Accident inference device, insurance fee examination device, on-vehicle equipment, terminal device, information processing system, accident inference method, accident inference program, ai model generation method, and ai generation program

Abstract

An accident inference device according to one embodiment of the present invention is provided with a controller. When a near-miss occurs, the controller infers an accident that is likely to happen in said near-miss situation, on the basis of the situation of the near-miss. The controller outputs accident information related to the inferred virtual accident.

Description

Accident estimation device, insurance premium review device, in-vehicle device, terminal device, information processing system, accident estimation method, accident estimation program, AI model generation method, and AI generation program

[0001] The disclosed embodiments relate to an accident estimation device, an insurance premium review device, an in-vehicle device, a terminal device, an information processing system, an accident estimation method, an accident estimation program, an AI model generation method, and an AI generation program.

[0002] There is a danger point information display device that contributes to traffic safety by notifying or presenting to vehicle drivers and others accident information such as the location, details, and surrounding environment of past accidents, as well as information on risk factors at the location of near misses that could lead to accidents (see, for example, Patent Document 1).

[0003] JP 2007-51973 A

[0004] However, because near misses do not become accidents, information about near misses is somewhat less stimulating. Also, information about past accidents and past accidents and near misses is something that does not concern those other than those involved in the accidents. For these reasons, it is insufficient as a teaching material to improve the effectiveness of traffic safety education.

[0005] One aspect of the embodiment has been made in consideration of the above, and aims to provide an accident estimation device, an insurance premium review device, an in-vehicle device, a terminal device, an information processing system, an accident estimation method, an accident estimation program, an AI model generation method, and an AI generation program that can contribute to improving the effectiveness of traffic safety education.

[0006] According to one aspect of the embodiment, the accident estimation device includes a controller. When an attempted accident, known as a near miss ("a dangerous event that fortunately did not result in a disaster") occurs, the controller estimates an accident likely to occur in the near miss situation based on the situation. The controller outputs accident information relating to the accident.

[0007] According to one aspect of the embodiment, an accident estimation device, an insurance premium review device, an on-board device, a terminal device, an information processing system, an accident estimation method, an accident estimation program, an AI model generation method, and an AI generation program estimate and generate information on accidents that are likely to occur in response to near misses that involve the user. Therefore, it is possible to realize information notification that is relevant to the user, thereby contributing to improving the effectiveness of traffic safety education.

[0008] FIG. 1 is an explanatory diagram showing the configuration of an accident estimation device and a vehicle according to an embodiment. FIG. 2 is an explanatory diagram showing a method for generating an accident estimation AI model according to an embodiment. FIG. 3 is an explanatory diagram showing a method for creating training data according to an embodiment. FIG. 4 is an explanatory diagram showing the configuration of an accident information generation unit according to an embodiment. FIG. 5 is an explanatory diagram showing an example of a table stored in a video DB according to an embodiment. FIG. 6 is an explanatory diagram showing an example of a table stored in a video DB according to an embodiment. FIG. 7 is an explanatory diagram showing an example of a table stored in a video DB according to an embodiment. FIG. 8 is an explanatory diagram showing an example of a table stored in an accident DB according to an embodiment. FIG. 9 is an explanatory diagram showing the configuration of an accident information generation unit equipped with AI according to an embodiment. FIG. 10 is an explanatory diagram showing an example of a training dataset for a video information generation AI model according to an embodiment. FIG. 11 is an explanatory diagram showing an example of a training method for a video information generation AI model according to an embodiment. FIG. 12 is an explanatory diagram showing an example of the operation of a video information generation AI model according to an embodiment. FIG. 13 is an explanatory diagram showing an example of a training dataset for a damage information generation AI model according to an embodiment. FIG. 14 is an explanatory diagram showing an example of a training method for a damage information generation AI model according to an embodiment. Fig. 15 is an explanatory diagram showing an example of the operation of the damage information generation AI model according to the embodiment. Fig. 16 is a flowchart showing processing executed by a controller of the accident estimation device according to the embodiment. Fig. 17 is a flowchart showing processing executed by a controller of the accident estimation device according to the embodiment. Fig. 18 is an explanatory diagram of an information processing system according to the embodiment.

[0009] Hereinafter, with reference to the accompanying drawings, embodiments of an accident estimation device, an insurance premium review device, an in-vehicle device, a terminal device, an information processing system, an accident estimation method, an accident estimation program, an AI model generation method, and an AI generation program will be described in detail. Note that the present invention is not limited to the embodiments shown below.

