Data processing algorithm evaluation device

Abstract

To appropriately evaluate the performance of a data processing algorithm for determining the situation of a vehicle.SOLUTION: A data processing algorithm evaluation device comprises: a data storage section which stores a plurality of radio wave range finding data items detected from a vehicle; a data generation section by which, in a case where disturbance information indicating a disturbance to target data in the plurality of radio wave range finding data items stored in the data storage section is inputted, the target data are acquired and comprehended and synthetic data are generated by processing the target data on the basis of the comprehension such that the target data are reflected with the disturbance; and a data processing section which evaluates the performance of a data processing algorithm for determining the situation of the vehicle on the basis of the generated synthetic data.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present disclosure relates to a data processing algorithm evaluation device. [Background technology]

[0002] Vehicles equipped with driving support functions such as autonomous driving are being developed. Vehicles equipped with such driving support functions are equipped with an image processing algorithm that captures the surrounding environment of the vehicle using an on-board camera or the like and judges the vehicle's situation based on the captured images.

[0003] Various performance evaluations are being carried out on image processing algorithms to realize appropriate driving support functions. For example, a technique is known in which the performance of an image processing algorithm is evaluated by performing image processing using a composite image in which an image of weather disturbance created by computer graphics is superimposed on an actual image taken from a vehicle (see, for example, Patent Document 1). [Prior art documents] [Patent documents]

[0004] [Patent Document 1] JP 2010-033321 A Summary of the Invention [Problem to be solved by the invention]

[0005] In the technology described in Patent Document 1, the generated image of weather disturbances, etc. is not generated to reflect the state of objects, etc., shown in the actual image. Therefore, if the image processing algorithm is evaluated using the above-mentioned synthetic image, it may not be evaluated appropriately. In addition, it has been proposed to judge the state of the vehicle based on information other than the above-mentioned image, and in this case, it is required to appropriately evaluate the performance of the information processing algorithm.

[0006] The present disclosure has been made in consideration of the above, and has an object to provide a data processing algorithm evaluation device capable of appropriately evaluating the performance of a data processing algorithm that determines the situation of a vehicle. [Means for solving the problem]

[0007] The data processing algorithm evaluation device of the present disclosure includes a data storage unit that stores multiple radio ranging data detected from a vehicle, a data generation unit that, when disturbance information indicating a disturbance to target data among the multiple radio ranging data stored in the data storage unit is input, acquires and understands the target data, and processes the target data based on the understanding so that the disturbance is reflected in the target data to generate synthetic data, and a data processing unit that evaluates the performance of a data processing algorithm that determines the situation of the vehicle based on the generated synthetic data. Effect of the Invention

[0008] According to the present disclosure, it becomes possible to appropriately evaluate the performance of a data processing algorithm that determines the state of a vehicle. [Brief description of the drawings]

[0009] [Figure 1] FIG. 1 is a functional block diagram showing an example of a data processing algorithm evaluation device according to the first embodiment. [Diagram 2] FIG. 2 is a diagram showing an example of radio wave ranging data. [Diagram 3] FIG. 3 is a diagram illustrating an example of a process in the data generating unit. [Figure 4] FIG. 4 is a flowchart showing the flow of the process for generating data. [Diagram 5] FIG. 5 is a functional block diagram illustrating an example of a data processing algorithm evaluation device according to the second embodiment. [Figure 6] FIG. 6 is a conceptual diagram illustrating an example of the data learning unit. [Figure 7] FIG. 7 is a conceptual diagram showing another example of the data learning unit. [Figure 8] FIG. 8 is a diagram illustrating an example of a process performed by the data learning unit. [Figure 9] FIG. 9 is a diagram illustrating an example of a process performed by the data learning unit. [Figure 10] FIG. 10 is a flowchart showing the flow of the process for generating data. [Figure 11] FIG. 11 is a conceptual diagram illustrating an example of a data learning unit according to the third embodiment. [Figure 12] FIG. 12 is a flowchart showing the flow of the process for generating data. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0010] Hereinafter, an embodiment of a data processing algorithm evaluation device according to the present disclosure will be described with reference to the drawings. Note that the present invention is not limited to this embodiment. In addition, the components in the following embodiments include those that are replaceable and easy for a person skilled in the art, or those that are substantially the same.

[0011] [First embodiment] Fig. 1 is a functional block diagram showing an example of a data processing algorithm evaluation device according to the first embodiment. The data processing algorithm evaluation device 100 shown in Fig. 1 includes a calculation device, i.e., a CPU (Central Processing Unit), and a storage device, i.e., a memory for storing calculation contents, program information, and the like. The memory includes at least one of a RAM (Random Access Memory), a ROM (Read Only Memory), and an external storage device such as an HDD (Hard Disk Drive). As shown in Fig. 1, the data processing algorithm evaluation device 100 includes a data storage unit 10, a data generation unit 20, and a data processing unit 30.

[0012] The data storage unit 10 stores radio wave ranging data detected from a vehicle. The radio wave ranging data is data measured by a radio wave ranging method such as RADAR (Radio Detecting and Ranging). The radio wave ranging data is data obtained by, for example, emitting radio waves from a vehicle and receiving reflected waves of the radio waves.

