Three-dimensional data estimation device and data processing algorithm evaluation device

Abstract

A three-dimensional data estimation device (100) comprises: an acquisition unit that acquires, as input information, three-dimensional optical distance measurement data (D1) detected for a detection target; and an estimation unit that estimates three-dimensional radio wave distance measurement data (D2), which corresponds to the input information, by means of a learning model obtained by machine learning of a correspondence relationship between the three-dimensional optical distance measurement data detected for the detection target and three-dimensional radio wave distance measurement data detected for the detection target.

Description

Three-dimensional data estimation device and data processing algorithm evaluation device

[0001] The present disclosure relates to a three-dimensional data estimation device and a data processing algorithm evaluation device.

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

[0003] Various performance evaluations are conducted on image processing algorithms to realize appropriate driving assistance functions. For example, a technique is known for evaluating the performance of image processing algorithms by performing image processing using a composite image in which an image of weather disturbances created by computer graphics is superimposed on an actual image taken from a vehicle (see, for example, Patent Document 1).

[0004] JP 2010-033321 A

[0005] In the above-mentioned technology, it is required to accurately determine the state of the detection target, such as a vehicle.

[0006] The present disclosure has been made in view of the above, and aims to provide a three-dimensional data estimation device and a data processing algorithm evaluation device that are capable of accurately determining the state of a detection target.

[0007] The three-dimensional data estimation device according to the present disclosure includes an acquisition unit that acquires three-dimensional optical ranging data detected for a detection target as input information, and an estimation unit that estimates the three-dimensional radio wave ranging data corresponding to the input information using a learning model that has been machine-learned to determine the correspondence between the three-dimensional optical ranging data detected for the detection target and the three-dimensional radio wave ranging data detected for the detection target.

[0008] The data processing algorithm evaluation device according to the present disclosure includes a data storage unit that stores the radio ranging data estimated by the above-described three-dimensional data estimation device, a data generation unit that, when disturbance information indicating a disturbance to target data among the radio ranging data stored in the data storage unit is input, acquires and understands the target data, processes the target data based on the understanding so that the disturbance is reflected in the target data, and generates 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.

[0009] According to the present disclosure, the situation of the detection target can be determined with high accuracy.

[0010] FIG. 1 is a diagram schematically illustrating an example of a three-dimensional data estimation device according to this embodiment. FIG. 2 is a functional block diagram illustrating an example of a three-dimensional data estimation device according to this embodiment. FIG. 3 is a diagram schematically illustrating a learning model. FIG. 4A is a diagram schematically illustrating an example of the learning model. FIG. 4B is a diagram schematically illustrating a processing flow when K points are extracted using multiple KPConv layers. FIG. 4C is a diagram schematically illustrating an example of the learning model. FIG. 4D is a diagram schematically illustrating a processing flow when K points are extracted using multiple KPConv layers, a fully connected layer, and voxel feature sampling processing. FIG. 5A is a diagram schematically illustrating an example of the learning model. FIG. 5B is a diagram schematically illustrating a processing flow in the VFE layer. FIG. 6A is a diagram schematically illustrating an example of the learning model. FIG. 6B is a diagram schematically illustrating a processing flow in the voxel feature extraction layer. Fig. 7A is a diagram schematically showing an example of a learning model. Fig. 7B is a diagram schematically showing a processing flow in learning model 40D. Fig. 8 is a functional block diagram showing an example of a data processing algorithm evaluation device according to this embodiment.

[0011] Hereinafter, embodiments of a three-dimensional data estimation device and a data processing algorithm evaluation device according to the present disclosure will be described with reference to the accompanying drawings. Note that the present invention is not limited to these embodiments. Furthermore, the components in the following embodiments include those that are easily replaceable by those skilled in the art, or those that are substantially identical.

[0012] FIG. 1 is a diagram schematically illustrating an example of a three-dimensional data estimation device 100 according to this embodiment. FIG. 2 is a functional block diagram illustrating an example of a three-dimensional data estimation device 100 according to this embodiment. As shown in FIG. 2, the three-dimensional data estimation device 100 includes a processing unit 10, a storage unit 20, and a communication unit 30. When three-dimensional optical ranging data D1 detected for a detection target is input as input information, the three-dimensional data estimation device 100 estimates and outputs three-dimensional radio wave ranging data D2 corresponding to the input information.