[0010] [1. Overview of the accident estimation method] First, an overview of the accident estimation method performed by the accident estimation device according to the embodiment will be described. The accident estimation device includes a computer, and the computer executes an accident estimation program stored in a memory to realize the accident estimation method according to the embodiment.

[0011] The accident estimation device according to the embodiment is a device that, when a near-miss accident (hereinafter referred to as a "near miss") that is likely to develop into an accident occurs in a traveling vehicle, estimates an accident that is likely to occur in the circumstances of the near miss and outputs accident information related to the hypothetical accident, i.e., hypothetical accident information based on the near miss. The accident estimation device estimates an accident that is likely to occur based on the circumstances of the near miss.

[0012] In this way, the accident estimation device outputs and provides accident information about accidents that are likely to occur as a result of the near miss to drivers who have actually experienced a near miss, making them aware of the danger of accidents. Therefore, drivers are provided with accident information that is relevant to them, and the accident estimation device can contribute to improving the effectiveness of traffic safety education.

[0013] 2. Accident estimation device An accident estimation device 1 according to an embodiment will be described with reference to Fig. 1. Fig. 1 is an explanatory diagram showing the configuration of the accident estimation device 1 and a vehicle 2 according to an embodiment. First, the configuration of the vehicle 2 according to the embodiment will be described.

[0014] 1 , a vehicle 2 is equipped with a drive recorder 20 having a communication function with an external device. The drive recorder 20 is connected to an accident estimation device 1 via a communication network N such as the Internet so as to be able to communicate information with the vehicle 2. A steering angle sensor 22 and an emotion sensor 23 installed in the vehicle 2 are also connected to the drive recorder 20 via an in-vehicle local area network (LAN). The drive recorder 20 includes an acceleration sensor 21, an in-vehicle camera 24, an in-vehicle microphone 25, a flash memory 26, a controller 27, and a communication unit 28.

[0015] The controller 27 acquires steering angle information of the vehicle detected by the steering angle sensor 22 and emotion information of the driver detected by the emotion sensor 23. The controller 27 also acquires acceleration of the vehicle detected by the acceleration sensor 21, image information captured by the in-vehicle camera 24, and audio information collected by the in-vehicle microphone 25.

[0016] The controller 27 stores the image information captured by the in-vehicle camera 24 and the audio information collected by the in-vehicle microphone 25 in the flash memory 26. The controller 27 also reads out the image information and audio information stored in the flash memory 26, uses them for various processes, and transmits them to an external device such as the accident estimation device 1 via the communication unit 28.

[0017] The acceleration sensor 21 is a sensor that detects the acceleration of the vehicle, and is made up of, for example, an elastic body and a mass body, and detects the acceleration by measuring the force applied to the mass body due to acceleration as the displacement of the elastic body, and can be configured as an acceleration sensor of a frequency change type, piezoelectric type, piezo-resistance type, capacitance type, etc. The steering angle sensor 22 is a sensor that detects the rotation state of the steering wheel of the vehicle 2.

[0018] The emotion sensor 23 is a sensor that detects biological information to estimate the driver's emotion, such as a heart rate sensor or an electroencephalogram sensor, and the emotion is estimated based on the electroencephalogram and heart rate detected by these sensors. The heart rate sensor or the like is provided on the steering wheel or the like of the vehicle 2. Note that a method of estimating emotion from the driver's facial information (facial expression) using artificial intelligence or the like can also be applied. In this case, the in-vehicle camera 24 can be used as a sensor to acquire the driver's facial information.

[0019] The in-vehicle camera 24 includes an interior imaging device that captures video images inside the cabin of the vehicle 2, and an exterior imaging device that captures video images around the vehicle 2. The in-vehicle microphone 25 is a sound collecting device that collects audio from inside the cabin of the vehicle 2. The flash memory 26 is a recording device that stores the video images inside the cabin captured by the in-vehicle camera 24 and the audio from inside the cabin collected by the in-vehicle microphone 25, as well as the video images around the vehicle 2 captured by the in-vehicle camera 24 and the audio around the vehicle 2 collected by the in-vehicle microphone 25.

[0020] The controller 27 includes various circuits and a microcomputer having a CPU (Central Processing Unit), ROM (Read Only Memory), RAM (Random Access Memory), etc. The controller 27 includes a near-miss detection unit 29 that functions when the CPU executes a program stored in the ROM using the RAM as a work area.

