JP6026959B2 - Driving scene label estimation device - Google Patents

Description

  The present invention relates to a driving scene label estimation apparatus that estimates a label to be applied to a driving scene.

  In recent years, various driving support systems have been researched, developed, and commercialized, but a new driving support system that can handle a wider range of situations if various driving support systems are used in appropriate combinations according to the situation. Realization of is expected. In order to appropriately propose a driving support system to be used to the driver, a technique for estimating the situation (driving scene) of the host vehicle is required. In such a technique, it is important not only to appropriately divide the driving scene, but also to clarify what situation the divided driving scene represents.

  As one of the devices that realize this kind of technology, data (vehicle speed, vehicle position, etc.) that represents the vehicle state is continuously monitored, and those data are stored in the registered driving scenes (for example, “Waiting for signal”, “Sudden stop”). Is known, and a current driving scene is determined based on the matching rate (see, for example, Patent Document 1).

JP 2011-129010 A

  By the way, in the above-mentioned conventional apparatus, since it is used to determine which of the predefined driving scenes the current situation is similar to, the design of what kind of situation the driving scene to be identified represents. Need to be clearly defined in advance.

  However, since it is difficult to pre-define and define every driving scene from the start to the end of driving, the problem is that it can only be used to detect a limited specific driving scene. was there.

  In addition, although a method for dividing a driving scene without a teacher that does not require a designer's driving scene definition has been proposed, in this case, although it is not necessary to define a driving scene in advance, an extracted driving scene is not required. There is a problem that it is difficult to assign a scene label that explains what kind of situation is represented.

  In order to solve the above-described problems, an object of the present invention is to provide a driving scene label estimation apparatus that generates a scene label that matches a divided driving scene.

  The driving scene label estimation apparatus of the present invention includes data collection means, scene division means, feature quantity distribution generation means, topic storage means, ratio calculation means, estimation distribution setting means, and label estimation means. .

  The data collecting means repeatedly collects data relating to the state of the vehicle, and the scene dividing means divides the data series collected by the data collecting means into a plurality of partial series each representing some driving scene. The feature quantity distribution generation means, for one or more types of feature quantities obtained from the data collected by the data collection means, is a feature quantity representing an appearance frequency for each index, with an index obtained by dividing a range of the feature quantity into a plurality of features. A distribution is generated for each partial series. The topic storage means uses a plurality of base feature distributions used as a base distribution for expressing the feature amount distribution, and a plurality of tags prepared in advance to explain the driving scene as indexes, and appears for each index. A driving topic including a base tag distribution representing a probability is stored. The ratio calculating means obtains a topic ratio that is a mixing ratio of driving topics when the feature quantity distribution generated by the feature quantity distribution generating means is expressed using driving topics. The estimation distribution setting means sets the label estimation distribution for each partial series based on the topic ratio and the base tag distribution, and the label estimation means uses the set label estimation distribution to perform the operation represented by the partial series. Estimate a scene label suitable for the scene.

  According to such a configuration, using the driving topic, a label estimation distribution corresponding to the driving scene is generated for each driving scene, and a scene label suitable for the driving scene is generated using the label estimation distribution. In order to estimate, there is no need to pre-define the judgment conditions for each driving scene or the correspondence between each driving scene and the scene label, and the estimated scene labels are automatically assigned to all the divided driving scenes. can do.

  As a result, for example, the driver can intuitively understand what the current driving scene is by checking the scene label assigned to each driving scene. In addition, since a scene label suitable for the driving scene is given throughout the driving behavior including a driving scene that is not assumed in advance, the driving assistance is interactively proposed to the driver using the scene label. It can be applied to a device, or a series of drives (ie, time series of driving scenes) can be left as a log expressed in a natural language using scene labels.