[0013] FIG. 2 is a diagram showing an example of radio wave ranging data. As shown in FIG. 2, the radio wave ranging data includes reflection cross-sectional area data DT1 indicating the reflection cross-sectional area of ​​radio waves reflected by an object, and speed data DT2 indicating the relative speed between the vehicle and the object. In the data in FIG. 2, the circumferential direction centered on the reference position P indicates an azimuth angle based on the case of facing forward from the reference position P. Also, the direction perpendicular to the circumferential direction centered on the reference position P indicates the distance from the reference position P. The radio wave ranging data includes data arranged two-dimensionally in which coordinates are set for each unit azimuth angle and unit distance for the azimuth angle (θ) at which the radio waves emitted from the vehicle and reflected by the object arrive and the distance (r) to the object. Note that the radio wave ranging data may be data arranged two-dimensionally in which coordinates are set for each unit distance for positions in the forward / rearward direction (y) and left / right direction (x) with respect to the vehicle.

[0014] The reflection cross-sectional area data DT1 and the velocity data DT2 are generated in a normalized state, for example. In this embodiment, normalization includes setting the value indicating the reflection cross-sectional area in the reflection cross-sectional area data DT1 to M stages, and the value indicating the velocity in the velocity data DT2 to N stages, so that M and N are the same or approximately the same value. For example, when the value indicating the reflection cross-sectional area in the reflection cross-sectional area data DT1 is in the range of 0 to 250 and the value indicating the velocity in the velocity data DT2 is in the range of 0 to 60000, the range of values ​​that each data can take is significantly different. Therefore, by making the range of values ​​that each data can take the same or approximately the same (normalizing), the weight of each data can be made the same or approximately the same. For example, normalization can be performed by dividing each value of the reflection cross-sectional area data DT1 and the velocity data DT2 by a predetermined value α. In this case, α can be set to, for example, a value obtained by dividing the maximum value of the velocity data DT2 by the maximum value of the reflection cross-sectional area data DT1, or the like. Also, each value of the reflection cross-sectional area data DT1 and the velocity data DT2 may be divided by the above-mentioned predetermined value α, and then the root of at least one of the reflection cross-sectional area data DT1 and the velocity data DT2 may be taken. Note that, when normalization is performed, the embodiment is not limited to adjusting both the reflection cross-sectional area data DT1 and the velocity data DT2, and may be adjusting at least one of the values.

[0015] The data generating unit 20 generates synthetic data by combining the target data input from the data storage unit 10 with disturbances based on disturbance information input from an input unit (not shown) or the like. The data generating unit 20 has a data understanding unit 21 and a data processing unit 22. The data understanding unit 21 understands the data stored in the data storage unit 10. The data processing unit 22 processes the target data input to the data generating unit 20.

[0016] The data processing unit 30 performs data processing based on the generated composite data, and evaluates the performance of the data processing algorithm that judges the vehicle's situation. The data processing unit 30 processes the composite data using the data processing algorithm, and calculates judgment information for judging the vehicle's situation. The data processing unit 30 stores the calculated judgment information in a storage unit (not shown). The judgment information may be, for example, an approach intersection time, which is the time it takes for the vehicle to reach the position of an object in front of the vehicle from the detection position while the vehicle is traveling. The data processing unit 30 can evaluate the performance of the data processing algorithm based on whether or not the judgment information is significantly different between when data not including a disturbance is processed and when data including a disturbance is processed.

[0017] Fig. 3 is a diagram showing an example of processing in the data generating unit 20. As shown in Fig. 3, target data and disturbance information are input to the data generating unit 20. The target data may be, for example, radio wave ranging data detected ahead of the vehicle. The disturbance information may be, for example, weather conditions such as rain and fog. In this embodiment, a case where the disturbance is rain will be described as an example.

[0018] The data understanding unit 21 has calculation units 21a and 21b. In this embodiment, the data understanding unit 21 understands the spatial position (distance and azimuth angle) of the object. The calculation unit 21a acquires the input object data. The calculation unit 21a acquires the spatial position of the object constituting the object data for each coordinate from the detection position of the object data by data processing. The calculation unit 21a generates spatial position information that associates the acquired spatial position with the coordinate, and outputs the spatial position information.

[0019] The calculation unit 21b acquires the spatial position information output from the calculation unit 21a. The calculation unit 21b also acquires disturbance information input via an input unit (not shown). Examples of disturbance information include rain data generated by a simulator. This data may have the same resolution as the target data (radio wave ranging data), for example.

[0020] The calculation unit 21b calculates the strength of the disturbance for each coordinate based on the spatial position information. For example, in an environment where rain occurs, the influence of the rain increases as the distance from the detection position increases. Therefore, the disturbance can be adjusted to a state close to the actual environment by adjusting the strength of the disturbance for each coordinate according to the distance included in the spatial position information. After adjusting the disturbance for each coordinate, the calculation unit 21b outputs disturbance information including the adjusted disturbance.

[0021] The data processing unit 22 acquires the input target data. The data processing unit 22 acquires the disturbance information output from the calculation unit 21b. The data processing unit 22 generates synthetic data by superimposing the disturbance included in the acquired disturbance information on the target data for each coordinate. In other words, the disturbance information input to the data processing unit 22 is information in which the strength of the disturbance is adjusted for each coordinate according to the distance, in contrast to the disturbance information input from an input unit (not shown). By superimposing such adjusted disturbance on the target data, appropriate synthetic data closer to the actual disturbance environment can be generated.