[0021] The controller 27 may be partially or entirely configured with hardware such as an ASIC (Application Specific Integrated Circuit) or an FPGA (Field Programmable Gate Array). The communication unit 28 is a communication interface that communicates information with the accident estimation device 1 via the communication network N.

[0022] Next, we will explain the operation of each component of the vehicle 2. The acceleration sensor 21 detects acceleration in the front-rear, left-right, up-down directions of the vehicle 2, assuming that the forward direction of the vehicle 2 is the front-rear direction, and outputs the detected acceleration to the controller 27.

[0023] The steering angle sensor 22 uses the state of the steering wheel when the vehicle 2 is traveling straight as a reference position, detects the rotation angle, rotation speed, and rotation acceleration of the steering wheel from the reference position, and outputs them to the controller 27. Note that the rotation speed and rotation acceleration of the steering wheel may be calculated by the controller 27 through arithmetic processing (differential processing and second-order differential processing) based on the rotation angle.

[0024] The emotion sensor 23 detects the driver's heartbeat from the pulsation of the driver's palms gripping the steering wheel. When a near miss occurs, the driver's emotions change suddenly, and so do biosignals correlated with emotions, such as heartbeat. Therefore, the drive recorder 20 uses the heartbeat to estimate emotions (emotional state at the time of the near miss). The emotion sensor 23 detects the driver's heartbeat and outputs it to the controller 27. Note that the emotion sensor 23 can also be a sensor that detects biosignals other than heartbeat that are correlated with emotions, such as brain waves.

[0025] The in-vehicle camera 24 outputs to the controller 27 images of the interior of the vehicle 2 and images of the surroundings of the vehicle 2. The in-vehicle microphone 25 collects audio from the interior of the vehicle 2 and outputs the audio to the controller 27.

[0026] The controller 27 stores the video input from the in-vehicle camera 24 and the audio input from the in-vehicle microphone 25 in the flash memory 26. The controller 27 associates the input video with the date and time when the video was captured, and the input audio with the date and time when the audio was collected, and stores these in the flash memory 26.

[0027] The controller 27 may also store the steering wheel rotation angle, rotation speed, and rotation acceleration detected by the steering angle sensor 22, as well as emotion information (heart rate information) detected by the emotion sensor 23, in the flash memory 26, in association with the date and time.

[0028] The near-miss detection unit 29 detects a near-miss that occurs while the vehicle 2 is traveling. The near-miss detection unit 29 detects a near-miss based on at least one of the acceleration of the vehicle 2 input from the acceleration sensor 21, the rotation angle, rotation speed, and rotation acceleration of the steering wheel input from the steering angle sensor 22, and the driver's heart rate.

[0029] The near-miss detection unit 29 determines that a near-miss has occurred when the acceleration of the vehicle 2 exceeds an acceleration threshold. In other words, when a near-miss occurs, sudden braking or abrupt steering is often performed, and some kind of impact is often applied even if it does not result in an accident, so the occurrence of a near-miss is detected based on the detection status of acceleration.

[0030] The lower threshold of acceleration for determining whether a near miss has occurred is the upper limit (with an appropriate offset) of the acceleration of the vehicle 2 that can occur when the vehicle 2 is being driven safely, and the upper threshold is the lower limit (with an appropriate offset) of the acceleration when an accident occurs, and is set based on experiments during design and development, for example. The deployment status of the airbags of the vehicle 2 may also be used to determine whether a near miss has occurred (when the airbags are deployed, it is determined that an accident has occurred but not a near miss).

[0031] Furthermore, the near-miss detection unit 29 determines that a near-miss has occurred when the rotational speed and rotational acceleration of the steering wheel exceed the corresponding rotational speed threshold and rotational acceleration threshold, respectively. In other words, when a near-miss occurs, the driver often makes an abrupt turn of the steering wheel, so it is determined that a near-miss has occurred when the rotational speed and rotational acceleration of the steering wheel are greater than the threshold. The rotational speed threshold and rotational acceleration threshold are set to values ​​that provide good determination accuracy, for example, based on experiments during design and development.

[0032] The rotational speed threshold and rotational acceleration threshold are the upper limit values ​​(with appropriate offsets) of the rotational speed and rotational acceleration of the steering wheel operated by the driver when the vehicle 2 is being driven safely, and are set to values ​​that provide good judgment accuracy based on experiments, for example, during design and development.

