocessor, ROM, RAM, etc. A computer program is stored in the RAM. The system LSI achieves its function by operating the microprocessor in accordance with the computer program.

[0077] (3) At least some of the components included in the traffic scenario generation device 10 described above may consist of an IC card or a standalone module that can be attached to or removed from each device. The IC card or module is a computer system consisting of a microprocessor, ROM, RAM, etc. The IC card or module may include the multi-functional LSI described above. The IC card or module achieves its function by operating the microprocessor in accordance with the computer program. The IC card or module may be tamper-resistant.

[0078] (4) Furthermore, at least a portion of the components included in the traffic scenario generation device 10 described above may be recorded on a recording medium that can be read by a computer, such as a flexible disk, hard disk, CD-ROM, MO, DVD, DVD-ROM, DVD-RAM, BD (Blu-ray® Disc), semiconductor memory, etc. Alternatively, the digital signal may be recorded on one of these recording media.

[0079] Furthermore, at least some of the components included in the traffic scenario generation device 10 described above may transmit the computer program or the digital signal via telecommunications lines, wireless or wired communication lines, networks such as the Internet, data broadcasting, etc.

[0080] (5) The disclosure may be a method for generating the traffic scenarios described above. Alternatively, it may be a computer program for implementing the method for generating the traffic scenarios described above using a computer, or it may be a digital signal consisting of the computer program. Furthermore, the disclosure may be implemented as a non-temporary computer-readable recording medium such as a CD-ROM on which the computer program is recorded.

[0081] (6) The Disclosure may also provide a computer system comprising a microprocessor and memory, wherein the memory stores the computer program, and the microprocessor operates in accordance with the computer program.

[0082] (7) Alternatively, the program or the digital signal may be carried out by another independent computer system by recording it on the recording medium and transferring it, or by transferring the program or the digital signal via the network or the like.

[0083] (8) The above embodiments and the above modified examples may be combined.

[0084] (Note) Furthermore, the following technologies are disclosed based on the above description.

[0085] (Technical 1) A method for generating a traffic scenario, performed by a computer, comprising: a placement step of placing anchor boxes at each of a plurality of different locations included in map information corresponding to the traffic scenario; and a representation step of representing the position of one or more objects, each of which is a moving body included in the traffic scenario, using an offset relative to the anchor box closest to the object.

[0086] (Technical 2) A method for generating a traffic scenario according to Technical 1, comprising an input step of inputting the position of each of the one or more objects represented using an offset relative to the anchor box to an autoencoder for reconstructing the traffic scenario.

[0087] (Technical 3) The method for generating a traffic scenario according to Technical 2, wherein in the input step, map information indicating a region including the position of each of the one or more objects is input to the autoencoder.

[0088] (Technology 4) The method for generating traffic scenarios according to Technology 2 or 3, wherein the autoencoder is a variational autoencoder.

[0089] (Technical 5) The method for generating a traffic scenario described in Technical 3, which includes an orthophoto of the map information.

[0090] (Technology 6) A program for causing a computer to execute one of the traffic scenario generation methods described in Technology 1 to 5.

[0091] (Technical 7) An autoencoder that is performed by a computer to reconstruct a traffic scenario, wherein an anchor box is placed at each of a plurality of different locations included in map information corresponding to the traffic scenario, and the position of each of one or more objects, each of which is a moving body included in the traffic scenario, is represented using an offset to the nearest anchor box to the object, and the autoencoder receives the positions of each of the one or more objects represented using the offset to the anchor box as input and outputs the reconstructed traffic scenario.

[0092] (Technical 8) A traffic scenario generation device for generating traffic scenarios, comprising: an arrangement unit for arranging anchor boxes at each of a plurality of different locations included in map information corresponding to the traffic scenario; and an expression unit for representing the position of one or more objects, each of which is a moving body included in the traffic scenario, using an offset relative to the anchor box closest to the object. [Industrial applicability]

[0093] A method for generating traffic scenarios according to one aspect of the present invention can be applied, for example, to an evaluation device for the autonomous driving of a vehicle. [Explanation of Symbols]

[0094] 10. Traffic Scenario Generator 20 Placement section 30 Expression part 40 Map Information Acquisition Unit 50 Generation part 50M Autoencoder 51 First Encoder 52 Second Encoder 53 Latent space 54 Decoders 60 Output section 70 Storage section AB, AB1, AB2, AB3 Anchor Boxes B1 Correct Answer Box B1r Reconfiguration Box V1 object

Claims

1. A method for generating traffic scenarios, which is performed by a computer, The placement step involves placing anchor boxes at each of several different locations included in the map information corresponding to the aforementioned traffic scenario, The process includes a representation step of representing the position of one or more objects, each of which is a moving body included in the traffic scenario, using an offset relative to the anchor box closest to the object. Method for generating traffic scenarios.

2. The input step includes inputting the position of each of the one or more objects, expressed using the offset relative to the anchor box, to an autoencoder that reconstructs the traffic scenario. A method for generating a traffic scenario according to claim 1.

3. In the input step, map information indicating a region including the position of each of the one or more objects is input to the autoencoder. A method for generating a traffic scenario according to claim 2.

4. The autoencoder is a variational autoencoder. A method for generating a traffic scenario according to claim 2 or 3.

5. The aforementioned map information includes orthophotos. A method for generating a traffic scenario according to claim 3.

6. To cause a computer to execute the traffic scenario generation method according to any one of claims 1 to 3, 5. program.

7. An autoencoder that is executed by a computer and reconstructs traffic scenarios, Anchor boxes are placed at each of several different locations included in the map information corresponding to the aforementioned traffic scenario. The position of each of the one or more objects, each of which is a moving body included in the traffic scenario, is expressed using an offset relative to the anchor box closest to the object. The aforementioned autoencoder is The position of each of the one or more objects, expressed using the offset relative to the anchor box, is input. Output the reconstructed traffic scenario. Autoencoder.

8. A traffic scenario generation device that generates traffic scenarios, A placement unit for placing anchor boxes at each of several different locations included in the map information corresponding to the aforementioned traffic scenario, The system includes a representation unit that represents the position of one or more objects, each of which is a moving body included in the traffic scenario, using an offset relative to the anchor box closest to the object. Traffic scenario generation device.

Images