[0013] The processing unit 10 performs various types of information processing. The processing unit 10 includes a processor such as a CPU (Central Processing Unit) and memories such as a ROM (Read Only Memory) and a RAM (Random Access Memory). The processing unit 10 includes an acquisition unit 11 and an estimation unit 12.

[0014] The acquisition unit 11 acquires, as input information, three-dimensional optical ranging data D1 detected for a detection target. The detection target includes, for example, a three-dimensional structure around a vehicle. The optical ranging data is measurement data obtained by an optical ranging method such as LiDAR (Laser Imaging Detection and Ranging). The optical ranging data is three-dimensional data, such as point cloud data obtained by emitting a beam of measurement light from a vehicle and receiving the reflected light of the measurement light at the detection target. The optical ranging data is acquired for each beam of measurement light emitted from the vehicle. The optical ranging data is data indicating three-dimensional position information (e.g., X coordinate, Y coordinate, and Z coordinate in an XYZ Cartesian coordinate system) of a point at which the measurement light is reflected and the reflection intensity of the reflected light.

[0015] The estimation unit 12 estimates three-dimensional radio wave ranging data corresponding to the input information (optical ranging data) acquired by the acquisition unit 11 based on the input information. For example, the data is equivalent to measurement data obtained by a radio wave ranging method such as RADAR (Radio Detecting and Ranging). That is, the radio wave ranging data is measurement data simulated by the three-dimensional data estimation device 100 so as to be equivalent to measurement data obtained by the radio wave ranging method. The radio wave ranging data is three-dimensional data such as point cloud data. The radio wave ranging data is data indicating three-dimensional position information of a target point (e.g., X coordinate, Y coordinate, and Z coordinate in an XYZ coordinate system based on a certain point) and an RCS value. The estimation unit 12 estimates the radio wave ranging data using a learning model 40 described below. The estimation unit 12 outputs the estimated radio wave ranging data.

[0016] The learning model 40 is a learning model that uses three-dimensional optical ranging data detected of a detection target and three-dimensional radio wave ranging data detected of the detection target as data sets and machine-learns the correspondence between the two. FIG. 3 is a diagram schematically illustrating an example of the learning model 40. As shown in FIG. 3 , in this embodiment, the learning model 40 is machine-learned to calculate local shape features and the reflection intensity of radio wave ranging data from the three-dimensional optical ranging data (processing block 41), extract a point cloud where the calculated reflection intensity is equal to or greater than a predetermined value (processing block 42), and output three-dimensional radio wave ranging data for the extracted point cloud as an estimated result. The reflection intensity depends on the shape of the reflective portion, and the reflection intensity tends to be higher in characteristic portions such as recesses, protrusions, and corners. Therefore, the local shape feature can be a value that indicates a characteristic shape, such as the degree of concavity of a recess, the degree of protrusion of a protrusion, or the angle of a corner of an object. The shape features and reflection intensities calculated in processing block 41 are information extracted from both the positional information and reflection intensity information of the point clouds of the optical ranging data to select the point clouds to be converted into radio wave ranging data. That is, the learning model 40 is trained in processing block 41 to extract the local shapes and reflection intensities required to select the point clouds to be converted into radio wave ranging data from both the positional information and reflection intensity information of the point clouds of the optical ranging data. The estimation unit 12 estimates the radio wave ranging data using the learning model 40 based on the optical ranging data acquired by the acquisition unit 11.

[0017] The learning model 40 in this embodiment refers to a learning model in AI (Artificial Intelligence). Specifically, the learning model 40 is a learning model learned by deep learning, and is composed of variables and a model (configuration information of the neural network) that defines a neural network that constitutes a classifier learned by deep learning.