[0033] The near-miss detection unit 29 determines that a near-miss has occurred if the increase in the driver's heart rate within the determination time exceeds the heart rate threshold. The determination time here is, for example, 10 seconds. Note that the determination time here is not limited to 10 seconds. The heart rate threshold is the upper limit (with an appropriate offset) of the increase in the driver's heart rate within the determination time while the vehicle 2 is being driven safely. Note that the determination time and the heart rate threshold are set to appropriate values ​​that provide good determination accuracy, for example, based on experiments during design and development.

[0034] In addition, the determination of the occurrence of a near miss may be made by a combination of the above-mentioned determination based on the acceleration state of the vehicle 2, the determination based on the steering wheel rotation state, and the driver's heart rate (emotional) state (determination based on the results of logical calculations of these determination results).

[0035] In this way, if the occurrence of a near miss can be detected by the near miss detection unit 29, information on the location where the near miss occurred and the circumstances of the near miss can be provided to, for example, an institution that provides traffic safety education.The institution that provides traffic safety education can then carry out traffic safety awareness activities by, for example, creating a near miss map that clearly shows the locations where near misses occurred and the circumstances of the near miss and providing it to drivers, etc.

[0036] However, drivers do not make a strong impression when they see a near miss, partly because no accident actually occurs. For example, when a near miss occurs, they do not know what kind of accident it could have become, the extent of the damage, or the percentage of responsibility for the accident, so it is somewhat unrealistic.

[0037] Meanwhile, insurance that compensates for damages caused by accidents can be divided into distance-linked (PAYD) insurance, where the premium is determined according to the distance driven, and behavior-linked (PHYD) insurance, where the premium is linked to driving operations. PHYD uses information from various sensors in the in-vehicle device to determine whether the driver is driving safely based on the number and frequency of sudden braking, sudden acceleration, sudden steering, etc. Compared to PAYD, PHYD is said to be able to grasp the risk of driving accidents more accurately and reflect this in insurance premiums.

[0038] As such, insurance companies are making various efforts to set appropriate premiums. However, the information used to set premiums is mainly information about accidents that have occurred, and the reality is that information about near misses that did not result in accidents is not being used sufficiently.

[0039] For example, drivers who have a high frequency of near misses are predicted to have a high probability of causing an accident, and drivers who have a high frequency of near misses that lead to accidents causing major damage are predicted to have a high probability of causing a major accident, but the reality is that this near miss information is not used to calculate insurance premiums.

[0040] Therefore, when the near-miss detection unit 29 according to the embodiment determines that a near-miss has occurred, it transmits near-miss information related to the near-miss to the accident estimation device 1. The accident estimation device 1 estimates a likely accident scene based on the near-miss information received from the near-miss detection unit 29 of the vehicle 2, and outputs accident information related to the accident scene.

[0041] For example, the accident estimation device 1 outputs accident type information indicating the type of accident concerning an accident that is feared to occur from a near miss, and hypothetical accident information indicating an accident event that is feared to occur from a near miss.

[0042] This allows the driver, for example, to check the estimated accident information and find out what kind of accident could have occurred when a near miss occurs, the extent of the damage, the percentage of responsibility for the accident, and the severity and impact of the accident if the near miss escalates into an accident.

[0043] On the other hand, by checking the estimated accident information, insurance companies can estimate the potential probability of an accident occurring and the scale of the accident (scale of damage), which is expected to enable them to calculate more appropriate insurance premiums.

[0044] Specifically, when the near miss detection unit 29 determines that a near miss has occurred, it reads from the flash memory 26 the date and time data and location data of the near miss, as well as the video and audio recorded a predetermined time before and after the date and time of the near miss.

[0045] Furthermore, the near-miss detection unit 29 reads the acceleration data detected by the acceleration sensor 21, the steering speed data detected by the steering angle sensor 22, and the driver's emotion data detected by the emotion sensor 23 from the flash memory 26.

[0046] The specified time here is the length of time for various pieces of information necessary to understand the characteristics of the near-miss occurrence situation, for example, 5 seconds, but it is not limited to 5 seconds and can be an appropriate time determined based on experiments, etc. during the design and development of the product system.

[0047] The near-miss detection unit 29 outputs these various data when a near-miss occurs to the communication unit 28. The communication unit 28 transmits these various data when a near-miss occurs, which are input from the controller 27, to the accident estimation device 1 as near-miss information.

[0048] Next, we will explain the configuration of the accident estimation device 1. The accident estimation device 1 includes a controller 10, a communication unit 11, and an accident database 12 (hereinafter referred to as "accident DB 12"). The communication unit 11 is a communication interface that communicates information with the vehicle 2 via the communication network N.

