JP6511982B2 - Driving operation discrimination device - Google Patents

Description

  The present invention relates to a technology for determining a driving operation corresponding to a driving scene.

  Conventionally, a driving behavior distribution, which is a distribution of feature quantities extracted from data belonging to divided sections (driving scenes), is automatically divided in advance by automatically dividing a time series of data in which a driver's operation or vehicle behavior is detected. There is known a technique for approximating with a weighted sum of a plurality of characteristic distributions, obtaining the weight as a topic ratio, and using this topic ratio to assign a label that matches each divided driving scene (Patent Document 1).

JP 2014-235605 A

  However, in the prior art in which labels are provided using topic proportions determined for each driving scene, it is possible to understand individual driving scenes, but in comparison with the preceding and following scenes, driving performed only in that scene There was a problem that only the operation could not be extracted.

  For example, in a driving scene in which a lane is changed while turning a curve, the driving operation in turning a curve is superimposed on the driving operation in changing the lane. In this case, a driving operation to be focused on (a driving operation at the time of a lane change) occurs simultaneously with a driving operation spanning a plurality of scenes (a driving operation at the time of turning a curve). It becomes difficult to correctly determine the driving operation at the time of a lane change distorted by the driving operation at the time of turning.

  The present invention has been made in view of these problems, and an object of the present invention is to provide a technique for accurately determining driving operation under complicated conditions.

A driving operation determination apparatus according to the present invention includes a data collection unit, a scene division unit, a behavior feature quantity generation unit, a driving feature quantity generation unit, and a determination unit.
The data collection unit repeatedly collects vehicle behavior data related to the behavior of the vehicle. The scene division unit divides the series of vehicle behavior data collected by the data collection section into a plurality of partial series each representing any driving scene. The behavior feature quantity generation unit generates a behavior feature quantity representing a feature of the vehicle behavior data for each of the partial sequences divided by the scene division unit. The operation feature quantity generation unit generates an operation feature quantity including at least the behavior feature quantity generated by the behavior feature quantity generation section and a change amount of the behavior feature quantity obtained from the time series of the behavior feature quantity. The determination unit uses the driving feature amount generated by the driving feature amount generation unit and the driving feature amount model prepared in advance for each driving operation to be determined, and the driving operation corresponding to the driving scene represented by the partial sequence To determine

  According to such a configuration, not only the behavior feature amount in each driving scene to be focused on, but also the change amount obtained from the time series of the behavior feature amount is used to determine the driving operation. The change amount of the behavior feature value removes the feature of the steady operation across multiple scenes, so the driving operation to be focused on (for example, the driving operation when changing lanes) is a driving operation across multiple scenes (for example, a curve is bent) Even if the driving operation is distorted, the driving operation corresponding to the driving scene can be determined more accurately.

  In addition, the code | symbol in the parentheses described in the claim shows correspondence with the specific means as described in embodiment mentioned later as one aspect, Comprising: The technical scope of this invention is limited. is not.

FIG. 1 is a block diagram showing an overall configuration of a driving support device. It is explanatory drawing which shows the division process of a driving | operation scene. It is a flowchart which shows the process in a topic ratio calculating part. It is explanatory drawing which shows the calculation process of a topic ratio. It is a flowchart which shows the process in a driving | operation feature-value production | generation part. It is explanatory drawing which shows the production | generation process of a driving | operation feature-value. It is a flowchart which shows the process in a driving operation discrimination | determination part. It is a flowchart which shows the process in a learning process part. It is explanatory drawing which shows dimension selection of a driving | operation feature-value. It is a figure which illustrates the driving | operation evaluation screen provided by the driving assistance provision part.

Hereinafter, a driving support apparatus according to an embodiment to which a driving operation determination apparatus of the present invention is applied will be described with reference to the drawings.
[1. overall structure]
As shown in FIG. 1, the driving support device 1 includes a vehicle behavior data collecting unit 2, a driving scene dividing unit 3, a topic ratio calculating unit 4, a related data collecting unit 5, and a driving feature quantity generating unit 6. A driving operation model database 7, a driving operation determination unit 8, a model learning unit 9, and a driving support providing unit 10 are provided.

