Method and apparatus for processing accident relevance information

Abstract

The present invention relates to a method and an apparatus for processing accident relevance information. The method, according to an embodiment of the present invention, is performed by an electronic apparatus and comprises the steps of: generating an embedding vector by performing an embedding process on raw information which is text information describing an accident occurrence situation of another vehicle; classifying into a plurality of clusters by performing a clustering process on the generated embedding vector; extracting, for each cluster, a feature element related to the occurrence of a vehicle accident with respect to raw information belonging to the corresponding cluster; and when driving description information, which is text information related to a driving situation of a target vehicle, is input as input information, generating an embedding vector for the driving description information and then generating accident relevance information regarding the target vehicle according to the driving description information on the basis of the cluster, the feature element, and the embedding vector of the generated driving description information.

Description

Method and device for processing accident-related information

[0001] The present invention relates to an accident-related information processing technology, and more particularly, to a technology for generating accident-related information that quantitatively indicates the accident-relatedness (or risk) of the driving description information when driving description information, which is text information related to the driving situation of a vehicle, is given.

[0002] For the development of automobile driving and autonomous driving technologies, the technology to predict the relevance (or risk) of accidents when additional information about a specific vehicle driving situation is input is becoming increasingly important.

[0003] For example, by analyzing real-time video footage taken while driving in a car equipped with a camera, if the situation is determined to be highly likely to cause an accident, the vehicle can take preemptive measures, such as slowing down or warning the driver, thereby helping to ensure safe driving.

[0004] As another example, even if real-time analysis is not possible, if textual information describing driving situations (i.e., driving description information) can be input, providing a quantitative measure of the accident relevance of that input. This could be useful for creating driving scenarios for driving simulations or analyzing traffic accident investigation documents. Furthermore, technology is also needed to determine the relevance of input driving description information to pre-grouped vehicle accident cases.

[0005] Meanwhile, conventional techniques pre-secure a large amount of accident investigation data to calculate the correlation between text-based data and extract key features from the data. For example, accident investigation data may include information such as {weather, time of day, location, traffic volume, traffic conditions, vehicle status, road conditions, risk factors, driver behavior, and driver status}. By extracting and statistically analyzing each of these pieces of information to identify factors influencing accidents, the correlation between accident occurrences can be calculated. For example, considering a single feature such as {weather}, a method could be used to calculate the correlation between accident occurrences for {sunny days, cloudy days, snowy days}.

[0006] However, prior art techniques can miss important features in text describing accident situations, and they also struggle to extract or utilize information generated by the synergy of multiple features. For example, while prior art can extract {weather, time zone} information from a sentence, it can miss {driver behavior} information, which may also be present in the input sentence. Furthermore, extracting only a few features, which can only be gleaned from the overall context, presents limitations in accurately representing an accident situation.

[0007] In addition, since conventional techniques require assumptions such as statistical independence or conditional independence between feature elements when extracting feature elements and statistically identifying their relationships, correct analysis is impossible for situations that do not meet these assumptions.

[0008] However, the above-described content merely provides background information on the present invention and does not correspond to previously disclosed technology.

[0009] In order to solve the problems of the prior art as described above, the purpose of the present invention is to provide a technology for generating accident relevance information that quantitatively indicates the relevance (or risk) of an accident occurrence for driving description information when driving description information, which is text information related to the driving situation of a vehicle, is given.

[0010] However, the problems to be solved by the present invention are not limited to the problems mentioned above, and other problems not mentioned can be clearly understood by a person having ordinary skill in the technical field to which the present invention belongs from the description below.

[0011] According to an embodiment of the present invention for solving the above-described problem, a method is performed by an electronic device, comprising: a step of generating an embedding vector through embedding processing of raw information, which is text information describing an accident situation of another vehicle; a step of classifying the generated embedding vector into a plurality of clusters through clustering processing; a step of extracting a feature element related to the occurrence of a vehicle accident with respect to raw information belonging to each cluster for each cluster; and a step of generating an embedding vector for the driving description information, which is text information related to a driving situation of a target vehicle, and then generating accident-related information for the target vehicle according to the driving description information based on the cluster, the feature element, and the generated embedding vector of the driving description information, wherein the accident-related information includes first information of a quantitative numerical value regarding the risk or possibility of an accident occurring for the target vehicle.

[0012] The method according to one embodiment of the present invention may further include a step of setting a representative vector for each cluster for embedding vectors belonging to each cluster.

[0013] In the step of generating the above accident-related information, the first information can be generated using the distance between the representative vector of each cluster and the embedding vector of the driving description information.

[0014] In the above setting step, any one of the average value, maximum density value, center value, mode value, maximum value, and minimum value of the embedding vectors forming each cluster can be set as the representative vector of the cluster.

[0015] In the above setting step, multiple values ​​among the average value, maximum density value, center value, mode value, maximum value, and minimum value for the embedding vectors forming each cluster can be calculated, and the weighted average result of the multiple values ​​calculated can be set as the representative vector.

[0016] In the above setting step, after dimension reduction of the vector space according to the clustering process, the center of the reduced vector space can be set as the representative vector.

[0017] In the step of generating the accident relevance information, the accident relevance information can be generated by restoring the dimensionally reduced representative vector to the vector space size of the embedding vector of the driving description information, or by reducing the dimension of the embedding vector of the driving description information by the size of the dimensionally reduced representative vector, and comparing the embedding vector of the driving description information with the representative vector.

[0018] A method according to one embodiment of the present invention may further include a step of setting an overall representative vector for all of the clustered embedding vectors.

[0019] In the step of generating the above accident-related information, the first information can be generated using the distance between the entire representative vector and the embedding vector of the driving description information.

[0020] The above accident-related information further includes second information in the form of text information about the type of vehicle accident that has a risk or possibility of causing the accident, and the second information may include characteristic elements related to the first information.

[0021] The method according to one embodiment of the present invention may further include a step of generating a control signal for transmitting text about the second information to the driver of the target vehicle together with a warning signal when the first information has a numerical value exceeding a reference value.