[0049] The controller 10 includes a microcomputer having a CPU, ROM, RAM, etc., and various circuits. The controller 10 is equipped with an accident estimation model 13 that functions when the CPU executes a program stored in the ROM using the RAM as a work area, and an accident information generation unit 14.

[0050] The controller 10 may be configured partially or entirely with hardware such as an ASIC or FPGA. The accident DB 12 is an information storage device such as a data flash, and stores information about accidents that have occurred in the past. An example of the information stored in the accident DB 12 will be described later with reference to FIG. 4.

[0051] Next, a description will be given of the operation of each component of the accident estimation device 1. When the communication unit 11 receives near-miss information from the vehicle 2, it outputs the received near-miss information to the accident estimation model 13 of the controller 10.

[0052] The accident inference model 13 is an inference model that, when near-miss information is input, outputs an accident scene that is likely to occur in the near-miss situation corresponding to the near-miss information. The accident inference model 13 can be realized by a method such as a matching process using a database (accident type database) made up of accident scene information (information that identifies the type of accident) using various information from the near-miss information as parameters, but because accident scenes come in a wide variety of patterns (types), it is more realistic to realize it using an AI (Artificial Intelligence) model.

[0053] This AI model is an AI model that has been machine-learned to estimate accident scenes based on input near-miss information. For ease of understanding, the accident estimation model 13 using the AI ​​model will be referred to as the accident estimation AI model 13A.

[0054] The circumstances of the accident (such as the level of damage) are estimated based on the relative speed with respect to the object colliding with, the direction of the collision (with respect to the vehicle (e.g., the forward direction of the vehicle) as the reference point), the point of collision, etc. The circumstances of the accident (such as the level of damage) also change depending on the environment around the accident site; for example, accident footage, etc., changes significantly depending on the surrounding scenery.

[0055] Therefore, increasing the number of types of parameters used in the estimation enables more detailed estimation that takes into account the data for each item. By increasing the types of information input to the accident estimation AI model 13A and the level of detail of that information, and by training the accident estimation AI model 13A with a large amount of detailed data, the accident estimation AI model 13A can perform more detailed estimation.

[0056] However, in practice, it is necessary to select appropriate items and accuracy for the data used for estimation, taking into consideration the performance of the device (CPU, etc.) and the AI ​​learning time. For this reason, and for ease of understanding, this embodiment will be described using an example in which the data items used for estimation are limited to items with a large impact. Note that even if other items are used, the accident situation can be estimated using the same processing.

[0057] The accident estimation AI model 13A is a model that estimates the accident situation based on near-miss information. This model estimates the accident situation that will occur if appropriate avoidance actions are not taken in the near-miss that has occurred, based on the similarity between the situation immediately before the near-miss information occurred (before the accident avoidance action was taken) and the situation immediately before the accident occurred (the time required for the accident avoidance action).

[0058] Therefore, the learning data (accident data) during learning of the accident estimation AI model 13A has input data as "status data such as vehicle driving immediately before the accident" and correct answer data as "accident result (accident situation)." Therefore, vehicle status data for an appropriate time period (an appropriate time is set based on experiments, etc.) ending at an appropriate time before the accident occurs (an appropriate time is set based on experiments, etc.) becomes input data for the learning data.

[0059] Hereinafter, this time period will be referred to as the “estimated data time period.” Furthermore, the input data when using the accident estimation model 13 (estimating operation) is “operational state data before the avoidance operation in the near miss,” and in reality, just as during learning, the input data for the learning data is the vehicle state data for the estimated data time period (an appropriate time period (an appropriate time period is set based on experiments, etc.) ending at an appropriate time before the near miss occurs (an appropriate time period is set based on experiments, etc.)).

[0060] In this embodiment, the near-miss information includes the detection result (acceleration sensor information) of the acceleration sensor 21 during the estimated data time period when the near-miss occurred. The near-miss information also includes the detection result (steering angle sensor information) of the steering angle sensor 22 during the estimated data time period.

[0061] The near-miss information includes the detection result (emotion sensor information) of the emotion sensor 23 during the estimated data time period when the near-miss occurred. The near-miss information includes an image (in-vehicle camera video) captured by the in-vehicle camera 24 during the estimated data time period. The near-miss information includes audio (in-vehicle microphone audio) picked up by the in-vehicle microphone 25 during the estimated data time period.