  The driving support device 1 is realized by one or more microcomputers (microcomputers). The function of each unit is realized by the CPU (not shown) included in the microcomputer executing a predetermined program stored in the memory, and the database is realized on the memory. That is, FIG. 1 illustrates the memory storing various functions realized by the CPU and information used by the CPU divided into functional blocks. However, the units representing these functions do not necessarily have to be realized by software, and all or part of them may be realized by hardware such as a logic circuit.

[2. Vehicle Behavior Data Collection]
The vehicle behavior data collection unit 2 repeatedly collects data representing the operation by the driver and data representing the behavior of the vehicle appearing as a result of the operation through various sensors mounted on the vehicle, and also respectively collects these data. Differentiated differential data (dynamic features) are generated, and multidimensional data obtained by putting them together are output as vehicle behavior data relating to vehicle behavior.

  As data representing the operation by the driver, for example, the operation amount of the accelerator pedal, the operation amount of the brake pedal, the operation amount of the steering wheel (steering angle), the operation state of the direction indicator, the shift position of the transmission, etc. are used be able to. Further, as data representing the behavior of the vehicle, for example, the speed of the vehicle, a yaw rate, etc. can be used.

[3. Driving scene division section]
The driving scene division unit 3 statistically analyzes the vehicle behavior data obtained from the vehicle behavior data collecting unit 2 using a model from the driver's environmental awareness to the operation, and the driving scene felt by the driver individual (or a general driver) By extracting the switching points of, the time series of the vehicle behavior data is segmented (discretized) into a plurality of partial series each representing any driving scene. Specifically, a double segmentation analyzer (DAA: Double Articulation Analyzer) that performs segmentation by an unsupervised driving scene segmentation method using a double segmentation structure is used.

  As shown in FIG. 2, the DAA first makes a cluster representing various vehicle states grasped from the vehicle behavior data in a multi-dimensional space representing a range of vehicle behavior data, and a transition probability between each cluster in advance. By defining the cluster and statistically processing which cluster the vehicle behavior data acquired from the vehicle behavior data collection unit 2 belongs to by using these pieces of information, the time series of the vehicle behavior data can be divided into unit units. It divides into every state (that is, every cluster) of vehicles which become. However, by associating identification symbols with each cluster in advance, the time series of vehicle behavior data is converted into a symbol string indicating which cluster it belongs to. For generation of this symbol string, for example, a hierarchical Dirichlet process Hidden Markov Model (HDP-HMM), which is one of a hidden state and a model represented by a probabilistic transition between the states, can be used.

  Next, DAA uses the generated string to generate a predetermined driving scene using Nested Pitman-Yor Language Model (NPYLM), which is an example of an unsupervised chunking method of discrete character strings using statistical information. Segmentation into meaningful subsequences. At this time, the dictionary size (i.e., the number of partial sequences) is as small as possible, and the generation probability of the entire symbol string consisting of the arrangement of partial sequences is maximized. This makes it possible to segment vehicle behavior data into a driving scene. However, the transition probability between partial sequences and the generation probability of partial sequences use those previously generated by learning.

  As for DAA to which HDP-HMM and NPYLM are applied, T. Taniguchi et al, "Semiotic Prediction of Driving Behavior using Unsupervised Double Articulation Analyzer", IEEE Intelligent Vehicles Symposium, 2012, published separately by the present applicant. And 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. However, the method used for generation of symbols and segmentation of symbols is not limited to the HDP-HMM or NYLM, and other methods may be used.

[4. Topic ratio operation unit]
The processing in the topic ratio calculation unit 4 will be described using the flowchart shown in FIG. 3 and the explanatory view shown in FIG. The present process is performed for each driving scene generated by the driving scene division unit 3.