It is a whole block diagram of a driving scene label estimation apparatus. It is a graphical model assumed by Latent Dirichlet Allocation (LDA). It is explanatory drawing which shows the production | generation process of a driving | running topic, and the calculation process of the topic ratio using a driving | running topic. It is explanatory drawing which shows the setting method of the distribution for estimation. It is explanatory drawing which shows the estimation method of the scene label using distribution for estimation. It is explanatory drawing which shows the example of the label estimated by the driving scene label estimation apparatus, the image corresponding to the label, the driving behavior data produced | generated in the process of estimating a label, a partial series, and a topic ratio. It is a graph which illustrates the feature-value distribution produced | generated by the driving | running feature calculation part in 2nd Embodiment. It is explanatory drawing which illustrated distribution regarding driving behavior data among base feature distribution. It is explanatory drawing which illustrated the base label extracted from the distribution which converted the appearance frequency in a base feature distribution into likelihood ratio, and the distribution. It is explanatory drawing which shows the example of application of the scene label based on a base label. It is explanatory drawing which shows the difference of an effect when a driving | running scene is divided | segmented with a different method.

Embodiments of the present invention will be described below with reference to the drawings.
[First Embodiment]
<Overall configuration>
As shown in FIG. 1, the driving scene label estimation apparatus 1 includes a driving behavior data collection unit 2, a driving scene discretization unit 3, a feature amount distribution generation unit 4, a driving topic database 5, and a topic ratio calculation unit 6. And an estimation distribution setting unit 7 and a scene label estimation unit 8.

  The driving scene label estimation apparatus 1 is realized by a microcomputer (microcomputer), and each unit except the driving topic database 5 is realized by a CPU (not shown) included in the microcomputer executing a predetermined program. That is, FIG. 1 shows various functions realized by the CPU divided into functional blocks. However, these units do not necessarily have to be realized by software, and all or a part thereof may be realized by hardware such as a logic circuit.

<Driving behavior data collection unit>
The driving behavior data collection unit 2 repeatedly collects driving data related to driving operations by the driver and behavior data related to vehicle behaviors appearing as a result of the driving operations via various sensors mounted on the vehicle. And differential data (dynamic features) obtained by differentiating the behavior data and the multi-dimensional data composed of the operation data, behavior data, and differential data are output as the operation behavior data.

  As the driving data, for example, an accelerator pedal operation amount, a brake pedal operation amount, a steering wheel operation amount (steering angle), a direction indicator operation state, a transmission shift position, and the like can be used. Moreover, as behavior data, the speed of a vehicle, a yaw rate, etc. can be used, for example.

<Operation scene discretization unit>
The driving scene discretization unit 3 statistically analyzes the driving behavior data obtained from the driving behavior data collection unit 2 using a model from the driver's environmental recognition to the operation, and the driving felt by the individual driver (or a general driver). By extracting scene change points, the time series of driving behavior data is segmented (discretized) into a plurality of partial series each representing some driving scene.

Specifically, a double segment analyzer (DAA: Double Articulation Analyzer) that performs segmentation by an unsupervised driving scene division method using a double segment structure is used.
DAA first defines in advance a cluster representing various vehicle states grasped from driving behavior data in a multidimensional space representing a range of driving behavior data, and transition probabilities between the clusters. By using the information to statistically determine which cluster the driving behavior data acquired from the driving behavior data collection unit 2 belongs to, the time series of the driving behavior data is obtained for each vehicle state (unit: That is, it is classified into each cluster). However, the time series of the driving behavior data is converted into a symbol string indicating which cluster belongs by associating each cluster with a symbol for identification in advance. For generation of this symbol string, for example, a hierarchical Dirichlet process hidden Markov model (HDP-HMM) which is one of models expressed by a hidden state and a stochastic transition between the states can be used.

  Next, DAA uses a Nested Pitman-Yor Language Model (NPYLM), which is an example of an unsupervised chunking technique for discrete character strings using statistical information, to generate a predetermined operation scene. Segment into meaningful subsequences. At this time, the dictionary size (that is, the number of partial sequences) is made as small as possible, and the generation probability of the entire symbol string composed of the partial sequences is maximized. This makes it possible to segment the driving behavior data into driving scenes. However, transition probabilities between partial sequences and partial sequence generation probabilities are generated in advance by learning.