[0018] 4A and 4C are diagrams schematically illustrating learning models. The learning model 40A shown in FIG. 4A illustrates an example in which a KPConv layer 51 is applied to a processing block 41 that calculates local shape features and reflection intensities of radio wave ranging data from three-dimensional optical ranging data. In the example shown in FIG. 4A, multiple KPConv layers 51 are provided. Each KPConv layer 51 performs a single process to reduce point cloud information. The multiple KPConv layers 51 extract K points (where K is, for example, a preset threshold) from the point cloud of the three-dimensional optical ranging data. FIG. 4B schematically illustrates the processing flow when extracting K points using multiple KPConv layers 51. In the example shown in FIG. 4A, a fully connected layer 52 is applied to a processing block 42 that calculates the reflection intensities of the extracted K points. The fully connected layer 52 extracts the reflection intensities of K points from the point cloud of the three-dimensional optical ranging data. That is, the learning model 40A is trained to extract the reflection intensity of a predetermined K point based on the point cloud of the three-dimensional optical ranging data. The learning model 40A is trained so that the coordinates and number of the point cloud of the radio wave ranging data output based on the input information approach or match the coordinates and number of the point cloud of the radio wave ranging data that is the training data. This learning model 40A can extend the convolution operation defined based on the grid to a discrete point cloud, allowing for appropriate extraction of local shape features.

[0019] Instead of the learning model 40A, for example, as shown in FIG. 4C , a learning model 40A2 may be used that is trained to calculate local shape features for a point cloud of 3D optical ranging data and reflection intensity of radio wave ranging data using the KP Conv layer 51 and fully connected layer 52. That is, in the learning model 40A2, the KP Conv layer 51 and fully connected layer 52 are used as the processing block 41. In this case, an example of the processing corresponding to the processing block 42 is a voxel feature sampling process 53 that extracts the top K points of reflection intensity based on the output of the learning model. Note that the voxel feature sampling process is not part of deep learning. FIG. 4D schematically illustrates the processing flow when extracting K points using multiple KP Conv layers 51, fully connected layers 52, and voxel feature sampling processes 53. In this configuration, the number of point clouds is not narrowed down to K points in the KPConv layer 51 and the fully connected layer 52, and the top K points are extracted by voxel feature sampling processing. Therefore, even if the value of K is changed, there is no need to retrain the learning model 40A.

[0020] FIG. 5A is a diagram schematically illustrating an example of a learning model. The learning model 40B illustrated in FIG. 5A illustrates an example in which a grouping layer 61, a random sampling layer 62, and multiple VFE (Voxel Feature Encoding) layers 63 are applied to a processing block 41 that calculates local shape features and reflection intensities of radio wave ranging data from three-dimensional optical ranging data. The multiple VFE layers 63 are arranged hierarchically. FIG. 5B is a diagram schematically illustrating the processing flow in the VFE layer 63. The VFE layer 63 includes a fully connected layer 63a, a MAX pooling layer 63b, and a Point-wise Concatenated Features layer 63c. The processing results of the multiple VFE layers 63 are output as the processing results of the processing block 41. Furthermore, processing block 42, which extracts point clouds whose calculated reflection intensities are equal to or greater than a predetermined value, is subjected to voxel feature sampling processing 67, which extracts K points with the highest intensities (K is, for example, a preset threshold value). Learning model 40B learns by using the Kullback-Leibler divergence as a loss value between the distribution of coordinates and numbers of point clouds of the radio wave ranging data output based on the input information and the distribution of coordinates and numbers of point clouds of the radio wave ranging data that are the learning data, so that the two distributions approach or match each other.

[0021] Fig. 6A is a diagram schematically showing an example of a learning model. The learning model 40C shown in Fig. 6A has a configuration in which a Voxel Feature Extraction layer 64 is applied instead of the VFE (Voxel Feature Encoding) layer 63 in the learning model 40B shown in Fig. 5A. Fig. 6B is a diagram schematically showing the processing flow in the Voxel Feature Extraction layer 64. The Voxel Feature Extraction layer 64 has a fully connected layer 64a, a MAX pooling layer 64b, and a Point-wise Concatenated Features layer 64c. Unlike the VFE layer 63, the Voxel Feature Extraction layer 64 preserves three-dimensional position information of the optical ranging data. In other words, the Voxel Feature Extraction layer 64 preserves the three-dimensional position information of the optical ranging data and performs processing using the fully connected layer 64a, the MAX pooling layer 64b, and the Point-wise Concatenated Features layer 64c. The remaining configuration is the same as that of the learning model 40B shown in FIG. 5A. This configuration uses the Voxel Feature Extraction layer 64 to preserve three-dimensional position information of the optical ranging data, allowing three-dimensional position information of the radio wave ranging data to be output, thereby enabling appropriate calculation of local shape features. In other words, this learning model 40B separates the position information and intensity information of the optical ranging data and allows them to be input and processed separately. This allows the location information of the radio wave ranging data to be estimated based on the location information of the optical ranging data, and the RCS value of the radio wave ranging data to be estimated based on the intensity information of the optical ranging data, thereby avoiding information loss and enabling appropriate extraction of local shape features.