[0022] The data processing unit 30 evaluates the performance of a data processing algorithm for determining the vehicle status based on the synthetic data. Since the performance of the data processing algorithm is evaluated based on appropriate synthetic data that is closer to the actual surrounding environment, appropriate evaluation results can be obtained without using a large-scale system, compared to the case where the vehicle status is determined using a simulator that reproduces the actual surrounding environment in detail in three dimensions.

[0023] Fig. 4 is a flowchart showing the flow of the process of generating data. As shown in Fig. 4, the data generating unit 20 acquires input target data in the data understanding unit 21 (step S101). The data understanding unit 21 acquires the spatial position of the target object corresponding to the coordinates constituting the target data from the detection position of the target data (step S102), and adjusts the disturbance by calculating the strength of the disturbance according to the distance based on the acquired spatial position (step S103). The data processing unit 22 generates composite data by combining the attenuated disturbance with the target data based on the input target data and the strength of the disturbance calculated by the calculation unit 21b (step S104).

[0024] As described above, the data processing algorithm evaluation device 100 of this embodiment comprises a data storage unit 10 that stores multiple radio ranging data detected from a vehicle, a data generation unit 20 that, when disturbance information indicating a disturbance to target data among the multiple radio ranging data stored in the data storage unit 10 is input, acquires and understands the target data, and processes the target data based on the understanding so that the disturbance is reflected in the target data to generate synthetic data, and a data processing unit 30 that evaluates the performance of a data processing algorithm that determines the vehicle situation based on the generated synthetic data.

[0025] According to this configuration, the data generator 20 processes the target data based on the result of understanding the target data so that the disturbance is reflected in the target data to generate synthetic data, so that synthetic data closer to the actual surrounding environment can be generated compared to the case where a disturbance generated by a simulator is simply superimposed. This allows the performance of the data processing algorithm for judging the vehicle situation to be appropriately evaluated.

[0026] In the data processing algorithm evaluation device 100 according to this embodiment, the data generation unit 20 acquires the spatial position of the object constituting the target data from the detection position of the target data, calculates the strength of the disturbance due to the spatial position based on the acquired spatial position, and generates synthetic data based on the calculation result. Therefore, synthetic data closer to the actual surrounding environment can be generated.

[0027] In the data processing algorithm evaluation device 100 according to this embodiment, the radio wave ranging data is data arranged two-dimensionally with respect to the azimuth angle at which the radio wave emitted from the vehicle and reflected by the object arrives and the distance to the object. Therefore, the radio wave ranging data can be handled in the same way as image data.

[0028] In the data processing algorithm evaluation device 100 according to the present embodiment, the radio wave ranging data includes reflection cross-sectional area data DT1 indicating the reflection cross-sectional area of ​​the radio wave reflected by the object, and speed data DT2 indicating the relative speed between the vehicle and the object. Therefore, since one radio wave ranging data includes two types of data, the reflection cross-sectional area data DT1 and the speed data DT2, the vehicle situation can be more appropriately determined.

[0029] [Second embodiment] Next, a second embodiment will be described. FIG. 5 is a functional block diagram showing an example of a data processing algorithm evaluation device according to the second embodiment. The data processing algorithm evaluation device 100 according to the second embodiment is configured to include a data storage unit 10, a data generation unit 20, and a data processing unit 30, similar to the first embodiment. In this embodiment, the processing content in the data generation unit 20 is different from that in the first embodiment. Also, a learning data storage unit 40 and a data learning unit 50 are added. In this embodiment, the radio wave ranging data includes reflection cross section data and velocity data, similar to the first embodiment.

[0030] In this embodiment, the learning data storage unit 40 stores, as learning data, reference data, which is radio distance measurement data that does not include disturbances, and disturbance data, which is radio distance measurement data that includes disturbances. An example of the reference data is radio distance measurement data detected under conditions where disturbances are the least, such as during the daytime on a clear day. An example of the disturbance data is radio distance measurement data detected under conditions where disturbances are greater than the reference data, such as during rain, snow, or fog. The learning data storage unit 40 may store reference data and disturbance data for the same or corresponding detection target in association with each other. For example, the learning data storage unit 40 may store reference data detected at a specific location during the daytime on a clear day, and disturbance data detected at the specific location during rain, snow, fog, or the like, in association with each other.

[0031] In this embodiment, when target data and disturbance information indicating disturbance to the target data are input, the data learning unit 50 can generate synthetic data for learning, for example, by learning using a neural network. For example, as a method for realizing a predetermined data processing using a neural network in the data learning unit 50, a technique called Generative Adversarial Networks (GAN) or a technique called Cycle Generative Adversarial Networks (Cycle GAN) can be used.

[0032] FIG. 6 is a conceptual diagram showing an example of the data learning unit 50 (data learning unit 50A). As shown in FIG. 6, the generative adversarial network used in the data learning unit 50A is composed of two neural networks, a data generating unit 51 and an authenticity determining unit 52. The data generating unit 51 has the same configuration as the data generating unit 20 described above, and includes a data understanding unit 51a and a data processing unit 51b. Note that the data understanding unit 51a in this embodiment understands the characteristics of the target data by a neural network. The data generating unit 51 generates synthetic data for learning by combining a disturbance with the reference data by the same process as the process in which the data generating unit 20 generates synthetic data based on the target data and disturbance information. The authenticity determining unit 52 determines the authenticity of the synthetic data for learning generated by the data generating unit 51 based on the synthetic data for learning generated by the data generating unit 51 and the disturbance data associated with the reference data. The data generating unit 51 generates synthetic data for learning so that the synthetic data is determined to be closer to the true data by the authenticity determining unit 52. Moreover, the authenticity determination unit 52 tries to detect differences between the generated synthetic data for learning and the true data. By alternately making the two networks compete with each other and proceeding with learning, the data generation unit 51 can generate synthetic data for learning that is close to the true disturbance data.