[0062] Here, a generation (learning) method of the accident estimation AI model 13A will be described with reference to Fig. 2. Fig. 2 is an explanatory diagram showing a generation method of the accident estimation AI model 13A according to an embodiment. As shown in Fig. 2, the accident estimation AI model 13A includes a DNN (Deep Neural Network) 13B that performs predictive calculations from input values ​​to derive output values.

[0063] The accident estimation AI model 13A is generated, for example, by supervised learning. In supervised learning, a learning data set is prepared that is composed of a large number of learning data in which input values ​​and correct answers are paired. In the case of the learning data of the accident estimation AI model 13A, predictive information, which is information similar to near-miss information immediately before an accident or near-miss occurs, is used as an input value, and the accident scene of the accident (or an accident predicted to occur from the near-miss) is used as a correct answer.

[0064] The predictive information to be input includes, as described above, acceleration sensor information, steering angle sensor information, emotion sensor information, in-vehicle camera footage, and in-vehicle microphone audio during the estimated data time period at the time of the accident. The correct answer values ​​for the accident scenes are each associated with identification information (accident scene ID) that identifies the respective accident scenes.

[0065] Next, three examples of how to create this supervised learning data will be described.

[0066] First learning data creation method: Near-miss predictive states in the near-miss occurrence process are extracted based on near-miss information, and accident information on accidents with accident predictive states similar to the near-miss predictive states is extracted. Learning data is then created using the predictive information corresponding to the near-miss predictive states as input data, with the corresponding accident information as correct answer data.

[0067] Specifically, as shown in Figure 3, the correct data is accident information (accident scene) that includes predictive information similar to that of a near miss (information similar to the information immediately before the accident occurs in the event information) during an appropriate period before the accident avoidance operation (sudden deceleration, sudden steering, avoidance operation by the other vehicle (determined from images, etc.)) in the process of the near miss occurrence. Note that the determination of the time of the accident avoidance operation can be made based on the vehicle's running (operation) state (speed sensor information, steering angle sensor information, etc.), and the data collection period can be determined at the time of the accident avoidance operation, but as explained above, the estimated data time period when the near miss occurs (time of the accident avoidance operation) can also be determined by experiment, etc.

[0068] In other words, since it is believed that an accident will occur if no action (operation) to avoid an accident is taken in the event of a near miss, accident information corresponding to the near miss information is extracted by comparing the similarity of data in a state in which no action to avoid an accident is taken. Note that this learning data creation method can be created manually (by looking at near miss information and accident information and creating it based on the above-mentioned perspective) or by computer processing (by comparing the similarity of near miss information and accident information based on the above-mentioned perspective).

[0069] Second learning data creation method: The second learning data creation method extracts accident precursor states (information) in the accident occurrence process based on accident information of actual accidents, and extracts near-miss information of near-misses that have near-miss precursor states (information) similar to the accident precursor states in question.The extracted near-miss information is then used as input data to create learning data with the corresponding accident information as correct answer data.

[0070] Specifically, as shown in Figure 3, accident information is used as correct answer data, and near-miss information for near misses containing information similar to accident prediction information for an accident prediction period (a non-avoidance state period before the accident (a period during which the accident could have been prevented if evasive action had been taken, determined, for example, by an incident)) ending an appropriate time before the accident occurred in the accident corresponding to the accident information is used as input data. In other words, since it is considered that an accident can be limited to a near miss if accident avoidance action (operation) is taken in the event of an accident, near-miss information corresponding to the accident information is extracted by comparing the similarity of data before the non-avoidance state.

[0071] This learning data creation method can also be done manually (by looking at near-miss information and accident information and creating it based on the above-mentioned perspective) or by computer processing (by comparing the similarities between near-miss information and accident information based on the above-mentioned perspective).

[0072] Third learning data creation method: Near-miss information before evasive maneuvers are extracted from near-miss information during the near-miss occurrence process. Then, vehicle behavior if the behavior before the evasive maneuver is maintained is predicted (and the behavior of the other vehicle is also predicted if there is an other vehicle), and accident information is predicted and created. Data for predictive generation of accident information (various data for predictive generation of accident images and vehicle damage conditions) is created based on past accident information. For example, accident images are created using computer graphics technology using image components of vehicle images and landscape images, as well as various mechanical formulas and coefficients related to object movement.

[0073] In this embodiment, an identification code is assigned to each accident and the identification code is used as the correct answer data. Various information about each accident is managed and stored in a database in which the various information is stored in a data record using the identification code as key data. In this case, the correct answer data in the learning data is the identification code for each accident.