[0022] The method according to one embodiment of the present invention may further include a step of extracting, for each cluster, a representative set of feature elements, which is a set representing the feature elements extracted from each cluster.

[0023] In the step of generating the above accident relevance information, the embedding vector of the driving description information may be mapped to one of each cluster, and the accident relevance information including a representative set of feature elements for the mapped cluster may be generated.

[0024] The method according to one embodiment of the present invention may further include a step of extracting, for each cluster, a representative set of feature elements, which is a set representing the feature elements extracted from each cluster.

[0025] In the step of generating the above accident relevance information, the embedding vector of the driving description information is mapped to one of the clusters, feature elements are extracted from the driving description information, and then the accident relevance information can be generated by comparing a representative set of feature elements of the mapped cluster with the feature elements of the driving description information.

[0026] The above accident-related information may further include second information in the form of text information about the type of vehicle accident that has a risk or possibility of causing the accident.

[0027] In the step of generating the above accident relevance information, the accident relevance information may further include second information on a feature element related to a cluster corresponding to a distance within a reference value among the distances between the representative vector of each cluster and the embedding vector of the driving description information.

[0028] The method according to one embodiment of the present invention may further include, when the input information is an image or video of a driving image of the target vehicle, a step of extracting the driving description information, which is text information describing the driving situation, from the image or video.

[0029] A device according to one embodiment of the present invention includes a memory storing raw information, which is text information describing an accident situation involving another vehicle; and a control unit performing control using the information stored in the memory.

[0030] The above control unit controls to generate an embedding vector through embedding processing on the raw information, controls to classify the generated embedding vector into a plurality of clusters through clustering processing, controls to extract feature elements related to the occurrence of a vehicle accident for the raw information belonging to each cluster for each cluster, and controls to generate an embedding vector for the driving description information, which is text information related to the driving situation of the target vehicle, and then controls to generate accident-related information for the target vehicle according to the driving description information based on the cluster, the feature elements, and the generated embedding vector of the driving description information.

[0031] The present invention, configured as described above, has the advantage of being able to generate accident relevance information that quantitatively indicates the relevance (or risk) of an accident occurrence for driving description information when driving description information, which is text information related to the driving situation of a vehicle, is given.

[0032] In particular, conventional techniques that extract and analyze features (e.g., weather, time zone, location, etc.) related to accident occurrence from input sentences (i.e., text information) require the assumption that each extracted element is statistically independent or conditionally independent. This assumption simplifies the analysis and facilitates the application of various statistical techniques. However, conventional techniques cannot properly analyze situations that do not meet this assumption.

[0033] In contrast, the present invention uses a nonlinear artificial neural network technique to convert driving description information, which is text information related to the driving situation of a vehicle, into a high-dimensional embedding vector, process it, cluster it, and analyze the feature elements. Accordingly, the present invention can better understand the relationships between feature elements without the aforementioned assumptions, can be applied to situations that are not possible to analyze with conventional techniques, and has the advantage of increasing the simplicity of the calculations required to derive accident-related information because calculations only need to be performed on newly input driving description information.

[0034] In addition, the present invention has the advantage of being able to increase the usability of the calculated accident-related information by being able to produce quantitative figures on accident-related information in various ways based on clustered accident cases.

[0035] In addition, the present invention extracts text-based driving description information from an image or video input to generate accident-related information, thereby enabling the driving safety of the vehicle to be improved by identifying the accident-relatedness of the current driving status of the vehicle in real time through a black box of the vehicle in operation and feeding it back to the vehicle.

[0036] The effects that can be obtained from the present invention are not limited to the effects mentioned above, and other effects not mentioned can be clearly understood by a person having ordinary skill in the art to which the present invention belongs from the description below.

[0037] Figure 1 shows a block diagram of a device (100) according to one embodiment of the present invention.

[0038] FIG. 2 shows a block diagram of a control unit (150) in a device (100) according to one embodiment of the present invention.

[0039] Figure 3 shows a flowchart of a method according to one embodiment of the present invention.

[0040] Figure 4 shows a detailed flowchart for S310.

[0041] Figure 5 shows examples of results for S311 and S312 for raw information.

[0042] Figure 6 shows an example of each cluster processed by clustering according to S312 and the representative vector of each cluster.

[0043] Figure 7 shows examples of embedding vectors q1 and q2 of driving description information for each cluster according to Figure 6.

[0044] The above-described objects, means, and resulting effects of the present invention will become more apparent through the following detailed description, taken in conjunction with the accompanying drawings. Accordingly, those skilled in the art will be able to readily implement the technical concepts of the present invention. Furthermore, in describing the present invention, if a detailed description of known technology related to the present invention is deemed to unnecessarily obscure the gist of the invention, such detailed description will be omitted.

[0045] The terminology used herein is for the purpose of describing embodiments and is not intended to limit the present invention. In this specification, singular forms also include plural forms, unless specifically stated otherwise. In this specification, terms such as "include," "provide," "provide," or "have" do not exclude the presence or addition of one or more other components other than the mentioned components.

[0046] In this specification, terms such as "or", "at least one", etc. may refer to one of the words listed together, or to a combination of two or more. For example, "A or B", "at least one of A and B" may include only one of A or B, or may include both A and B.

[0047] In this specification, descriptions using the phrase "for example" or the like should not be construed as limiting the embodiments of the invention in terms of the effects of variations such as tolerances, measurement errors, limitations of measurement accuracy, and other commonly known factors, including the information presented, such as cited characteristics, variables, or values, may not be exact matches.

[0048] In this specification, when a component is described as being "connected" or "connected" to another component, it should be understood that it may be directly connected or connected to that other component, but that there may be other components in between. Conversely, when a component is described as being "directly connected" or "directly connected" to another component, it should be understood that there are no other components in between.

[0049] In this specification, when a component is described as being "on" or "in contact with" another component, it should be understood that it may be directly on or connected to the other component, but there may be another component in between. Conversely, when a component is described as being "directly on" or "in direct contact with" another component, it should be understood that there is no other component in between. Other expressions that describe the relationship between components, such as "between" and "directly between", can be interpreted similarly.