  When the process starts, the CPU functioning as the topic ratio computing unit 4 first uses the vehicle behavior data belonging to the partial series corresponding to the driving scene to be focused on at S110 to use the vehicle behavior histogram corresponding to the feature quantity distribution. Generate

  In the following S120, the topic ratio obtained by expressing the generated vehicle behavior histogram by a weighted sum of a plurality of characteristic distributions (driving topics) prepared in advance is calculated, and the process is ended. In the present embodiment, this topic ratio corresponds to the behavior feature quantity.

  The vehicle behavior histogram represents, as a distribution, feature quantities of vehicle behavior data belonging to a partial series corresponding to a driving scene of interest. For example, when multi-dimensional data including an accelerator pedal operation amount, a brake pedal operation amount, a steering operation amount, a vehicle speed, differential data of each of these, etc. is used as a feature amount, a feature space representing the range of the multi-dimensional data A histogram obtained by representing the frequency of appearance of data for each index is generated by using a plurality of divisions as an index. Note that as shown in FIG. 4, a histogram may be generated for each piece of data, and a combination of the generated histograms may be referred to as a vehicle behavior histogram.

  The driving topic consists of K (for example, 100) basis feature distributions Topic 1 to Topic K used as a basis distribution when the vehicle behavior histogram generated in S110 is represented by a mixture of a plurality of distributions (histograms) .

  The topic ratio is a mixing ratio determined as a vehicle behavior histogram represented by mixing base feature distributions, and represents the content ratio of each driving topic to the vehicle behavior histogram.

  To generate topic proportions, consider each partial series corresponding to any of the driving scenes as “one document” and the observed feature distribution (here, the vehicle behavior histogram) as “one word”. Create using topic estimation methods used in the natural language processing field. Here, the latent Dirichlet Allocation Method (LDA: Latent Dirichlet Allocation) is used. When the vehicle behavior histogram to be processed is configured by a plurality of types of histograms, multimodal LDA, which is an extended method of LDA, is used. Specifically, using an approximation method such as variational Bayesian and Gibbs sampling to realize LDA (or multimodal LDA), only the E step is executed among the E step and M steps performed in these processes. It can be determined by

  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. Since it is described in of Sciences, 2004 etc., description is abbreviate | omitted here. In addition, since a method of generating the basis feature distribution is a known method described in, for example, Japanese Patent Application Laid-Open No. 2014-235605, the description is omitted.

[5. Related Data Collection Department]
The related data collection unit 5 includes a video data collection unit 51, a position information data collection unit 52, and a peripheral object data collection unit 53.

  The video data collection unit 51 collects video data output from an on-vehicle camera installed to photograph the vehicle periphery such as the front, the side, and the rear of the vehicle. The position information data collecting unit 52 collects position information data on the position of the vehicle acquired by a global positioning system (GPS) or an inertial measurement unit (IMU). The peripheral object data collection unit 53 collects peripheral object data output from various radar sensors installed so as to set the search area to the vehicle periphery such as the front, the side, and the rear of the vehicle. In addition, for example, data collected by communication with an infrastructure may be used as these data.

[6. Operation feature generator]
The processing in the driving feature quantity generation unit 6 will be described using the flowchart shown in FIG. This process is activated each time the topic ratio is calculated by the topic ratio calculating unit 4.

  When the present process is activated, the CPU functioning as the driving feature quantity generating unit 6 first calculates the amount of change in the topic ratio in S210. Here, as shown in FIG. 6, it is assumed that the driving scene to be processed is i, the driving scene one cycle before is i-1, and the driving scene one cycle after is i + 1, the driving scene i and the driving scene i. A first variation, which is the difference between the topic ratio of -1, and a second variation, which is the difference between the topic ratio of the driving scene i + 1 and the driving scene i, are obtained.