[0033] FIG. 7 is a conceptual diagram showing another example (data learning unit 50B) of the data learning unit 50. As shown in FIG. 7, the cycle generative adversarial network used in the data learning unit 50B has data generating units 51 and 53 and authenticity determining units 52 and 54. The data generating units 51 and 53 have the same configuration as the data generating unit 20 described above. The data generating unit 51 has a data understanding unit 51a and a data processing unit 51b. The data generating unit 53 has data understanding units 53a and 53b. The data generating unit 51 generates first synthetic data for learning by combining a disturbance with reference data. In addition, the data generating unit 53 generates second synthetic data for learning by removing the disturbance from the disturbance data. The authenticity determining unit 52 determines the authenticity of the first synthetic data for learning based on the true disturbance data. In addition, the authenticity determining unit 54 determines the authenticity of the second synthetic data for learning based on the true reference data. The data generating units 51, 53 generate the first synthetic data for training and the second synthetic data for training so that the authenticity determining units 52, 54 determine that the data is closer to the true data, i.e., so that the accuracy rate is higher. The authenticity determining units 52, 54 also attempt to detect differences between the generated first synthetic data for training and the second synthetic data for training and the true data. By alternately having these two networks compete with each other and progressing with learning, the data generating unit 51 is able to generate synthetic data for training that is closer to the true disturbance data, i.e., has a high accuracy rate for authenticity.

[0034] In the cycle generative adversarial network, it is sufficient to have reference data that does not include disturbances and disturbance data that includes disturbances, and it is not necessary for the reference data and disturbance data to be associated with each other. Therefore, compared to the generative adversarial network, it is easier to collect training data. In addition, when using the cycle generative adversarial network, multiple disturbance data with different levels of disturbance can be stored in the training data storage unit 40.

[0035] When using the above-mentioned generative adversarial network or cycle generative adversarial network, the data learning unit 50 processes the radio ranging data using a network such as KPConv. Such a network can extract features from the spatial arrangement of a point cloud, and is therefore effective for data with sparse information, such as radio ranging data from RADAR, etc.

[0036] FIG. 8 is a diagram showing an example of processing of radio wave distance measurement data. As shown in FIG. 8, the data learning unit 50 can extract data on target areas R1 and R2 set in advance for some azimuth angles and distances from the reflection cross-sectional area data DT1 and the speed data DT2, and use the extracted data as learning data. In the example shown in FIG. 8, the data learning unit 50 sets rectangular target areas R1 and R2 based on coordinates indicating the left-right and front-back positions (x, y), but this is not limited to this case. The data learning unit 50 may set the target areas R1 and R2 based on coordinates indicating the distance from the vehicle and the azimuth angle (r, θ), for example. The data learning unit 50 sets the target areas R1 and R2 so that the left-right and front-back positions (x, y), or the distance from the vehicle and the azimuth angle (r, θ), are in the same range.

[0037] In addition, the reflection cross-sectional area data DT1 and the velocity data DT2 may have a portion (non-detection portion) where the data of the object is not detected. In the target regions R1 and R2, no data exists at the coordinates corresponding to the non-detection portion D0. When such a non-detection portion is included in the target regions R1 and R2, the data learning unit 50 can set the value of the coordinates corresponding to the non-detection portion in the target regions R1 and R2 as a predetermined non-detection value. The non-detection portion of the reflection cross-sectional area data DT1 has the same coordinates as the non-detection portion in the velocity data DT2. Therefore, the data learning unit 50 only needs to determine whether or not there is a non-detection portion in the reflection cross-sectional area data DT1.

[0038] FIG. 9 is a schematic diagram showing an example of processing of radio wave distance measurement data. FIG. 9 is an enlarged view showing values ​​for each coordinate so that they can be distinguished. As shown in FIG. 9, in the target regions R1 and R2, there are more non-detected portions D0 that are not detected than the actual measurement portion D1 that is detected by actual measurement, when viewed for each coordinate. In the radio wave distance measurement data (reflection cross-sectional area data DT1, velocity data DT2) including the target regions R1 and R2 detected in this embodiment, the actual measurement portion D1 is sparse data. In this way, when the non-detected portion D0 exists in the target regions R1 and R2, the data learning unit 50 can set a value that can be distinguished from the value (actual measurement value) of the actual measurement portion D1 detected by actual measurement as the value (non-detection value) of the non-detected portion D0. For example, the data learning unit 50 can set a value smaller than the smallest actual measurement value among the actual measurement values ​​set in the data in the radio wave distance measurement data as the non-detection value.