[0050] In this specification, terms such as "first" and "second" may be used to describe various components, but the components should not be limited by these terms. Furthermore, these terms should not be construed to limit the order of each component, but rather may be used to distinguish one component from another. For example, a "first component" may be referred to as a "second component," and similarly, a "second component" may also be referred to as a "first component."

[0051] Unless otherwise defined, all terms used herein may be used in their common sense by those of ordinary skill in the art to which the present invention pertains. Furthermore, terms defined in commonly used dictionaries are not to be interpreted ideally or excessively unless explicitly and specifically defined otherwise.

[0052]

[0053] Hereinafter, a preferred embodiment according to the present invention will be described in detail with reference to the attached drawings.

[0054] Figure 1 shows a block diagram of a device (100) according to one embodiment of the present invention.

[0055] A device (100) according to one embodiment of the present invention (hereinafter referred to as “the device”) is a device that performs as its basic function the function of generating information on the relevance (or risk) of an accident occurrence with respect to driving description information (hereinafter referred to as “accident relevance information”). This accident relevance information may include quantitative numerical value information (hereinafter referred to as “first information”) on the relevance of an accident occurrence (i.e., the risk or possibility of an accident occurrence of a target vehicle), and may additionally include text information (hereinafter referred to as “second information”) on the type of vehicle accident with a high relevance of an accident occurrence.

[0056] At this time, the driving description information is information related to the driving situation of the target vehicle and corresponds to text-based information. In other words, the driving description information may be information describing, in text form, the driving behavior of the vehicle. For example, the driving description information may include information describing {driving speed, driving direction, weather, time of day, driving location, traffic volume, traffic conditions, vehicle condition, road conditions, surrounding vehicles, surrounding objects, surrounding signs, and surrounding road markings}.

[0057] To this end, the device (10) may perform a function (hereinafter referred to as the "first function") that processes raw driving description raw information related to a vehicle accident (hereinafter referred to as "raw information"). At this time, the raw information corresponds to any text-type data describing the situation of a previous accident involving another vehicle. For example, the raw information may be information extracted from a traffic accident investigation document created when an accident involving another vehicle occurs, a newspaper article about the accident, traffic safety education materials, or a novel related to traffic accidents.

[0058] These raw data have one thing in common: they assume a car accident. However, each raw data may have unique characteristics, and they may not necessarily represent similar car accident situations. Nevertheless, using the method described below, we can determine which car accident case each driving description belongs to, and provide quantitative figures for its relevance to the accident. Here, in determining which car accident case each driving description belongs to, we assume that the raw data is not explicitly labeled. In the method described below, clustering, a type of unsupervised learning method, can be used to classify similar cases for each raw data into clusters.

[0059] In other words, the first function corresponds to the function of processing raw information to generate some standard information (hereinafter referred to as "standard information"). In other words, the standard information may be information used as a reference to generate accident-related information from given driving description information.

[0060] That is, according to the first function, the device (100) receives raw information and performs a predetermined processing, which will be described later. In terms of actual service, the raw information can be input and processed in its entirety in advance, but it can also be input and processed gradually over time. This is because, as will be described later, when performing the first function, a method of internally calculating and processing an embedding vector can be introduced, and streaming-based clustering can be utilized.

[0061] Additionally, the device (10) can perform a function (hereinafter referred to as the "second function") of extracting text-based driving description information from input information. That is, the input information may not be in text form, but may be an image or video related to vehicle driving. In this case, the second function of extracting driving description information by converting the input information of the image or video into text form can be performed.

[0062] At this time, the image or video can be obtained from a driving video of the target vehicle. The driving video may be a video captured by a video collection device such as a black box installed in the target vehicle while the target vehicle is driving, but is not limited thereto.

[0063] For example, the second function may be performed using an image processing technique or a pre-trained machine learning model (hereinafter referred to as “model”) that converts an image or video into text information describing the driving situation of a target vehicle containing the image or video.

[0064] At this time, the model may be a model trained using a machine learning technique of supervised learning through training data of input and output data pairs (datasets). That is, the model can be trained using training data that each includes input data and output data. Accordingly, the model has a function for the relationship between input data and output data through training, and expresses this function using various parameters. In other words, the trained model can express the relationship between input data and output data using parameters such as weights and biases. Accordingly, when performing inference, the trained model can output the output data for the corresponding function as the result value when the target input data is input as the input value.

[0065] For example, machine learning techniques applied to train a model may include, but are not limited to, Artificial neural network, Boosting, Bayesian statistics, Decision tree, Gaussian process regression, Nearest neighbor algorithm, Support vector machine, Random forests, Symbolic machine learning, Ensembles of classifiers, or Deep learning.

[0066] In particular, when the model is a deep learning model trained using deep learning techniques, the relationship between input data and output data is expressed as multiple layers, and these multiple representation layers are also referred to as a "neural network."

[0067] For example, deep learning techniques may include, but are not limited to, Deep Neural Network (DNN), Convolutional Neural Network (CNN), Recurrent Neural Network (RNN), Restricted Boltzmann Machine (RBM), Deep Belief Network (DBN), and Deep Q-Networks.

[0068] Specifically, the training data used to train the model includes images or videos related to vehicle accident situations as input data. Furthermore, the training data includes textual information describing the situation in the image or video as output data. In other words, depending on the training data, the model can be trained to optimally express the relationship between the input data (image or video) and the output data (textual driving description information describing the vehicle accident situation in the image or video) using optimal parameters.

[0069] In addition, the present device (100) can perform a third function corresponding to the basic function described above. That is, the third function is a function of generating and outputting accident-related information for the driving description information using the reference information prepared according to the first function when driving description information is input in the form of a query. Of course, in this case, if the input information is not text-type information but image or video-type information, the second function is performed to extract driving description information for the image or video, and then the third function is performed, thereby producing accident-related information for the extracted driving description information.

[0070] In addition, the present device (100) may also perform a function of utilizing accident-related information generated according to the third function (hereinafter referred to as “the fourth function”).