[0074] In the case of the accident estimation model 13 using an accident type database, for example, a group of information associating input data in the learning data with corresponding correct answer data becomes the accident type database.

[0075] For example, a certain accident scene ID is associated with an accident scene (various data in the accident scene) called "a collision caused by an oncoming vehicle making an unreasonable right turn." Other accident scenes include, for example, "a collision with an oncoming vehicle going straight due to an unreasonable right turn," "a rear-end collision with a vehicle in front due to sudden braking by the vehicle in front," "a rear-end collision with a vehicle in front due to inattention to the road ahead," and "a collision caused by unreasonable overtaking."

[0076] The DNN 13B sequentially derives estimated values ​​of accident scenes that are likely to occur from the near-miss information (sign information) of each piece of learning data in the sequentially input learning data set, and outputs the estimated values ​​as estimation results. The controller 10 causes the accident estimation AI model 13A to perform machine learning by updating weighting coefficients, etc. of each layer in the DNN 13B using processing such as backpropagation so as to reduce the error between the estimated results of the accident scene output from the DNN 13B and the correct values ​​of the learning data corresponding to the input values.

[0077] In other words, by executing the AI ​​generation program, the controller 10 performs machine learning on the accident estimation AI model 13A so that the accident scene estimated from the information about the near-miss situation input to the accident estimation AI model 13A becomes an accident scene (correct data) that occurred in the input near-miss situation.

[0078] This allows the controller 10 to generate an accident estimation model 13 that can estimate, from the input near-miss information, an accident scene that is likely to occur in the near-miss situation corresponding to the near-miss information.

[0079] 1 , the controller 10 outputs the near-miss information that has occurred in the vehicle 2 and that has been input via the communication unit 11 to the accident estimation model 13 configured using the accident estimation AI model 13A generated in this way, etc. Then, the controller 10 estimates an accident scene that is likely to occur in the near-miss situation through estimation processing of the accident estimation model 13.

[0080] That is, when near-miss information is input, the accident inference model 13 infers an accident scene that is likely to occur in the near-miss situation corresponding to the near-miss information, and outputs the inference result to the accident information generation unit 14. At this time, the accident inference model 13 outputs the near-miss information used to infer the accident to the accident information generation unit 14 together with the accident scene ID of the inferred accident scene.

[0081] The accident information generation unit 14 generates accident information about the accident scene estimated by the accident estimation model 13 and outputs the information to the information output device 15. The information output device 15 is a display capable of displaying an image of the accident information. The information output device 15 also includes a speaker capable of outputting the audio of the accident information. The information output device 15 also includes a camera that captures images of viewers watching the accident information.

[0082] The information output device 15 outputs a captured image of the viewer viewing the accident information to the controller 10 of the accident estimation device 1. The information output device 15 may be configured to be connected to a heart rate sensor that measures the heart rate of the viewer viewing the accident information. In this case, the information output device 15 outputs information indicating the heart rate of the viewer to the controller 10.

[0083] Next, the configuration and operation of the accident information generation unit 14 will be described with reference to Figures 4 to 8. As shown in Figure 4, the accident information generation unit 14 includes a video information generation unit 31 and a damage information generation unit 32. The video information generation unit 31 includes a video database (DB) 33, a video generation model 34, and an allowance determination unit 35.

[0084] Next, a method for generating an accident video will be described. First accident video generation method: The video DB 33 stores live-action video, VR (Virtual Reality) video, and CG (Computer Graphics) video of various accident scenes. For example, the video DB 33 stores data such as that shown in FIG. 5, in which accident scene IDs are associated with accident videos. The data for the video DB 33 is collected and created in advance by design developers and the like, and stored in the video DB 33. In addition to the accident video, data such as accident audio may also be stored, so that the accident audio is played back synchronously when the accident image is played back.

[0085] The video generation model 34 searches the data in the video DB 33 using the accident scene ID input from the accident estimation model 13, and extracts the accident video corresponding to the near-miss information. The video generation model 34 uses this extracted accident video as a simulation video (image) of an accident scene that is likely to occur in the near-miss situation encountered by the vehicle 2. Note that the video of the near-miss encountered by the vehicle 2 may be joined with the simulation video of this accident scene to create a simulation video of the accident scene. In this case, the image contains real images (video), which increases realism and results in an effective image.