[0039] The data learning unit 50 may set the non-detection value to different values ​​in the reflection cross-sectional area data DT1 and the velocity data DT2. For example, in the reflection cross-sectional area data DT1, the value of "0" detected by actual measurement is the same as the value when radio waves reflected from a human body are detected. If the non-detection value is set to 0 in the reflection cross-sectional area data DT1, the data will indicate that "a person is present" in the non-detection portion D0, so it is preferable to avoid such a situation. Therefore, when setting the non-detection value of the reflection cross-sectional area data DT1, the data learning unit 50 can set a value smaller than 0 as the non-detection value even if 0 corresponds to a value smaller than the minimum value detected by actual measurement. In addition, the data learning unit 50 can set the non-detection value in the reflection cross-sectional area data DT1 to, for example, a lower limit value detectable by actual measurement (for example, about -tens of dB, about -hundreds of dB, etc.).

[0040] Furthermore, the data learning unit 50 determines that the value of "0" detected in the speed data DT2 by actual measurement is the same value as when radio waves reflected from an object traveling at the same speed as the vehicle are detected. If the non-detection value is set to 0 in the speed data DT2, the non-detection portion D0 will contain data indicating that "an object traveling at the same speed as the vehicle is present." In this case, when setting the non-detection value of the speed data DT2, the data learning unit 50 can set the non-detection value to a value smaller than the lower limit of detectability (e.g., -80km / h, -150km / h) that is set based on the speed limit (e.g., 60km / h, 100km / h, etc.) of the road on which the vehicle travels in actual measurement.

[0041] In this way, by extracting the target regions R1 and R2 by the data learning unit 50 and setting the value of the non-detection part D0 to a non-detection value, the size and format of the radio wave ranging data can be standardized, and a network such as KPConv can be easily applied. In addition, by applying a network such as KPConv to such radio wave ranging data, it becomes possible to extract spatial arrangement features from meaningful sparse data. Furthermore, by setting a non-detection value according to the type of data (reflection cross-sectional area data DT1, velocity data DT2), it is possible to prevent the data from being learned with a meaning different from the original meaning.

[0042] FIG. 10 is a flowchart showing the flow of the process of generating data. As shown in FIG. 10, the data learning unit 50 acquires learning data stored in the learning data storage unit 40 (step S201). The data generating unit 51 understands the acquired learning data in the data understanding unit 51a (step S202). The data processing unit 51b generates synthetic learning data in which a disturbance is added to the learning data based on the understanding result (step S203). The authenticity determination unit 52 determines the authenticity of the synthetic learning data based on the true disturbance data (step S204). The data learning unit 50 makes the data generating unit 51 and the authenticity determination unit 52 compete with each other alternately to proceed with learning (step S205). The data learning unit 50 determines whether learning is completed (step S206), and if it is determined that learning is completed (Yes in step S206), proceeds to step S207. Moreover, when it is determined that the learning is not completed (No in step S206), the process from step S202 onward is repeated. In step S206, the data learning unit 50 can determine that the learning is completed, for example, when the difference between the synthetic data for learning and the true disturbance data is equal to or smaller than a predetermined value, that is, when the accuracy rate is equal to or larger than a predetermined value.

[0043] After the learning is completed, if the target data and disturbance information are input, the data generating unit 20 acquires the target data from the data storage unit 10 (step S207). The data understanding unit 21 understands the data based on the learning result (step S208). The data processing unit 22 generates synthetic data in which the disturbance is added to the target data based on the understanding result (step S209).

[0044] As described above, in the data processing algorithm evaluation device 100 according to this embodiment, the learning data storage unit 40 stores reference data, which is radio ranging data that does not include disturbances, and disturbance data, which is radio ranging data that includes disturbances, the data learning unit 50 learns using a generative adversarial network or a cycle generative adversarial network based on the reference data and the disturbance data, and the data generation unit 20 understands the target data based on the learning results of the data learning unit 50 and generates synthetic data.

[0045] In this configuration, the data learning unit 50 learns the radio ranging data using a generative adversarial network or a cycle generative adversarial network, and the data generating unit 20 understands the target data and generates synthetic data based on the learning results, so that it is possible to generate synthetic data that is close to the actual surrounding environment. This makes it possible to appropriately evaluate the performance of the data processing algorithm that judges the vehicle situation.

[0046] Moreover, in the data processing algorithm evaluation device 100 according to this embodiment, the learning data storage unit 40 stores a plurality of disturbance data including disturbances of the same type but different degrees, the data learning unit 50 performs learning based on the plurality of disturbance data including disturbances of different degrees, and when disturbance information including the degree of disturbance is input, the data generating unit 20 understands target data based on the learning result of the data learning unit 50 and generates synthetic data such that the target data reflects the disturbance of the degree corresponding to the disturbance information. This makes it possible to appropriately generate synthetic data with different degrees of disturbance, thereby making it possible to widely evaluate the performance of data processing algorithms.

[0047] In the data processing algorithm evaluation device 100 according to this embodiment, the data learning unit 50 sets a predetermined non-detection value to data that does not have an actual measurement value set among the data constituting the radio ranging data. This makes it possible to standardize the size and format of the radio ranging data, and to easily apply a network such as KPConv.

[0048] In the data processing algorithm evaluation device 100 according to this embodiment, the data learning unit 50 sets a value smaller than the smallest actual measurement value among the actual measurement values ​​set in the data in the radio wave ranging data as a non-detection value. This makes it possible to set a value that is clearly distinguishable from the value of the actual measurement part detected by actual measurement (actual measurement value) as a non-detection value.