  As for DAA to which HDP-HMM and NPYLM are applied, non-patent literature separately disclosed by the present applicant, T. Taniguchi et al, “Semiotic Prediction of Driving Behavior using Unsupervised Double Articulation Analyzer” IEEE Intelligent Vehicles Symposium, 2012, K. Takenaka et al., “Contextual Scene Segmentation of Driving Behavior based on Double Articulation Analyzer” IEEE / RSJ International Conference on Intelligent Robots and Systems, 2012, etc., will not be described here. However, the method used for symbol generation and symbol segmentation is not limited to HDP-HMM or NYLM, and other methods may be used.

<Feature distribution generation unit>
The feature amount distribution generation unit 4 includes a driving feature calculation unit 41 and a moving image feature calculation unit 42.
The driving feature calculation unit 41 is based on the time-series driving behavior data output from the driving behavior data collection unit 2 for each partial series segmented by the driving scene discretization unit 3 (that is, any driving scene). Then, a distribution of feature quantities appearing in the partial series is generated. Here, the accelerator pedal operation amount, the brake pedal operation amount, the steering operation amount, the vehicle speed, and eight-dimensional data composed of differential data of these are set as the generation targets of the feature amount distribution. The feature amount distribution includes a histogram that expresses the appearance frequency of the 8D data for each index obtained by dividing the feature space representing the range of the 8D data into a plurality of indexes.

  The moving image feature calculation unit 42 is a portion obtained by segmenting image data output from the in-vehicle camera 11 installed so as to photograph the vehicle periphery such as the front, side, and rear of the vehicle by the driving scene discretization unit 3. Segmentation is performed according to the series, and a distribution of feature quantities appearing in the partial series of the image data is generated. Here, the input image is compressed to a predetermined size (for example, 320 × 240 pixels), and the compressed image data has a plurality (for example, 24 × 24 pixels) having a predetermined size (for example, 24 × 24 pixels) at regular intervals (for example, 5 pixels). In this example, 3072 points) are divided into unit areas, and SIFT (Scale-Invariant Feature Transform) feature amounts obtained for the respective unit regions are used as generation targets of the feature amount distribution. The feature amount distribution includes a histogram that expresses the frequency of appearance of SIFT feature amounts for each index obtained by dividing a feature space representing a range of SIFT feature amounts into a plurality of indexes.

  Note that the index of the feature distribution is determined in advance by learning. For example, the data for which the feature quantity distribution is to be generated (the above-mentioned eight-dimensional data, SIFT feature quantity) is clustered using a clustering technique such as k-means, and each cluster obtained as a result is converted into the feature quantity distribution. Used as an index of

  Hereinafter, the feature amount distribution generated by the driving feature calculation unit 41 is referred to as a driving feature distribution, the feature amount distribution generated by the moving image feature calculation unit is referred to as a moving image feature distribution, and a combination thereof is referred to as an analysis target distribution. . That is, the feature amount distribution generation unit 4 generates an analysis target distribution (driving feature distribution and moving image feature distribution) for each partial series corresponding to the driving scene, and outputs it to the subsequent stage.

<Operation topic database>
The driving topic database 5 includes a plurality of (for example, 100) base feature distributions used when the analysis target distribution generated by the feature amount distribution generation unit 4 is expressed by a mixture of a plurality of distributions, and each base feature distribution. The driving topic including the base tag distribution associated with is stored.

Here, a method for generating a driving topic (basic feature distribution, base tag distribution) will be described.
Driving topics are generated by considering each subsequence corresponding to one of the driving scenes as “one document” and the observed features (analysis distribution etc.) as “one word”. Use the topic estimation method used in the processing field.