[0071] The present device (100) is an electronic device capable of computing according to the performance of the first to fourth functions. For example, the electronic device may be a general-purpose computing device such as a desktop personal computer, a laptop personal computer, a tablet personal computer, a netbook computer, a workstation, a personal digital assistant (PDA), a smartphone, a smartpad, or a mobile phone, or a dedicated embedded system, but is not limited thereto.

[0072] Specifically, the present device (100) may include an input unit (110), a communication unit (120), a display (130), a memory (140), and a control unit (150), as illustrated in FIG. 1. Of course, the present device (100) may also include an image collection device that captures driving images.

[0073] The input unit (110) generates input data in response to various user inputs and may include various input means. For example, the input unit (110) may include, but is not limited to, a keyboard, a keypad, a dome switch, a touch panel, a touch key, a touch pad, a mouse, a menu button, etc.

[0074] The communication unit (120) is a component that performs communication with other devices. For example, the communication unit (120) may perform wireless communication such as 5G (5th generation communication), LTE-A (long term evolution-advanced), LTE (long term evolution), Bluetooth, BLE (Bluetooth low energy), NFC (near field communication), and WiFi communication, or may perform wired communication such as cable communication, but is not limited thereto. For example, the communication unit (120) may receive raw information, driving description information, driving images, etc. from other devices. In addition, the communication unit (120) may transmit results according to the first to fourth functions or results according to the method to be described later, etc. to other devices. In particular, when receiving driving description information or driving images, etc. from other devices and then transmitting the processing result therefor to other devices, the present device (100) may operate as a server that transmits the processing result.

[0075] The display (130) is a configuration that displays various image data on the screen. For example, the display (130) may be configured as a non-luminous panel or a luminous panel. For example, the display (130) may include, but is not limited to, a liquid crystal display (LCD), a light emitting diode (LED) display, an organic light emitting diode (OLED) display, a micro electro mechanical systems (MEMS) display, or an electronic paper display. For example, the display (130) may display input information, generated results, etc. on the screen. In addition, the display (130) may be implemented as a touch screen or the like by being coupled with the input unit (110).

[0076] The memory (140) stores various information required for the operation of the device (100). For example, the stored information may include, but is not limited to, raw input information, driving description information, driving images, models, or program information related to a method to be described later. For example, the memory (140) may include, but is not limited to, a hard disk type, a magnetic media type, a compact disc read-only memory (CD-ROM), an optical media type, a magneto-optical media type, a multimedia card micro type, a flash memory type, a read-only memory type, or a random access memory type, depending on its type. In addition, the memory (140) may be, but is not limited to, a cache, a buffer, a main memory, an auxiliary memory, or a separately provided storage system, depending on its use / location.

[0077] The control unit (150) can perform various control operations of the device (100). That is, the control unit (150) can control the execution of the method described below, and can control the operations of the remaining components of the device (100), such as the input unit (110), the communication unit (120), the display (130), the memory (140), etc. For example, the control unit (150) may include, but is not limited to, a hardware processor or a software process executed on the processor.

[0078] FIG. 2 shows a block diagram of a control unit (150) in a device (100) according to one embodiment of the present invention.

[0079] The control unit (150) controls the execution of the method according to one embodiment of the present invention, and may include an embedding unit (151), a clustering unit (152), a feature processing unit (153), a reference information management unit (154), an accident relevance processing unit (155), and an input information processing unit (156), as illustrated in FIG. 2. For example, the embedding unit (151), the clustering unit (152), the feature processing unit (153), the reference information management unit (154), the accident relevance processing unit (155), and the input information processing unit (156) may be hardware components of the control unit (150), or components of software processes executed in the control unit (150), but are not limited thereto.

[0080] Hereinafter, the method according to the present invention will be described in more detail.

[0081] Figure 3 shows a flowchart of a method according to one embodiment of the present invention.

[0082] A method according to one embodiment of the present invention (hereinafter referred to as "the present method") may be performed under the control of a control unit (150) and may be a method for performing the first to fourth functions. This present method may include steps S310 to S330, as illustrated in FIG. 3.

[0083] Figure 4 shows a detailed flowchart for S310.

[0084] Step S310 is a step that performs the first function. That is, the control unit (150) controls processing of raw information describing the vehicle accident situation. This step S310 may include steps S311 to S313, as illustrated in FIG. 4.

[0085] First, in S311, the embedding unit (151) of the control unit (150) performs embedding processing on the input raw information in the form of text (i.e., text sentences) and controls the output (generation) of an embedding vector for the raw information.

[0086] Here, various embedding techniques can be used, but it may be preferable to use an embedding technique that utilizes a transformer structure that can well reflect contextual information in the input text sentence. For example, BERT (Bidirectional Encoder Representations from Transformers) or improved variants of BERT (e.g., Roberta, XLNet, DistilBERT, etc.) can be used. Techniques such as self-attention, multi-head attention, feedforward networks, and residual connections are used within the transformer structure. These technologies show excellent performance in identifying relationships between sets of input tokens (sets of sentences decomposed through a tokenizer). In other words, since they learn the meaning of the tokens themselves corresponding to the feature elements and the relationships between tokens, there is no need to introduce assumptions such as statistical independence or conditional independence, and they can internally utilize relationships between feature elements that were difficult to interpret using conventional statistical methods.

[0087] For example, it is possible to obtain an embedding vector containing corresponding information without having to calculate the complex joint probability distribution between multiple variables. Of course, in S311, a pre-trained Transformer-based embedding model can also be used. In this case, the pre-trained embedding model can be updated through transfer learning using data such as traffic accident data, and the updated model can then be used in S311.

[0088] By applying various similarity calculation metrics to the embedding vectors output according to S311, the similarity for each embedding vector can be calculated. For example, cosine similarity is a representative example of a similarity calculation metric that can be used. However, it is not limited to this, and a similarity calculation metric that uses the inner product between two vectors, a similarity calculation metric that uses a distance calculation method in vector space such as L0 norm, L1 norm, L2 norm, and a similarity calculation metric that considers the variance from the average position such as Mahalanobis distance can also be used.