[0022] FIG. 7A is a diagram schematically illustrating an example of a learning model. The learning model 40D shown in FIG. 7A illustrates an example in which a voxelization layer 71, a Gaussian approximation layer 72, an adjacent voxel integration layer 73, and multiple voxel feature extraction layers 74 are applied as a processing block 41 that calculates local shape features and reflection intensities of radio wave ranging data from three-dimensional optical ranging data. FIG. 7B is a diagram schematically illustrating the processing flow in the learning model 40D. The voxelization layer 71 voxels the point cloud data for each unit spatial lattice. The Gaussian approximation layer 72 performs Gaussian approximation, which is used in the 3D-NDT method, for each voxelized unit spatial lattice. The Gaussian approximation layer 72 calculates, for example, the center coordinates, variance-covariance matrix, and average reflection intensity of the point cloud for each unit spatial lattice. The adjacent voxel integration layer 73 integrates the unit spatial lattices that have undergone Gaussian approximation to calculate a feature vector. The multiple voxel feature extraction layers 74 perform processing similar to that of the VFE layer 63, while retaining position information based on the feature vectors. Furthermore, the processing block 42 that extracts point clouds with calculated reflection intensities equal to or greater than a predetermined value is subjected to a voxel feature sampling process 75 that extracts the K points with the highest intensities (where K is, for example, a preset threshold value). The learning model 40D learns to approximate or match the distribution of coordinates and numbers of point clouds of the radio wave ranging data output based on input information with the distribution of coordinates and numbers of point clouds of the radio wave ranging data used as training data, using the Kullback-Leibler divergence as a loss value. This learning model 40D enables appropriate extraction of local shape features.

[0023] 8 is a functional block diagram showing an example of a data processing algorithm evaluation device 200 according to this embodiment. The data processing algorithm evaluation device 200 shown in FIG. 8 includes a calculation device, i.e., a CPU, 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, a ROM, and an external storage device such as a HDD, for example. As shown in FIG. 8, the data processing algorithm evaluation device 200 includes a data storage unit 110, a data generation unit 120, and a data processing unit 130.

[0024] The data storage unit 110 stores the radio wave ranging data estimated and output by the three-dimensional data estimation device 100 according to this embodiment. That is, the radio wave ranging data is equivalent to measurement data obtained by a radio wave ranging method such as RADAR.

[0025] The data generation unit 120 generates composite data by combining object shape information that takes into account the shape of the object and is input from an input unit (not shown) or the like with the target data input from the data storage unit 110. The data generation unit 120 has a data understanding unit 121 and a data processing unit 122. The data understanding unit 121 understands the data stored in the data storage unit 110. The data processing unit 122 processes the target data input to the data generation unit 120.

[0026] The data processing unit 130 performs data processing based on the generated composite data and evaluates, for example, the performance of a data processing algorithm that determines the vehicle's status. The data processing unit 130 processes the composite data using the data processing algorithm and calculates determination information for determining the vehicle's status. The data processing unit 130 stores the calculated determination information in a storage unit (not shown). Examples of the determination information include 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 130 can evaluate the performance of the data processing algorithm based on whether or not the determination information differs significantly between when data that does not include a disturbance is processed and when data that includes a disturbance is processed.

[0027] As described above, the three-dimensional data estimation device according to the first aspect of the present disclosure includes an acquisition unit 11 that acquires three-dimensional optical ranging data detected for a detection target as input information, and an estimation unit 12 that estimates three-dimensional radio wave ranging data corresponding to the input information using a learning model 40 that has machine-learned the correspondence between the three-dimensional optical ranging data detected for the detection target and the three-dimensional radio wave ranging data detected for the detection target.

[0028] With this configuration, the learning model 40, which has machine-learned the correspondence between the three-dimensional optical ranging data detected about the detection target and the three-dimensional radio wave ranging data detected about the detection target, can estimate three-dimensional radio wave ranging data corresponding to the three-dimensional optical ranging data, which is the input information, thereby making it possible to accurately determine the situation of the detection target.