  In particular, here, a driving topic that spans three modalities including driving behavior data (driving feature distribution), image data (moving feature distribution), and a tag (tag distribution) attached to the image represented by the image data in a natural language is estimated. Therefore, multimodal LDA, which is a method that is an extension of Latent Dirichlet Allocation (LDA), is used.

In LDA, a graphical model as shown in FIG. 2 is assumed. Where D is the total number of scenes, M _d is the total number of frames in the d-th scene, K is the total number of topics, and w _{d and m} are the features observed in the m-th frame of the d-th scene (ie, analysis target distribution, tag distribution). ), Z _{d, m} is a topic indicating variable indicating a topic assigned to a feature (takes a natural number from 1 to K), θ _d is a multinomial distribution parameter indicating the proportion of topics included in the d-th scene, φ _k is a multinomial distribution parameter (i.e., base feature distribution, base tag distribution) indicating the ratio of each feature amount generated from the k-th topic, and α and β are parameters of θ and φ, respectively. Further, θ _d , φ _k , z _{d, m} w _{d, m} are defined by the equations (1) to (4), where Dir is a Dirichlet distribution and Multi is a multinomial distribution.

Θ and φ are estimated using this model, and approximation methods such as variational Bayes and Gibbs sampling can be used for these estimations. For details of these methods, see, for example, D. Blei et al, “Latent Dirichlet Allocation” Journal of Machine Learning Research, 2003, and T. Griffiths & M. Steyvers, “Finding Scientific Topics,” Proceedings of the National Academy. of Sciences, 2004, etc., and thus the description thereof is omitted here.

However, the analysis target distribution (driving feature distribution, moving image feature distribution) and tag distribution used for generating the driving topic are generated as follows, for example.
First, by actually running the vehicle and taking an image for learning with the in-vehicle camera 11, by operating the driving behavior data collection unit 2, the driving scene discretization unit 3, and the feature amount distribution generation unit 4, A learning analysis target distribution corresponding to the learning image is generated.

  Next, after instructing a plurality of subjects only on rough tag items (for example, the presence / absence and type of a preceding vehicle, the presence / absence of traffic jams, the number of lanes, etc.), the learning image is shown for each frame of the image data. Then, tagging is performed, and a result of the tagging is collected into a histogram for each driving scene divided by the driving scene discretization unit 3, thereby generating a tag distribution for learning. Since the tag given in this way varies depending on the subject's judgment on the timing of tag assignment (for example, how close the vehicle is considered as the preceding vehicle), the probability of tag assignment is required. .

The result of estimating the driving topic by the LDA using the analysis target distribution and the tag distribution prepared in this way is the base feature distribution and the base tag distribution.
As shown in the column “Drive Topic Estimation” in FIG. 3, at the time of learning, the analysis target distribution and the tag for each partial series (driving scene) generated by segmenting the driving feature data based on the driving feature data. Generate a distribution. Based on the generated analysis target distribution and tag distribution, a base feature distribution and a base tag distribution are generated using LDA.

<Topic ratio calculation unit>
In the topic ratio calculation unit 6, the analysis target distribution output from the feature amount distribution generation unit 4 is expressed by mixing the base feature distribution stored in the driving topic database 5. The ratio (content ratio of each driving topic with respect to the analysis target distribution) is calculated (see the column “Drive Annotation” in FIG. 3). Specifically, the topic ratio can be obtained by executing only the E step among the processes of the E step and the M step for realizing the multimodal LDA.

<Estimation distribution setting section>
As shown in FIG. 4A, the estimation distribution setting unit 7 sets, as the estimation distribution, a mixed distribution generated by superimposing the base tag distributions according to the topic ratio obtained by the topic ratio calculation unit 6. . However, the estimation distribution setting method is not limited to this. For example, as shown in FIG. 4B, the base tag distribution associated with the driving topic having the largest topic ratio is set as the estimation distribution. May be.