[0049] [Third embodiment] Next, a third embodiment will be described. The data processing algorithm evaluation device 100 according to the third embodiment includes a data storage unit 10, a data generation unit 20, a data processing unit 30, a learning data storage unit 40, and a data learning unit 50, similar to the second embodiment. In this embodiment, the type of radio ranging data stored in the learning data storage unit 40 and the processing contents in the data generation unit 20 and the data learning unit 50 are different from those in the second embodiment. In this embodiment, the radio ranging data itself includes reflection cross section data and velocity data, similar to the above embodiment.

[0050] The learning data storage unit 40 stores the radio ranging data including the reference data and the disturbance data as learning data, as in the second embodiment. In this embodiment, the learning data storage unit 40 stores radio ranging data including different attributes. Examples of the radio ranging data including different attributes include multiple radio ranging data with different attributes such as detection locations, such as radio ranging data detected in a downtown area, radio ranging data detected in a residential area, and radio ranging data detected on a mountain path. The learning data storage unit 40 can store such radio ranging data in association with label information indicating attributes.

[0051] In this embodiment, when target data and disturbance information for the target data are input, the data generating unit 20 extracts label information indicating the above-mentioned attributes from the target data. The data generating unit 20 can understand the target data based on the learning result of the data learning unit 50 described later and generate synthetic data reflecting the attributes. For example, when the target data is daytime reference data for a "downtown" and "rain" is input as disturbance information, conversion that reflects the material of the surrounding structure, for example, a change in reflectance due to wetting of concrete, becomes possible. Also, for example, when the target data is daytime reference data for a "mountain path" and "rain" is input as disturbance information, conversion that reflects the change in reflectance due to wetting of trees becomes possible.

[0052] 11 is a conceptual diagram showing an example of a data learning unit 50 (data learning unit 50C) according to the third embodiment. As shown in FIG. 11, the cycle generative adversarial network used in the data learning unit 50C has data generation units 51 and 53 and authenticity determination units 52 and 54, similar to the second embodiment. In the cycle generative adversarial network according to this embodiment, label information is input to the data generation units 51 and 53.

[0053] In this case, the data generating unit 51 generates first synthetic data by synthesizing the disturbance with the reference data. For example, if the extracted attribute is "downtown" and the input disturbance is "rain", the synthetic data is generated so that the disturbance of rain is reflected in the reference data of the downtown. The data generating unit 53 generates second synthetic data by removing the disturbance of rain from the disturbance data of the downtown rain. In this case, the data generating units 51 and 53 can convert the target data according to the given attribute by storing the learning result for each attribute, for example. The authenticity determining unit 52 determines the authenticity of the first synthetic data based on the true disturbance data and determines whether the attribute of "downtown" is appropriately reflected. The authenticity determining unit 54 determines the authenticity of the second synthetic data based on the true reference data and determines whether the disturbance of rain in the downtown rain is appropriately removed. In this manner, in this embodiment, the authenticity determining units 52 and 54 determine the authenticity of the composite data as well as the appropriateness of reflecting the attributes.

[0054] In this embodiment, the data generation units 51 and 53 generate the first and second synthetic data so that the authenticity determination units 52 and 54 determine that the data is closer to the true data and that the attributes are appropriately reflected. The authenticity determination units 52 and 54 also attempt to detect differences between the generated first and second synthetic data and the true data, and points where the attributes are not appropriately reflected. By alternately having these two networks compete with each other and progressing with learning, the data generation unit 51 becomes able to generate synthetic data with a high accuracy rate for authenticity and a high accuracy rate for reflecting attributes.

[0055] FIG. 12 is a flowchart showing the flow of the process of generating data. As shown in FIG. 12, the data learning unit 50 acquires learning data stored in the learning data storage unit 40 (step S301). The data learning unit 50 extracts label information indicating attributes included in the acquired learning data (step S302). The data generating unit 51 understands the acquired learning data in the data understanding unit 51a (step S303). The data processing unit 51b generates synthetic learning data in which a disturbance is added to the learning data based on the understanding result (step S304). The authenticity determination unit 52 determines the authenticity of the synthetic learning data based on the true disturbance data (step S305). The data learning unit 50 makes the data generating unit 51 and the authenticity determination unit 52 compete with each other alternately to proceed with learning (step S306). The data learning unit 50 determines whether learning is completed (step S307), and if it is determined that learning is completed (Yes in step S307), proceeds to step S308. Moreover, when it is determined that the learning is not complete (No in step S307), the processing from step S303 onward is repeated. In step S307, the data learning unit 50 can determine that the learning is complete when, for example, the difference between the synthetic data for learning and the true disturbance data is equal to or less than a predetermined value, and the difference between the synthetic data for learning and the disturbance data in the part related to the attributes is equal to or less than a predetermined value, that is, when the accuracy rate of the authenticity of the synthetic data for learning and the accuracy rate of the part related to the attributes are equal to or greater than a predetermined value.

[0056] When the target data and disturbance information are input, the data generating unit 20 acquires the input target data (step S308). The data understanding unit 21 extracts label information indicating attributes contained in the target data (step S309) and understands the target data (step S310). The data processing unit 22 generates synthetic data in which a disturbance is added to the target data based on the understanding result (step S311).