[0089] In S312, the clustering unit (152) of the control unit (150) controls the clustering process to group similar groups of embedded data (i.e., embedding vectors). The groups classified as similar through clustering in this manner are referred to as "clusters." In other words, similar cases for each raw information can be classified into each cluster through clustering, which is a type of unsupervised learning method.

[0090] For example, when performing clustering processing with K clusters, clustering methods such as K-Means, which clusters around the K means, K-neighbors, which determines K neighbors as clusters, DBSCAN, which forms clusters around points of dense data (i.e., embedding vectors), and Expectation Maximization, which repeatedly finds clusters after assuming a distribution model of the data (i.e., embedding vectors), can be used, but are not limited thereto.

[0091] In particular, in S312, it may be desirable to apply a streaming clustering method that performs clustering processing in a situation where data (i.e., embedding vectors) on new raw information can be continuously input. This allows clusters to be updated whenever data (i.e., embedding vectors) on new raw information are input, or clusters can be created after collecting data for a certain period of time. For example, streaming clustering methods such as streaming K-means, density-based streaming clustering, or hierarchical streaming clustering can be used, but are not limited thereto.

[0092] After going through this clustering process, a specific label is assigned to each embedding vector, and through the assigned label, it is possible to immediately know which group (cluster) a specific embedding vector belongs to.

[0093] Figure 5 shows examples of results for S311 and S312 for raw information.

[0094] For example, referring to Fig. 5, by performing embedding processing on raw information corresponding to sequences 1, 2, 3, and 4, embedding vectors corresponding to E1, E2, E3, and E4 can be obtained, respectively. Clustering is performed on the embedding vectors obtained in this manner. At this time, clustering label 1 corresponding to cluster 1 is assigned to sequences 1 and 2, and clustering label 2 corresponding to cluster 2 is assigned to sequences 3 and 4. At this time, raw information (i.e., embedding vectors) with the same clustering label corresponding to the same cluster can be identified as having similar information regarding vehicle accidents.

[0095] In other words, according to S312, similar raw information can be immediately identified and extracted simply by examining the clustering labels. Of course, using a similarity comparison method using embedding vectors allows for more detailed and sophisticated comparisons (i.e., comparisons of more precise similarities).

[0096] Meanwhile, in S312, a representative vector can be extracted for each cluster through clustering processing. This need for a representative vector for each cluster stems from the fact that the number of clusters is not fixed and can continuously increase. In other words, without such a representative vector, when performing S320 (described below), if a query regarding arbitrary driving description information is entered, all embedding vectors within the cluster must be individually compared with the embedding vector corresponding to the query. This process requires a significant amount of computation and is therefore impractical.

[0097] To improve these problems, the present invention can be implemented to extract and store representative vectors representing embedding vectors belonging to the same cluster (i.e., representing each cluster) in S312, and use each representative vector when necessary (i.e., when performing S320 described below). At this time, the representative vector is a vector that represents and expresses the embedding vector belonging to the cluster. This representative vector may have a value for any embedding vector belonging to the cluster, or may have a value resulting from additional processing performed on the embedding vector belonging to the cluster according to any one of the first to third methods described below.

[0098] For example, when there are K clusters (where K is a natural number greater than or equal to 2), a representative vector is extracted for each of the K clusters. At this time, the first to third methods for setting (extracting) the representative vector for each cluster are as follows.

[0099] - Method 1:

[0100] The first method is to set one of the mean, maximum density, center, mode, maximum, and minimum values ​​of the embedding vectors forming a cluster as the representative vector of the cluster.

[0101] - Method 2:

[0102] The second method is to calculate multiple values ​​among the mean, maximum density, center, mode, maximum, and minimum values ​​for the embedding vectors forming a certain cluster, and set the weighted average result of the calculated multiple values ​​(i.e., the average value using weights) as the representative vector.

[0103] - Third method:

[0104] The third method involves reducing the dimensionality of the vector space of each cluster and then setting the center of the reduced vector space as the representative vector. In this case, when performing S320 described below, in order to compare the embedding vector of a query for arbitrary driving description information with the representative vector, the dimensionally reduced representative vector can be restored to the vector space size of the embedding vector of the corresponding query and the comparison can be performed, or the embedding vector of the corresponding query can be reduced in dimensionality by the size of the dimensionally reduced representative vector and the comparison can be performed (e.g., distance comparison according to the second method described below). As a result of such a comparison, accident relevance information can be generated.

[0105] Meanwhile, in S312, a representative vector (hereinafter referred to as the "first overall representative vector") for all additionally clustered embedding vectors may be extracted. That is, the first overall representative vector is a representative vector for all embedding vectors belonging to the K clusters.

[0106] Of course, representative vectors (hereinafter referred to as "second overall representative vectors") can also be extracted for all remaining embedding vectors for which clustering processing has not been performed. In this case, the first and second overall representative vectors can be established using one of the first to third methods described above. These first and second overall representative vectors, along with the respective representative vectors for the K clusters, can be used in S320, described below.

[0107] Figure 6 shows an example of each cluster processed by clustering according to S312 and the representative vector of each cluster.

[0108] Referring to Figure 6, for some raw information, clustering processing was performed into clusters 1 to 3, and representative vectors c1, c2, and c3 were extracted for each cluster. That is, the representative vector of cluster 1 is c1, the representative vector of cluster 2 is c2, and the representative vector of cluster 3 is c3.

[0109] In S313, the feature element processing unit (153) of the control unit (150) controls the data that has completed clustering processing to perform additional analysis on raw information with the same clustering label to extract feature elements of the corresponding raw information. At this time, the feature elements correspond to information related to a class that affects vehicle operation or the occurrence of vehicle accidents in the text sentences of the raw information. Of course, feature elements for the raw information belonging to each cluster can be extracted for each cluster.

[0110] These features can be extracted using word-based natural language processing techniques, regular expressions, and named entity recognition (NER), which classifies words within a sentence into different classes. However, since clustering already assumes that data within a cluster are correlated and thus belong to a single cluster, the relationships between each feature are not analyzed separately, thereby reducing processing complexity.

[0111] For example, from the text sentences of raw information, feature elements such as {road type, driving location, driving environment, traffic flow, weather conditions, road conditions, emergency situations, driver characteristics, traffic rules and signal systems, time zone} can be extracted, and their descriptions are as follows.