<Scene label estimation unit>
As shown in FIG. 5A, the scene label estimation unit 8 is based on the estimation distribution set by the estimation distribution setting unit 7, and tags (label estimation distribution distribution) whose appearance frequency exceeds a preset threshold value. Index) is extracted, and the extracted tag is used as the estimation result of the scene label to be given to the driving scene. Note that the tag extraction method is not limited to this, and, for example, a part having a high appearance frequency to a predetermined number may be extracted. Further, as shown in FIG. 5B, the tag may be extracted by generating a tag probabilistically according to the label estimation distribution.

<Effect>
As described above, in the driving scene label estimation apparatus 1, the time series of driving behavior data is divided by DAA into partial series each representing some driving scene, and driving behavior data and image data are divided for each partial series. The analysis target distribution representing the characteristics of is generated. Then, a label estimation distribution is generated according to the mixing ratio (topic ratio) of the driving topics obtained for the analysis target distribution and the base tag distribution prepared in advance, and the label estimation distribution is used to adapt to the driving scene. The scene label is estimated.

  As described above, according to the driving scene label estimation apparatus 1, the driving scene is divided by the unsupervised division method, and the scene label is estimated using the label estimation distribution generated each time according to the driving scene. Therefore, it is not necessary to predefine the judgment conditions of each driving scene and the correspondence between each driving scene and the scene label, and all the divided driving including the driving scene that is difficult to define in advance. A scene label can be automatically assigned to a scene.

  Further, according to the driving scene label estimation apparatus 1, since the driving scene is extracted in a unit suitable for the driver's intuition by using DAA for dividing the driving scene, the driver's sense for the driving scene is extracted. It is possible to attach a scene label suitable for.

  FIG. 6 shows scene behaviors estimated for several driving scenes extracted by the driving scene label estimation apparatus 1, and driving behavior data generated in the process of estimating the labels corresponding to the scene labels. It is an example with series and topic ratio. From the figure, it can be seen that a scene label that is generally suitable for the image is extracted. The scene label indicates the case where the number of driving topics is 10 and the case where the driving topics are 100.

[Second Embodiment]
A second embodiment will be described.
In the present embodiment, the feature quantity distribution generated by the driving feature calculation unit 41, the data stored in the driving topic database 5, and the processing in the scene label estimation unit 8 are partially different from those in the first embodiment. Therefore, these points will be mainly described.

<Driving feature calculation unit>
In the driving feature calculation unit 41 of the present embodiment, eight types of data constituting the driving behavior data (accelerator pedal operation amount accel, brake pedal operation amount brake, steering operation amount steering, vehicle speed velocity, and their differential data) Instead of generating a feature quantity distribution of 8-dimensional data composed of Δaccel, Δbrake, Δsteering, and Δvelocity), a feature quantity distribution (driving feature distribution) is generated for each of the eight types of data. That is, the analysis target distribution output from the driving feature calculation unit 41 includes eight driving feature distributions and one moving image feature distribution.

  As shown in FIG. 7, each feature amount distribution is composed of a histogram expressing the appearance frequency for each bin obtained by dividing a feature value range (minimum value to maximum value) into 20 equal parts. In the figure, (a) shows the vehicle speed V, and (b) shows the characteristic amount distribution of the differential value (acceleration) ΔV of the vehicle speed.

<Operation topic database>
The driving topic database 5 stores a base label associated with the base feature distribution in addition to the base feature distribution and the base tag distribution.

The generation of the base feature distribution and the base tag distribution is performed in the same manner as in the first embodiment, except that the number of feature amount distributions constituting the analysis target distribution used for generation is different.
FIG. 8 illustrates a distribution related to driving behavior data among the generated base feature distributions. Here, for five types of driving topics TP1 to TP5, the appearance frequency is represented by a tone for each bin of the feature amount, and the whiter the tone, the higher the appearance frequency.