[0029] In the three-dimensional data estimation device according to the second aspect of the present disclosure, in the three-dimensional data estimation device according to the first aspect, the learning model 40 is machine-learned to calculate local shape features from three-dimensional optical ranging data, calculate the reflection intensity of the radio wave ranging data based on the calculated shape features, extract a point cloud where the calculated reflection intensity is equal to or greater than a predetermined value, and output three-dimensional radio wave ranging data for the extracted point cloud as an estimation result.

[0030] According to this configuration, the learning model 40 calculates local shape features from three-dimensional optical ranging data, calculates the reflection intensity of the radio wave ranging data based on the calculated shape features, extracts point groups where the calculated reflection intensity is above a predetermined level, and outputs three-dimensional radio wave ranging data for the extracted point groups as an estimated result, thereby making it possible to accurately determine the situation of the target to be detected.

[0031] In the three-dimensional data estimation device according to the third aspect of the present disclosure, the radio wave ranging data estimated by the estimation unit 12 is data indicating three-dimensional position information and an RCS value of a target point in the point cloud.

[0032] According to this configuration, it is possible to appropriately estimate three-dimensional data indicating three-dimensional position information and RCS values.

[0033] In a three-dimensional data estimation device according to a fourth aspect of the present disclosure, in the three-dimensional data estimation device according to any one of the first to third aspects, learning model 40A is machine-learned using KPConv layer 51 to calculate shape features and reflection intensity of radio wave ranging data from three-dimensional optical ranging data.

[0034] According to this configuration, the convolution operation defined based on the grid can be extended to a discrete point group, so that local shape features can be appropriately extracted.

[0035] In the three-dimensional data estimation device according to the fifth aspect of the present disclosure, in the three-dimensional data estimation device according to the fourth aspect, the learning model 40A is machine-learned using a fully connected layer 52 to extract a point cloud having a reflection intensity equal to or greater than a predetermined value.

[0036] According to this configuration, a point cloud can be appropriately extracted based on the content previously learned in the learning model 40A.

[0037] In a three-dimensional data estimation device according to a sixth aspect of the present disclosure, in the three-dimensional data estimation device according to any one of the first to third aspects, the learning model 40B is machine-trained to calculate shape features from three-dimensional optical ranging data using the voxel feature encoding layer 63.

[0038] According to this configuration, by using the VFE (Voxel Feature Encoding) layer 63, it is possible to appropriately calculate the shape feature amount and the reflection intensity of the radio wave ranging data.

[0039] In a three-dimensional data estimation device according to a seventh aspect of the present disclosure, in the three-dimensional data estimation device according to any one of the first to third aspects, learning model 40C is machine-learned using voxel feature extraction layer 64 to calculate shape features and reflection intensity of radio wave ranging data from the three-dimensional optical ranging data while retaining position information of the optical ranging data.

[0040] With this configuration, by using the Voxel Feature Extraction layer 64, the position information and intensity information of the optical ranging data can be separated and input and processed separately. As a result, the position information of the radio wave ranging data is estimated based on the position information of the optical ranging data, and the RCS value of the radio wave ranging data is estimated based on the intensity information of the optical ranging data. This makes it possible to avoid information loss and appropriately extract local shape features and the reflection intensity of the radio wave ranging data.

[0041] In a three-dimensional data estimation device according to an eighth aspect of the present disclosure, in the three-dimensional data estimation device according to any one of the first to third aspects, learning model 40D is machine-trained using voxel feature extraction layer 74 to calculate shape features and reflection intensity of radio wave ranging data from the results of Gaussian approximation used in the 3D-NDT method performed on three-dimensional optical ranging data.

[0042] According to this configuration, it is possible to calculate the shape feature amount and the reflection intensity of the radio wave ranging data that reflect the Gaussian approximation that is performed.

[0043] The data processing algorithm evaluation device 200 according to a ninth aspect of the present disclosure includes a data storage unit 110 that stores radio wave ranging data estimated by the three-dimensional data estimation device 100 according to any one of the first to seventh aspects, a data generation unit 120 that, when disturbance information indicating a disturbance to target data among the radio wave ranging data stored in the data storage unit 110 is input, acquires and understands the target data, processes the target data based on the understanding so that the disturbance is reflected in the target data, and generates synthetic data, and a data processing unit 130 that evaluates the performance of a data processing algorithm that determines the vehicle situation based on the generated synthetic data.