  In S220, based on the video data and the peripheral object data collected by the video data collection unit 51 and the peripheral object data collection unit 53, the inter-vehicle distance and the relative speed with the preceding vehicle traveling in front of the own vehicle are obtained.

  In the subsequent S230, the information on the intersection extracted from the position information data collected by the position information data collector 52, the video data collected by the video data collector 51 or the map data acquired from the on-vehicle navigation device (not shown) Based on the above, the distance to the intersection existing in the traveling direction of the host vehicle is determined.

  At S240, the topic ratio estimated by the topic ratio calculation unit 4, the first change amount and the second change amount obtained at S210, the inter-vehicle distance obtained at S220, and the distance to the intersection obtained at S230 As this, the operation feature quantity which combined these each element is generated, and this processing is ended.

[7. Driving operation model database]
The driving operation model database 7 is prepared for each driving operation to be determined, and when a driving feature amount is input, the driving operation model database 7 is a discriminator that outputs an identification result as to whether the driving operation corresponds to the driving operation to be determined. That is, the typical value of the driving feature value representing the driving operation to be discriminated is the driving operation model, and the discriminator discriminates the driving operation based on the driving operation model.

  Specifically, an identifier such as an AdaBoost identifier or an SVM (support vector machine) is used. In addition, it may be a discriminator that discriminates a data set obtained by learning according to the k-nearest neighbor method. In the k-nearest neighbor method, k points close to each other in the feature space are selected with respect to a certain input, and based on the data set to which the selected point belongs, the majority determines whether the input point belongs to which data set Do. As these classifiers, those learned by the model learning unit 9 described later are used.

In addition, as driving | operation operation used as discrimination | determination object, specifically, straight running, curve road driving | running | working, right turn, left turn, lane change, passing, start, a stop etc. are mentioned.
[8. Driving operation discrimination unit]
The process in the driving operation determination unit 8 will be described with reference to the flowchart shown in FIG. The present process is activated each time a driving feature quantity is generated by the driving feature quantity generation unit 6.

  When the present process is activated, the CPU functioning as the drive operation determination unit 8 first selects one of the models stored in the drive operation model database 7, that is, one of the drive operations to be determined in S310. Do.

At S320, the driving feature is applied to the selected model which is the model selected at S310 to calculate the score of the selected model. For example, assuming that the operation feature quantity is x = (x ₁ , x ₂ ..., X _j ...), H _j (x) a function (eg weak classifiers) of judgment on the j-th element x _j of the feature quantity The score S is obtained by equation (1), where ω _j is the weight (reliability) of The function h _j (x) and the weight ω _j are determined by the selection model.

In continuing S330, it is judged whether the above-mentioned processing of S310-S320 has been performed about all the models memorized by driving operation model database 7 or not. If there is an unprocessed model, the process returns to S310, and if all models have been processed, the process proceeds to S340.

In S340, a model in which the score S calculated in S320 is maximized is identified, and a driving operation corresponding to the identified model is extracted as a discrimination result.
In the subsequent S350, the extracted driving operation and the score S representing the reliability of the determination are output in association with each other, and the present process is ended. The larger the score S is, the higher the reliability of the determination is.

[9. Model learning section]
The model learning unit 9 includes a driving operation recording unit 91, a driving operation database 92, and a learning processing unit 93.

  Each time the driving operation recording unit 91 outputs the judgment result (the driving operation and the reliability of the judgment) from the driving operation judging unit 8, the driving operation recording unit 91 associates the judgment result with the driving feature amount used for the judgment. The operation data is recorded in the driving operation database 92.

  Next, processing in the learning processing unit 93 will be described along the flowchart shown in FIG. The present process is activated every predetermined period (for example, several hours, one day, one week, etc.) or every time when a certain amount of driving operation data is accumulated in the driving operation database 92.