  The base feature distribution relating to the driving behavior generated in this manner represents the driving operation and vehicle behavior characteristic of each driving topic. For example, it can be seen that the driving topic TP1 frequently performs a steering operation in a low speed region as compared with other driving topics.

  The base label expresses a characteristic with respect to the specified feature distribution that can be read from the base feature distribution. As the base label, the situation that can be read from the base feature distribution as described above may be used as the base label as it is, but in the present embodiment, the base label is set as follows.

First, based on the basis feature distribution of each driving topic, the frequency of the basis feature distribution is converted into a likelihood ratio in accordance with the equation (5) in order to extract features peculiar to the driving topic. Note that i is a feature quantity identifier, k is a driving topic identifier, w _i is a feature quantity identified by i, and φ _k is a multinomial distribution of feature quantities constituting a driving topic identified by k (basic feature distribution). It is. The likelihood ratio L _{i, k} is an index called so-called inverse document frequency (IDF) indicating the degree to which a certain feature is not included in another driving topic but is included in a specific driving topic.

FIG. 9 is a feature emphasis distribution that displays only those with a likelihood ratio of a certain value (for example, 5) or more for each driving topic. Here, only the accelerator, brake, steering, and speed are shown for easy understanding of the drawing. Yes.

  Base labels are extracted for each driving topic from this feature enhancement distribution. For example, in the case of speed, 20 bins are divided into five stages [BIN1] [BIN2-5] [BIN6-10] [BIN11-15] [BIN16-20], and labels “v0” and “v” are respectively provided. Assume that “v +”, “v ++”, and “v ++” are associated with each other, and a label associated with a bin having the maximum likelihood ratio in each driving topic is set as a base label. Base labels are similarly set for other feature amounts. In the figure, “s ++ / −” is one of the labels regarding the steering operation amount, “a ++” is one of the labels regarding the operation amount of the accelerator pedal, and “b ++++” is one of the labels regarding the operation amount of the brake pedal. .

In the driving topic TP1 shown in FIG. 9, remarkable features are observed in the steering operation amount and the speed, and two labels “v” and “s +/−” corresponding thereto are given.
<Scene label estimation unit>
In the scene label estimation unit 8, in addition to the scene label estimated from the distribution for estimation as in the first embodiment, the base label associated with the driving topic having the maximum topic ratio is used as the scene of the target driving scene. Give it as a label.

<Effect>
According to the present embodiment, not only the same effects as in the first embodiment can be obtained, but also a label representing driving characteristics is used. Therefore, using this, for example, as shown in FIG. The situation seen from the viewpoint of can be recorded in a natural language.

  FIG. 11 shows experimental results showing the effect of performing driving scene extraction using DAA. As a comparative example, the driving topic (likelihood) generated when the number of driving scenes is the same as the number of divisions in the driving scene discretization unit 3 and the lengths of the driving scenes are all equalized. And the base label extracted from the driving topic.

  In the comparative example (FL), compared to the present embodiment (DAA), a characteristic distribution with respect to a certain characteristic amount tends to exist sparsely across a plurality of driving topics. It can be seen that it is difficult to assign a scene label suitable for the. However, the present invention does not deny dividing the driving scene by the method of the comparative example, and may be applied to such an apparatus.

[Other Embodiments]
As mentioned above, although embodiment of this invention was described, it cannot be overemphasized that this invention can take a various form, without being limited to the said embodiment. For example, the functions of one component may be distributed to a plurality of components, or the functions of a plurality of components may be integrated into one component. Further, at least a part of the configuration of the above embodiment may be replaced with a known configuration having the same function.

  For example, in the above embodiment, driving behavior data and image data are used as the generation target of the feature amount distribution. However, the present invention is not limited to this. For example, a global positioning system (GPS) or inertial measurement is used. You may use the position data regarding the position of the own vehicle, etc. which are acquired by the apparatus (Inertial Measurement Unit: IMU).