[0044] According to this configuration, the data generator 120 processes the target data based on the results of its understanding of the target data so that the object shape is reflected in the target data to generate synthetic data. This allows for the generation of synthetic data that is closer to the actual surrounding environment than, for example, when simply superimposing disturbances generated by a simulator. Furthermore, because the target data is radio ranging data estimated by the three-dimensional data estimation device 100, the synthetic data can be generated with high accuracy. This allows for appropriate evaluation of the performance of the data processing algorithm for determining the vehicle's situation.

[0045] The technical scope of the present invention is not limited to the above-described embodiments, and appropriate modifications can be made without departing from the spirit of the present invention.

[0046] 10 Processing unit 11 Acquisition unit 12 Estimation unit 20 Memory unit 30 Communication unit 40, 40A, 40B, 40C, 40D Learning model 41, 42 Processing block 51 KPConv layer 52 Fully connected layer 53, 67, 75 Voxel Feature Sampling process 61 Grouping layer 62 Random sampling layer 63 Voxel Feature Encoding layer 64, 74 Voxel Feature Extraction layer 71 Voxelization layer 72 Gaussian approximation layer 73 Adjacent voxel integration layer 100 Three-dimensional data estimation device 110 Data memory unit 120 Data generation unit 121 Data understanding unit 122 Data processing unit 130 Data processing unit 200 Data processing algorithm evaluation device D1 Optical ranging data D2 Radio wave ranging data

Claims

1. An acquisition unit that acquires, as input information, three-dimensional optical ranging data detected for a detection target, and an estimation unit that estimates three-dimensional radio ranging data corresponding to the input information by a learning model that has learned a correspondence relationship between the three-dimensional optical ranging data detected for the detection target and the three-dimensional radio ranging data detected for the detection target. A three-dimensional data estimation device comprising:

2. The three-dimensional data estimation device according to claim 1, wherein the learning model calculates local shape feature amounts from the three-dimensional optical ranging data, calculates the reflection intensity of the radio ranging data based on the calculated shape feature amounts, extracts a point group in which the calculated reflection intensity is equal to or greater than a predetermined value, and outputs, as an estimation result, three-dimensional radio ranging data for the extracted point group by machine learning.

3. The three-dimensional data estimation device according to claim 2, wherein the radio ranging data estimated by the estimation unit is data indicating three-dimensional position information and an RCS value at a target point among the point group.

4. The three-dimensional data estimation device according to claim 2, wherein the learning model calculates the shape feature amounts and the reflection intensity of the radio ranging data from the three-dimensional optical ranging data by using a KPConv layer by machine learning.

5. The three-dimensional data estimation device according to claim 4, wherein the learning model extracts a point group in which the reflection intensity is equal to or greater than a predetermined value by using a fully connected layer by machine learning.

6. The three-dimensional data estimation device according to claim 2, wherein the learning model calculates the shape feature amounts and the reflection intensity of the radio ranging data from the three-dimensional optical ranging data by using a Voxel Feature Encoding layer by machine learning.

7. The three-dimensional data estimation device according to claim 2, wherein the learning model calculates the shape feature amounts and the reflection intensity of the radio ranging data from the three-dimensional optical ranging data while leaving the position information of the optical ranging data by using a Voxel Feature Extraction layer by machine learning.

8. The learning model is machine-learned to calculate the shape feature amount and the reflection intensity of the radio wave ranging data from the result of performing Gaussian approximation used in the 3D-NDT method on the three-dimensional optical ranging data using a Voxel Feature Extraction layer. The three-dimensional data estimation device according to claim 2.

9. A data storage unit that stores the radio wave ranging data estimated by the three-dimensional data estimation device according to any one of claims 1 to 8; when disturbance information indicating a disturbance to target data is input among the radio wave ranging data stored in the data storage unit, the target data is acquired and understood, and a data generation unit that generates synthetic data by processing the target data based on the understanding so that the disturbance is reflected in the target data; and a data processing unit that evaluates the performance of a data processing algorithm for determining the situation of a vehicle based on the generated synthetic data. A data processing algorithm evaluation device comprising:

Images