  When the present process is activated, the CPU functioning as the learning processing unit 93 first operates the driving operation data from the driving characteristic amount expressed as m-dimensional data in the driving operation data as shown in FIG. 9 at S410. Select n-dimensional data to be used for the model. However, m and n are positive integers such that m ≧ n. Specifically, n pieces of data are selected from among data with large factor loads obtained by subjecting m-dimensional operation feature quantities to principal component analysis. When the AdaBoost discriminator is used to discriminate the driving operation, data corresponding to the top n weak discriminators may be selected in accordance with the order in which the weak discriminators are adopted. That is, the selection may be made in the order of the highest contribution to discrimination.

At S420, the selected n-dimensional driving feature amount is used as learning data to learn the classifiers (h _j (x), ω _j ), and the model stored in the driving operation model database 7 according to the learning result To finish this process.

[10. Driving support provided]
The driving support providing unit 10 provides various driving support according to the determination result of the driving operation determination unit 8.

  As one of the driving support, for example, according to the determined driving operation, the operation of the existing driving system such as acceleration / deceleration or steering state of the vehicle may be controlled. Specifically, it is conceivable to end the driving support system presenting the lane change timing when the lane change completion is determined.

  Further, as one type of driving assistance, the determination result may be displayed on an on-vehicle display mounted on the vehicle to display various information, or the display content of the on-vehicle display may be switched according to the determination result.

  Furthermore, as one of the driving support, the smoothness of the driving operation is evaluated from the result of discrimination of the driving operation and the magnitude of the change amount of the vehicle behavior observed at that time, or the inter-vehicle distance with the surrounding vehicle during the driving operation The driver's skill may be evaluated based on whether the driver's attention is given to the surroundings or the driving behavior and the distance between the vehicles appropriate for each driving operation. And these evaluation results may be displayed on a vehicle-mounted display etc., for example in a format as shown in FIG.

[11. effect]
As described above, in the driving support device 1, not only the topic rate but also the change rate of the topic rate (the first change amount, the second change amount) is used as the driving feature amount used to determine the driving operation. . The variation of the topic ratio is the one in which the feature of steady driving operation across multiple scenes is removed, so even in a complicated situation where the driving operation to be focused on is distorted by the driving operation across multiple scenes The driving operation to which attention is paid can be accurately determined. Therefore, for example, it is possible to correctly determine the lane change while traveling on a curved road as shown in FIG.

  Further, in the driving support device 1, information such as video data, position information data, an inter-vehicle distance to a preceding vehicle obtained from surrounding object data, a distance to an intersection, etc. is added to the driving feature amount. As a result, it is possible to improve the determination accuracy of various operations that may occur at specific points or specific situations, such as turning operations at intersections and overtaking operations when the distance between the vehicle and the vehicle in front is blocked.

  Further, in the driving support device 1, since the feature amount useful for determining the driving operation is extracted to create the driving operation model, it is possible to suppress the decrease in detection performance due to the influence of the non-useful feature amount. The processing load required for model learning can be reduced.

The driving support device 1 excluding the driving support providing unit 10 corresponds to the driving operation determination device.
[12. Other embodiments]
As mentioned above, although embodiment of this invention was described, this invention can take various forms, without being limited to the said embodiment.

  (1) In the above embodiment, both the difference from the previous cycle and the difference from the last cycle are used as the change amount of the topic ratio, but only one of them may be used. Alternatively, a time series of differences in topic proportions obtained over a plurality of past cycles may be used.

  (2) In the above embodiment, the topic ratio and the change amount thereof are used as the driving feature amount, but for example, the vehicle behavior data itself or the change amount thereof may be used. However, since the topic ratio is information with a high degree of abstraction compared to the case where vehicle behavior data is used as it is, using the topic ratio can improve the robustness when determining the driving operation. it can.

  (3) In the above embodiment, when dividing the driving scene, only the vehicle behavior data collected by the vehicle behavior data collection unit 2 is used, but in addition to this, for example, the related data collection unit 5 collects Image data, position information data, peripheral object data, etc. may be used.