  In addition, the code | symbol in the parenthesis described in the claim shows the correspondence with the specific means as described in the embodiment as one aspect, and does not limit the technical scope of the present invention. .

  In addition to the driving scene label estimation device described above, the present invention also provides various systems including the driving scene label estimation device as components, a program for causing a computer to function as each means constituting the driving scene label estimation device, and a driving scene. It can be realized in various forms such as a label estimation method.

  DESCRIPTION OF SYMBOLS 1 ... Driving scene label estimation apparatus 2 ... Driving behavior data collection part 3 ... Driving scene discretization part 4 ... Feature-value distribution generation part 5 ... Driving topic database 6 ... Topic ratio calculation part 7 ... Distribution setting part 8 for estimation 8 ... Scene label Estimating unit 11 ... In-vehicle camera 41 ... Driving feature calculating unit 42 ... Moving image feature calculating unit

Claims (10)

Data collection means (2, 11) for repeatedly collecting data relating to the state of the vehicle;
Scene dividing means (3) for dividing the series of data collected by the data collecting means into a plurality of partial series each representing some driving scene;
A feature that generates, for each partial series, a feature quantity distribution representing an appearance frequency for each index obtained by dividing a range of the feature quantity into a plurality of feature quantities obtained from data collected by the data collection unit. A quantity distribution generating means (4);
A plurality of base feature distributions used as a base distribution for expressing the feature quantity distribution, and a base tag distribution representing an appearance probability for each of a plurality of tags prepared in advance for explaining the driving scene. Topic storage means (5) for storing driving topics;
A ratio calculation means (6) for calculating a topic ratio that is a mixture ratio when the feature quantity distribution generated by the distribution generation means is expressed using the base feature distribution;
For each partial series, an estimation distribution setting means (7) for setting a label estimation distribution based on the topic ratio and the base tag distribution;
Label estimation means (8) for estimating a scene label suitable for the driving scene represented by the partial series using the label estimation distribution;
A driving scene label estimation apparatus comprising:   2. The driving scene label estimation according to claim 1, wherein the estimation distribution setting unit sets a mixed distribution generated by mixing the base tag distribution according to the topic ratio as the label estimation distribution. apparatus.   The driving scene label estimation apparatus according to claim 1, wherein the estimation distribution setting unit sets a base tag distribution of the driving topic having the maximum topic ratio as the label estimation distribution.   4. The label estimation unit according to claim 1, wherein, based on the label estimation distribution, the tag whose appearance frequency is equal to or higher than a preset threshold is used as the scene label. 5. The driving scene label estimation device according to 1.   The driving scene according to any one of claims 1 to 3, wherein the label estimation means uses the tag that is probabilistically generated using the label estimation distribution as the scene label. Label estimation device.   6. The scene dividing means comprises a double segment analyzer (DAA: Double Articulation Analyzer) that performs segmentation by an unsupervised driving scene segmentation method using a double segment structure. The driving scene label estimation apparatus according to any one of the above.   The operation according to any one of claims 1 to 6, wherein the ratio calculation means obtains the topic ratio by a latent dirichlet allocation method (LDA) or a multimodal LDA. Scene label estimation device.   8. The feature amount distribution according to claim 1, wherein the feature amount distribution is expressed by a histogram using, as an index, each value range obtained by equally dividing a maximum value to a minimum value of the feature amount. The driving scene label estimation device described.   The driving scene label according to any one of claims 1 to 7, wherein the feature quantity distribution is expressed by a histogram using a range obtained by clustering the feature quantities as an index. Estimating device.   The feature quantity distribution generation target includes a feature quantity obtained from driving data relating to a driver's driving operation, a feature quantity obtained from behavior data relating to the behavior of the vehicle, a feature quantity obtained from image data taken by an in-vehicle camera, and the own vehicle The driving scene label estimation apparatus according to any one of claims 1 to 9, wherein at least one of feature quantities obtained from position data relating to the position of the vehicle is included.