[0112] - Road types, driving locations, driving environments: {tokens, words, phrases, sentences} containing contents such as city, urban road, rural, unpaved road, off-road, mountain road, highway, residential area, commercial area, etc.

[0113] - Traffic flow: {tokens, words, phrases, sentences} containing contents such as lane changes, intersections, traffic lights, congestion, smoothness, etc.

[0114] - Weather conditions: {tokens, words, phrases, sentences} containing content such as sunshine, backlight, clouds, rain, snow, fog, etc.

[0115] - Road conditions: {tokens, words, phrases, sentences} containing information such as slippery roads, controlled construction zones, and paving work.

[0116] - Emergency situations: {tokens, words, phrases, sentences} containing content such as sudden stops, sudden braking, and unexpected actions by other vehicles.

[0117] - Driver characteristics: {tokens, words, phrases, sentences} containing new drivers, experienced drivers, private taxi drivers, etc.

[0118] - Traffic rules and signal systems: {tokens, words, phrases, sentences} containing information such as child protection zones, no-turn zones, and one-way zones.

[0119] - Time zone: {token, word, phrase, sentence} containing contents such as dawn, morning, noon, afternoon, early evening, evening, etc.

[0120] Additionally, in S313, the feature element processing unit (153) may additionally perform statistical analysis on each feature element for the corresponding cluster for the extracted feature elements.

[0121] For example, with respect to the time zone feature, the frequency of extracted features can be calculated, such as {10 cases in the early morning, 20 cases in the morning, 10 cases in the afternoon, 42 cases in the afternoon, 10 cases in the early evening, 23 cases in the evening}. Of course, if some data (i.e., clustered embedding vectors) are added to a cluster, feature analysis can be performed on the raw information for the added embedding vectors to update the frequency for that content.

[0122] In addition, the feature element processing unit (153) can generate (output) a representative set of feature elements for each cluster as a result of processing feature element extraction. At this time, the representative set of feature elements corresponds to a set of feature elements representing the feature elements extracted from a certain cluster. In other words, a representative set of feature elements can be extracted for each of the K clusters. Of course, a representative set of feature elements can include at least one feature element, and can also include multiple feature elements.

[0123] For example, among the feature elements extracted from raw information belonging to a certain cluster, if it is a set of feature elements that has the highest frequency, or a frequency above a reference value, or is a representative set of feature elements for the cluster based on the results of natural language processing such as TF-IDF, word embedding, or topic modeling, it can be set as a representative set of feature elements for the cluster.

[0124] Meanwhile, various information (i.e., reference information) processed in S310 may be stored and managed in the memory (140) by the reference information management unit (154) of the control unit (150). That is, information processed according to the performance of S311 to S313 may be stored as reference information. This reference information may be utilized as reference material for generating accident-related information for driving description information when performing S320, which will be described later.

[0125] For example, information on embedding vectors, etc. according to the performance of S311, information on clusters, clustering labels, representative vectors, first and second overall representative vectors, etc. according to the performance of S312, and information on feature elements, feature element representative sets, etc. according to the performance of S313 may be stored as reference information in the memory (140). Of course, such reference information may be continuously updated according to the performance of S311 to S313 for additional raw information when additional raw information is input.

[0126] Next, S320 is a step for performing the third function. That is, when driving description information is input in the form of a query, the accident relevance processing unit (155) of the control unit (150) controls the generation and output of accident relevance information for the corresponding driving description information based on the reference information prepared according to the first function.

[0127] To this end, by performing embedding processing on the input driving description information, an embedding vector for the corresponding driving description information is generated (extracted), and by applying the embedding vector of the extracted driving description information to the provided reference information, the amount of accident-related information can be output. Of course, the embedding processing on the driving description information can be applied as described above in S131. At this time, the accident-related information can include first information of a quantitative numerical value for the accident occurrence relevance (i.e., the risk or possibility of a vehicle accident occurring), and may additionally include second information which is text information for the type of vehicle accident with a high accident occurrence relevance. This first information can be obtained using a representative vector, a first overall representative vector, etc., and the second information can be obtained using a feature element, a feature element representative set, etc.

[0128] For example, the accident-related information may include information according to any one of the following methods, or may include information combining multiple methods among the following methods.

[0129] - Method 1:

[0130] The first method calculates the distance between the embedding vector of the driving description information and the first global representative vector of all K clusters and outputs the calculated distance value. In other words, the first information regarding the corresponding distance value (i.e., the first distance value) can be included in the accident relevance information. In this case, a smaller distance value indicates a higher accident relevance (i.e., the risk or possibility of an accident occurring). Of course, the distance between the embedding vector of the driving description information and the second global representative vector can also be calculated and output the calculated distance value (i.e., the second distance value). In this case, the first information regarding the smaller of the first and second distance values ​​can be included in the accident relevance information.

[0131] - Method 2:

[0132] The second method calculates the distance between the embedding vector of the driving description information and each representative vector of each cluster (k clusters) and outputs the calculated distance value. In this case, as the distance for the representative vector of each cluster is calculated, k distance values ​​can be output. At this time, the smaller the distance value, the closer it is to the corresponding cluster, which means that the accident occurrence relevance for the vehicle accident type of the corresponding cluster is high. In other words, the first information for each of the k distance values ​​can be included in the accident relevance information, or the first information for the smallest distance value among the k distance values ​​can be included in the accident relevance information. Of course, in addition, the second information for the representative set of feature elements of the cluster corresponding to the smallest distance value among the k distance values ​​can be included in the accident relevance information.

[0133] - Method 3

[0134] The third method maps the embedding vector of driving description information to clusters (i.e., maps to which cluster it belongs) and outputs information about the cluster to which the driving description information belongs, or a clustering label. At this time, the cluster with the shortest distance between the embedding vector of the driving description information and the representative vectors of each of the K clusters can be mapped to the cluster to which the embedding vector of the driving description information belongs. This mapping of the embedding vector of driving description information to clusters can be applied in the same way in the method described below. Accordingly, secondary information about the mapped cluster or the clustering label of the cluster can be included in the accident relevance information. Accordingly, the accident relevance information can be used to infer which vehicle accident type is highly likely to be involved in an accident. Of course, additionally, primary information about the distance between the representative vector of the mapped cluster and the embedding vector of the driving description information can be included in the accident relevance information. In this case, a smaller distance value indicates a higher relevance for the vehicle accident type of the cluster.