  (4) The function of one component in the above embodiment may be distributed to a plurality of components or the function 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. In addition, part of the configuration of the above embodiment may be omitted. Further, at least a part of the configuration of the above-described embodiment may be added to or replaced with the configuration of the other above-described embodiment. In addition, all the aspects contained in the technical thought specified only by the words described in the claim are an embodiment of the present invention.

  (5) Other than the above-described driving operation determination device and the driving support device, a system including the driving operation determination device as a component, a program for causing a computer to function as the driving operation determination device, a medium storing the program The present invention can also be realized in various forms such as an operation determination method.

  DESCRIPTION OF SYMBOLS 1 ... driving assistance device, 2 ... vehicle behavior data collection part, 3 ... driving scene division part, 4 ... topic ratio operation part, 5 ... related data collection part, 6 ... driving feature quantity generation part, 7 ... driving operation model database, 8 ... driving operation determination unit, 9 ... model learning unit, 10 ... driving support providing unit, 51 ... video data collecting unit, 52 ... position information data collecting unit, 53 ... peripheral object data collecting unit, 91 ... driving operation recording unit, 92: driving operation database, 93: learning processing unit

Claims (7)

A data collection unit (2) that repeatedly collects vehicle behavior data related to vehicle behavior;
A scene division unit (3) for dividing the series of vehicle behavior data collected by the data collection section into a plurality of partial series each representing a driving scene of any type;
A behavior feature quantity generation unit (4) that generates a behavior feature quantity representing a feature of the vehicle behavior data for each of the partial sequences divided by the scene division unit;
An operation feature quantity generation unit (6) that generates an operation feature quantity including at least the behavior feature quantity generated by the behavior feature quantity generation unit and a change amount of the behavior feature quantity obtained from a time series of the behavior feature quantity When,
The driving operation corresponding to the driving scene represented by the partial series using the driving feature amount generated by the driving feature amount generation unit and the driving feature amount model prepared in advance for each driving operation to be discriminated A determination unit (7, 8) for determining
A driving operation determination device comprising: The behavior feature quantity generation unit
For one or more types of feature quantities obtained from the operation data collected by the data collection unit, a feature quantity distribution representing an appearance frequency for each index obtained by dividing the range of the feature quantities into a plurality is generated for each of the partial series A distribution generation unit (S110),
A mixture of a plurality of basis feature distributions used as a basis distribution for expressing the feature amount distribution as a driving topic, and the feature amount distribution generated by the distribution generation unit being expressed using the driving topic A ratio calculation unit (S120) that calculates a topic ratio that is a ratio as the behavior feature value;
The driving operation determination device according to claim 1, comprising:   The driving feature quantity generated by the driving feature quantity generation unit includes information extracted from at least one of video data, position information data, and peripheral object data. The driving operation determination device according to claim 1. A driving operation recording unit (91, 92) that records driving operation data in which the driving feature amount generated by the driving feature amount generation unit is associated with the determination result of the driving operation in the determination unit;
A learning processing unit (93) for learning the model used in the determination unit according to the driving operation data recorded by the driving operation recording unit;
The driving operation determination device according to any one of claims 1 to 3, further comprising:   5. The driving operation determination apparatus according to claim 4, wherein the driving operation data recorded by the driving operation recording unit includes a reliability of determination.   Among the m (m is a positive integer) elements constituting the driving operation data, the learning processing unit is selected from n (n is a positive integer less than m) elements selected in descending order of contribution to discrimination. The driving operation determination apparatus according to claim 4 or 5, wherein the learning is performed using n-dimensional data.   The degree of contribution to the discrimination is in order of the magnitude of the factor load amount obtained by performing principal component analysis of the driving operation data recorded by the driving operation recording unit for each of the discrimination results, or the weakness of the Adaboost discriminator The driving operation determination device according to claim 6, wherein the driving operation determination device is determined in accordance with the identifier adoption order.