[0057] As described above, in the data processing algorithm evaluation device 100 according to this embodiment, the learning data storage unit 40 stores a plurality of radio ranging data including different attributes in association with label information indicating the attributes, the data learning unit 50 learns using a cycle generative adversarial network so as to increase the accuracy rate regarding attributes between the generated synthetic data for learning and disturbance data, and when target data and disturbance information are input, the data generation unit 20 extracts label information indicating the attributes of the target data, understands the target data based on the learning result of the data learning unit 50, and generates synthetic data.

[0058] With this configuration, the cycle generative adversarial network can learn to increase the accuracy rate of determining the authenticity of the synthetic data used for training, and to increase the accuracy rate of reflecting attributes, thereby making it possible to generate synthetic data in a manner that corresponds to the attributes of the radio ranging data.

[0059] As described above, the data processing algorithm evaluation device according to the first aspect of the present disclosure includes a data storage unit 10 that stores multiple radio ranging data detected from a vehicle, a data generation unit 20 that, when disturbance information indicating a disturbance to target data among the multiple radio ranging data stored in the data storage unit 10 is input, acquires and understands the target data, and processes the target data based on the understanding so that the disturbance is reflected in the target data to generate synthetic data, and a data processing unit 30 that evaluates the performance of a data processing algorithm that determines the vehicle situation based on the generated synthetic data.

[0060] According to this configuration, the data generator 20 processes the target data based on the result of understanding the target data so that the disturbance is reflected in the target data to generate synthetic data, so that synthetic data closer to the actual surrounding environment can be generated compared to the case where a disturbance generated by a simulator is simply superimposed. This allows the performance of the data processing algorithm for judging the vehicle situation to be appropriately evaluated.

[0061] According to the second aspect of the present disclosure, in the data processing algorithm evaluation device according to the first aspect, the data generation unit 20 acquires the spatial position of the object constituting the target data from the detection position of the target data, calculates the strength of the disturbance due to the spatial position based on the acquired spatial position, and generates synthetic data based on the calculation result. Therefore, it is possible to generate synthetic data that is closer to the actual surrounding environment.

[0062] According to a third aspect of the present disclosure, in the data processing algorithm evaluation device according to the first or second aspect, the radio wave ranging data includes data arranged two-dimensionally on the azimuth angle at which the radio wave emitted from the vehicle and reflected by the object arrives and the distance to the object. Therefore, the radio wave ranging data can be handled in the same way as image data.

[0063] According to a fourth aspect of the present disclosure, in the data processing algorithm evaluation device according to any one of the first to third aspects, the radio wave ranging data includes reflection cross-sectional area data DT1 indicating a reflection cross-sectional area of ​​the radio wave reflected by an object, and speed data DT2 indicating a relative speed between the vehicle and the object. Therefore, since one radio wave ranging data includes two types of data, the reflection cross-sectional area data DT1 and the speed data DT2, the situation of the vehicle can be more appropriately determined.

[0064] According to the fifth aspect of the present disclosure, in the data processing algorithm evaluation device according to the third or fourth aspect, the learning data storage unit 40 stores reference data, which is radio ranging data that does not include disturbances, and disturbance data, which is radio ranging data that includes disturbances, and the data learning unit 50 learns using a generative adversarial network or a cycle generative adversarial network based on the reference data and the disturbance data, and the data generation unit 20 understands the target data based on the learning result of the data learning unit 50 and generates synthetic data. In this configuration, the data learning unit 50 learns the radio ranging data using a generative adversarial network or a cycle generative adversarial network, and the data generation unit 20 understands the target data based on the learning result and generates synthetic data, so that synthetic data close to the actual surrounding environment can be generated. This makes it possible to appropriately evaluate the performance of a data processing algorithm that determines the situation of a vehicle.

[0065] According to a sixth aspect of the present disclosure, in the data processing algorithm evaluation device according to the fifth aspect, the learning data storage unit 40 stores a plurality of disturbance data including disturbances of the same type but different degrees, the data learning unit 50 performs learning based on the plurality of disturbance data including disturbances of different degrees, and when disturbance information including the degree of disturbance is input, the data generating unit 20 understands target data based on the learning result of the data learning unit 50 and generates synthetic data such that the disturbance of the degree corresponding to the disturbance information is reflected in the target data. This makes it possible to appropriately generate synthetic data with different degrees of disturbance, thereby making it possible to widely evaluate the performance of data processing algorithms.

[0066] According to a seventh aspect of the present disclosure, in the data processing algorithm evaluation device according to the fifth or sixth aspect, the learning data storage unit 40 stores a plurality of radio ranging data including different attributes in association with label information indicating the attributes, the data learning unit 50 learns using a cycle generative adversarial network so as to increase the accuracy rate of attributes between the generated synthetic data for training and disturbance data, and the data generating unit 20, when receiving target data and disturbance information, extracts label information indicating the attributes of the target data, and generates synthetic data by understanding the target data based on the learning result of the data learning unit 50. With this configuration, the cycle generative adversarial network can learn so as to increase the accuracy rate of authenticity of the synthetic data for training and to increase the accuracy rate of reflection of the attributes, so that synthetic data can be generated in a manner according to the attributes of the radio ranging data.