[0135] - Method 4

[0136] The fourth method maps the embedding vector of driving description information to a cluster and outputs a representative set of feature elements of the cluster to which the driving description information belongs. Secondary information regarding the representative set of feature elements of the mapped cluster may be included in the accident relevance information. For example, the driving description information may belong to the kth cluster, and the representative set of feature elements of that cluster may be {rain, skid}. In this case, the accident relevance information of the representative set may be used to determine that the driving description information has a high relevance to the accident occurrence of the feature elements of {rain, skid}. Additionally, primary information regarding the distance between the representative vector of the mapped cluster and the embedding vector of the driving description information may also be included in the accident relevance information. A smaller distance value indicates a higher relevance to the accident occurrence of the vehicle accident type of the corresponding cluster.

[0137] - Method 5

[0138] The fifth method maps the embedding vector of driving description information to clusters, extracts features from the driving description information, and then outputs the results of comparing the representative set of features of the cluster to which the driving description information belongs with the extracted features of the driving description information. In other words, the first information regarding the comparison value can be included in the accident relevance information. Additionally, the second information regarding the representative set of features of the mapped cluster can also be included in the accident relevance information.

[0139] At this time, a similarity measurement method such as Jaccard similarity can be used for comparison between two feature elements, and this Jaccard similarity (J) can be defined by the following mathematical expression 1.

[0140] (Mathematical formula 1) J(A, B)=|A∩B| / |A∪B|

[0141] Here, A and B represent different sets of feature elements, respectively, and the first information about the calculated J value can be included in the accident relevance information.

[0142] For example, the representative set of features for the Kth cluster K may be A={rain, highway, slip, early evening}, and the set of features for the query of driving description information may be B={rain, highway, noon, rural}. In this case, since A∩B={rain, highway}, |A∩B|=2. Also, since A∪B={rain, highway, slip, early evening, noon, rural}, |A∪B|=6. Therefore, the Jaccard similarity (J)=|A∩B| / |A∪B|=2 / 6=1 / 3.

[0143] Of course, when using Jaccard similarity in this way, additionally, second information about feature elements corresponding to A∩B among the representative set of feature elements of the mapped cluster (i.e., common feature elements between the mapped cluster and the driving description information) may be included in the accident relevance information.

[0144] - Method 6

[0145] The sixth method outputs the results of examining whether the embedding vector of the driving description information belongs to a cluster within a threshold level (i.e., within a reference value). That is, the distance between the embedding vector of the driving description information and each representative vector of each cluster (k clusters) is calculated, and if each distance value is within the threshold level, secondary information about the corresponding cluster or the clustering label of the corresponding cluster can be included in the accident relevance information. At this time, secondary information about the representative set of feature elements of the corresponding cluster can also be included in the accident relevance information. Of course, additionally, primary information about the distance between the representative vector of a cluster that falls within the threshold level and the embedding vector of the driving description information can also be included in the accident relevance information. At this time, a smaller distance value indicates a higher relevance to the accident occurrence of the vehicle accident type of the corresponding cluster.

[0146] Figure 7 shows examples of embedding vectors q1 and q2 of driving description information for each cluster according to Figure 6.

[0147] For example, referring to Figure 7, q1 falls within the critical level based on the representative vector c1 for cluster 1, and is therefore determined to be an accident type for cluster 1. On the other hand, q2 is closest to cluster 3, but falls outside the critical level based on the representative vectors of all clusters, and therefore is not determined to be an accident type for any cluster. Of course, according to the third method described above, q2 is closest to cluster 3, and thus cluster 3 may map it.

[0148] However, in S320, prior to performing the third function, information in the form of an image or video related to a vehicle accident may be input as input information. In this case, the control unit (150) may control the second function to be performed first, followed by the third function.

[0149] That is, the input information processing unit (156) of the control unit (150) can control the execution of the second function to extract text-based driving description information from the input information of the corresponding image or video. Accordingly, the input information of the image or video regarding the corresponding vehicle accident can be converted into text-based driving description information, and the second function according to S320 described above can be performed on the extracted driving description information.

[0150] Next, step S330 performs the fourth function. That is, the control unit (150) can control the utilization of the generated accident-related information. In other words, feedback on the accident-related information can be utilized to deliver it to the driver of the target vehicle.

[0151] For example, if the first information in the accident-related information has a numerical value exceeding a reference value (i.e., a numerical value indicating a high risk or possibility of an accident), the control unit (150) can control a warning signal to be transmitted to the driver of the target vehicle. In this case, the warning signal can be transmitted to the user (driver) in the form of visual information through a display within the target vehicle, or in the form of auditory information through a speaker within the target vehicle. Accordingly, the control unit (150) can generate a control signal for the corresponding operation of the display or speaker.

[0152] Additionally, the control unit (150) can control the transmission of text regarding the second information from the accident-related information to the driver of the target vehicle. In this case, the text may be transmitted to the user as visual information via a display within the target vehicle, or as auditory information via a speaker within the target vehicle. Accordingly, the control unit (150) can generate a control signal for the corresponding operation of the display or speaker.

[0153] Of course, the control unit (150) can also control the warning signal for the first information and the text for the second information to be transmitted together to the driver of the target vehicle.

[0154] The present invention, configured as described above, has an advantage in that, when driving description information, which is text information related to the driving situation of a vehicle, is given, accident relevance information that quantitatively indicates the relevance (or risk) of an accident occurrence for the driving description information can be generated.

[0155] In particular, conventional techniques that extract and analyze features (e.g., weather, time zone, location, etc.) related to accident occurrence from input sentences (i.e., text information) require the assumption that each extracted element is statistically independent or conditionally independent. This assumption simplifies the analysis and facilitates the application of various statistical techniques. However, conventional techniques cannot properly analyze situations that do not meet this assumption.

[0156] In contrast, the present invention uses a nonlinear artificial neural network technique to convert driving description information, which is text information related to the driving situation of a vehicle, into a high-dimensional embedding vector, process it, cluster it, and analyze the feature elements. Accordingly, the present invention can better understand the relationships between feature elements without the aforementioned assumptions, can be applied to situations that are not possible to analyze with conventional techniques, and has the advantage of increasing the simplicity of the calculations required to derive accident-related information because calculations only need to be performed on newly input driving description information.

[0157] Furthermore, the present invention can generate quantitative figures on accident relevance information in various ways based on clustered accident cases, thereby increasing the usability of the generated accident relevance information. Furthermore, when an image or video is input, the present invention extracts text-based driving description information from the image or video to generate accident relevance information. Therefore, by identifying the relevance of an accident based on the current driving status of a vehicle in real time through a black box or the like, and providing feedback to the vehicle, the present invention has the advantage of increasing the driving safety of the vehicle.

[0158] While the detailed description of the present invention has described specific embodiments, it should be understood that various modifications are possible without departing from the scope of the present invention. Therefore, the scope of the present invention is not limited to the described embodiments, but should be determined by the claims and their equivalents.

[0159] The present invention relates to an accident-related information processing technology, and when driving description information, which is text information related to a driving situation of a vehicle, is given, accident-related information that quantitatively indicates the accident occurrence relevance of the driving description information can be generated, and thus has industrial applicability.

Claims

1. A method performed by an electronic device, A step of generating an embedding vector through embedding processing of raw information, which is text information describing the accident situation of another vehicle; A step of classifying the generated embedding vector into multiple clusters through clustering processing; A step of extracting feature elements related to the occurrence of a vehicle accident for each cluster and raw information belonging to the cluster; and When driving description information, which is text information related to the driving situation of the target vehicle, is input as input information, a step of generating an embedding vector for the driving description information, and then generating accident-related information for the target vehicle according to the driving description information based on the cluster, the feature elements, and the embedding vector of the generated driving description information is included; A method wherein the above accident relevance information includes first information in the form of a quantitative numerical value regarding the risk or possibility of an accident occurring for the target vehicle.

2. In paragraph 1, It further includes a step of setting a representative vector for each cluster for the embedding vectors belonging to each cluster. A method for generating the first information by using the distance between the representative vector of each cluster and the embedding vector of the driving description information in the step of generating the above accident relevance information.

3. In paragraph 2, In the above setting step, a method of setting one of the mean, maximum density, center, mode, maximum and minimum values ​​of the embedding vectors forming each cluster as the representative vector of the cluster.

4. In paragraph 2, In the above setting step, a method of calculating multiple values ​​among the mean, maximum density, center, mode, maximum and minimum values ​​for the embedding vectors forming each cluster, and setting the result of the weighted average of the multiple values ​​calculated as the representative vector.

5. In paragraph 2, In the above setting step, after dimension reduction of the vector space according to the clustering process, the center of the reduced vector space is set as the representative vector. A method for generating the accident relevance information by, in the step of generating the accident relevance information, restoring the dimensionally reduced representative vector to the vector space size of the embedding vector of the driving description information, or reducing the dimension of the embedding vector of the driving description information by the size of the dimensionally reduced representative vector, and comparing the embedding vector of the driving description information with the representative vector.

6. In paragraph 1, It further includes a step of setting an overall representative vector for all the above clustered embedding vectors, A method for generating the first information by using the distance between the entire representative vector and the embedding vector of the driving description information in the step of generating the above accident relevance information.

7. In paragraph 1, The above accident relevance information further includes second information in the form of text information about the type of vehicle accident that has a risk or possibility of causing the above accident, The second information includes a characteristic element related to the first information, A method further comprising the step of generating a control signal for transmitting text about the second information to the driver of the target vehicle together with a warning signal when the first information has a numerical value exceeding a reference value.

8. In paragraph 1, It further includes a step of extracting a representative set of feature elements, which is a set representing the feature elements extracted from each cluster, for each cluster. A method for generating the accident relevance information, wherein, in the step of generating the accident relevance information, the embedding vector of the driving description information is mapped to one of each cluster, and the accident relevance information includes a representative set of feature elements for the mapped cluster.

9. In paragraph 1, It further includes a step of extracting a representative set of feature elements, which is a set representing the feature elements extracted from each cluster, for each cluster. A method for generating the accident relevance information, wherein, in the step of generating the accident relevance information, the embedding vector of the driving description information is mapped to one of each cluster, feature elements are extracted from the driving description information, and then the accident relevance information is generated by comparing a representative set of feature elements of the mapped cluster with the feature elements of the driving description information.

10. In paragraph 2, The above accident relevance information further includes second information in the form of text information about the type of vehicle accident that has a risk or possibility of causing the above accident, A method for generating the accident relevance information, wherein the method further includes second information on a feature element related to a cluster corresponding to a distance within a reference value among the distances between the representative vector of each cluster and the embedding vector of the driving description information in the step of generating the accident relevance information.

11. In paragraph 1, A method further comprising a step of extracting driving description information, which is text information describing the driving situation, from the image or video, when the input information is an image or video of a driving image of the target vehicle.

12. Memory storing raw information, which is text information describing the situation of an accident involving another vehicle; and A control unit that performs control using information stored in the above memory; The above control unit, Control to generate an embedding vector through embedding processing of the above raw information, Controls the classification into multiple clusters through clustering processing on the generated embedding vector. Controls the extraction of feature elements related to vehicle accident occurrence for raw information belonging to each cluster, When driving description information, which is text information related to the driving situation of the target vehicle, is input as input information, an embedding vector for the driving description information is generated, and then, based on the cluster, the feature elements, and the generated embedding vector of the driving description information, accident-related information for the target vehicle according to the driving description information is generated. A device wherein the above accident relevance information includes first information in the form of a quantitative numerical value regarding the risk or possibility of an accident occurring for the target vehicle.

Images