[0067] According to the eighth aspect of the present disclosure, in the data processing algorithm evaluation device according to the fifth or sixth aspect, the data learning unit 50 sets a predetermined non-detection value to data that does not have an actual measurement value set among data constituting the radio ranging data. This makes it possible to standardize the size and format of the radio ranging data, and to easily apply a network such as KPConv.

[0068] According to a ninth aspect of the present disclosure, in the data processing algorithm evaluation device according to the eighth aspect, the data learning unit 50 sets a value smaller than the smallest actual measurement value among the actual measurement values ​​set in the data in the radio wave ranging data as the non-detection value. This makes it possible to set a value that is clearly distinguishable from the value of the actual measurement part detected by actual measurement (actual measurement value) as the non-detection value.

[0069] According to the tenth aspect of the present disclosure, in the data processing algorithm evaluation device according to the third aspect, the data generator 20 sets a predetermined non-detection value to data that does not have an actual measurement value set among the data constituting the radio ranging data. This makes it possible to standardize the size and format of the radio ranging data, and to easily apply a network such as KPConv.

[0070] According to an eleventh aspect of the present disclosure, in the data processing algorithm evaluation device according to the tenth aspect, the data generating unit 20 sets a value smaller than the smallest actual measurement value among the actual measurement values ​​set in the data in the radio wave ranging data as the non-detection value. This makes it possible to set a value that is clearly distinguishable from the value of the actual measurement part detected by actual measurement (actual measurement value) as the non-detection value.

[0071] The technical scope of the present invention is not limited to the above-described embodiment, and appropriate modifications can be made without departing from the spirit of the present invention. [Explanation of symbols]

[0072] 10 Data storage unit 20,51,53 Data Generation Section 21, 51a, 53a Data Understanding Department 21a,21b Arithmetic unit 22,51b Data Processing Department 30 Data Processing Section 40 Learning data storage unit 50, 50A, 50B, 50C Data learning section 52,54 Authenticity Determination Department 100 Data processing algorithm evaluation device D0 Non-detection part D1 Actual measurement part DT1 reflection cross section data DT2 Speed ​​Data P reference position R1, R2 target area

Claims

1. a data storage unit that stores radio wave ranging data detected from the vehicle; a data generating unit that, when disturbance information indicating a disturbance to target data among the radio wave ranging data stored in the data storage unit is input, acquires and understands the target data, and processes the target data based on the understanding to generate synthetic data so that the disturbance is reflected in the target data; a data processing unit that evaluates the performance of a data processing algorithm that determines the situation of the vehicle based on the generated synthetic data; Equipped with the data generation unit extracts a spatial arrangement of the objects constituting the target data from the detected positions of the target data, and generates the composite data based on the extraction result; the radio wave ranging data includes data arranged two-dimensionally regarding an azimuth angle at which a radio wave emitted from the vehicle and reflected by the object arrives and a distance to the object, The radio wave ranging data includes reflection cross-sectional area data indicating a reflection cross-sectional area of ​​the radio wave reflected by the object, and speed data indicating a relative speed between the vehicle and the object. Data processing algorithm evaluation device.

2. a learning data storage unit that stores, as learning data, reference data that is the radio wave ranging data that does not include the disturbance and disturbance data that is the radio wave ranging data that includes the disturbance; a data learning unit that generates synthetic data for training based on the reference data and the disturbance information by the same process as the process by which the data generation unit generates the synthetic data based on the target data and the disturbance information, and that performs learning using a generative adversarial network or a cycle generative adversarial network so as to increase the accuracy rate between the generated synthetic data for training and the disturbance data; Further provided with The data generation unit understands the target data based on the learning result of the data learning unit and generates the synthetic data.

2. The data processing algorithm evaluation device according to claim 1.

3. the learning data storage unit stores a plurality of disturbance data including the disturbances of the same type but different degrees; the data learning unit performs learning based on a plurality of disturbance data including the disturbances of different degrees, When the disturbance information including the degree of the disturbance is input, the data generation unit understands the target data based on a learning result of the data learning unit, and generates the synthetic data so that the disturbance of a degree corresponding to the disturbance information is reflected in the target data.

3. The data processing algorithm evaluation device according to claim 2.

4. the learning data storage unit stores a plurality of pieces of the radio wave ranging data including different attributes in association with label information indicating the attributes; the data learning unit learns using the cycle generative adversarial network so as to increase a rate of accuracy regarding the attribute between the generated synthetic data for learning and the disturbance data; When the target data and the disturbance information are input, the data generation unit extracts label information indicating the attributes of the target data, understands the target data based on a learning result of the data learning unit, and generates the synthetic data reflecting the attributes.

4. The data processing algorithm evaluation device according to claim 2 or 3.

5. The data learning unit sets a predetermined non-detection value to the data that does not have an actual measurement value set among the data that constitutes the radio wave ranging data.

4. The data processing algorithm evaluation device according to claim 2 or 3.

6. The data learning unit sets a value smaller than the smallest actual measurement value set in the radio wave ranging data as the non-detection value.

6. The data processing algorithm evaluation device according to claim 5.

7. The data generating unit sets a predetermined non-detection value to the data that does not have an actual measurement value set among the data that constitutes the radio wave ranging data.

2. The data processing algorithm evaluation device according to claim 1.

8. The data generating unit sets a value smaller than the smallest actual measurement value set in the radio wave ranging data as the non-detection value.

8. The data processing algorithm evaluation device according to claim 7.

Images