[0086] Second accident video generation method: The first accident video generation method has the problem of large data volume because accident video for each accident type, taking into consideration the surrounding environment such as the background of the accident location, is stored in the video DB 33. The second accident video generation method generates accident video by synthesizing a database of accident video, including video of the accident itself, that is, video of the accident vehicle and accident objects (people, objects, etc.), which are the objects involved in the accident, and background video, in this case, unchanging video (unchanging video over the duration of the accident, for example, images of still objects such as buildings) and changing video (people, animals, displays and traffic lights (changes due to displayed images or flashing)).

[0087] For example, the video generation model 34 performs image recognition on the background image of a stationary object in the video captured by the onboard camera 24 in the near-miss information, such as information about the shape of an intersection, and selects the near-miss location and background image, such as a background video with a similar intersection shape, from the video DB 33 based on the recognized information. The information about the intersection shape here includes, for example, information about the road configuration, such as a T-junction, the number of lanes, and information indicating whether or not there is a traffic light. Note that a stationary object can be determined, for example, by comparing the moving speed of the object in the video with the moving speed of the vehicle filming it (if the moving speeds of both are the same, the object is still).

[0088] In addition, the video generation model 34 performs image recognition from the video of the in-vehicle camera 24 to determine whether or not there are people who are assumed to have been involved in the near miss, and to obtain party information regarding their location, and based on the recognized party information, selects accident video (video of the parties involved that does not include background, etc.) from the video DB 33 that contains party information similar to the situation of the parties involved in the near miss.

[0089] Furthermore, the video generation model 34 performs image recognition on the presence or absence of vehicles assumed to have been involved in the near miss and information on their locations from the video from the in-vehicle camera 24, and based on the recognized involved vehicle information, selects accident video (video of the involved vehicle not including the background, etc.) with information similar to that of the involved vehicle in the near miss situation from the video DB 33. The video generation model 34 then outputs a video obtained by combining the selected background video, video of the parties involved, and video of the involved vehicle to the information output device 15, for viewing by, for example, the driver of the party involved in the near miss.

[0090] In the case of this second method of generating accident footage, the footage DB 33 will include a database for selecting background images (a database of information associating accident background image elements (intersection information, etc.) with background images for generating simulation images), a database for selecting images of parties involved in the accident (a database of information associating accident party elements (location information of each party, etc.) with images of parties involved for generating simulation images), and a database for selecting related vehicle images (a database of information associating accident-related vehicle elements (location information of each related vehicle, etc.) with related vehicle images for generating simulation images).

[0091] In the case of a video, the video generation model 34 may output the video generated by stitching a near-miss image with a simulation image of the accident (splicing the simulation image at the timing just before the avoidance action in the near-miss image) to the information output device 15.

[0092] As a result, the accident estimation device 1 can contribute to improving the effectiveness of traffic safety education by helping drivers understand the horror of accidents that can occur from near-miss situations that the drivers themselves have experienced.

[0093] Furthermore, the video generation model 34 may be configured to superimpose a real-life scenery image (a real-life photographed image or a real-life scenery image of the accident site included in the map data) at the time of the near-miss onto the accident object (an image of the person involved and an image of a related vehicle (without a background image)) in the accident video selected from the video DB 33, and output the video to the information output device 15. This allows the scenery and the like in the simulation video to be real-life images, which can enhance the impact on the viewer.

[0094] Third accident video generation method: The second accident video generation method can reduce the data volume compared to the first accident video generation method, but has the problem of a large data volume because the video of the parties involved in the accident and the accident-related vehicles is stored in the video DB 33. In the third accident video generation method, images of the accident subject objects (parties involved and related vehicles) are generated based on physical characteristics, etc., estimated by image recognition processing of the images captured at the time of the near-miss. The generated images of the accident vehicles and accident objects (people, objects, etc.) are then combined with the background video to generate an accident image.

[0095] Specifically, the position, movement speed, and weight of the parties involved in the near-miss and related vehicles are estimated based on images taken at the time of the near-miss. Note that the position and movement speed are estimated based on the detection of each object using image recognition and the detected position and position change. Furthermore, the weight is estimated by detecting the type and size of each object using image recognition and estimating the weight of each object based on the detected type and size. Note that the weight of each object can be estimated, for example, using a database in which weight is associated with the type and size of the object.

[0096] Then, based on the estimated position, movement speed, and weight of each object, the laws of physics (mainly the laws of mechanics) are used to estimate the collision situation and post-collision damage of each object, and computer graphics technology is used to generate images of accidents that are feared to occur due to near misses.

[0097] 6 and 7 may be stored in the video DB 33, and accident images may be generated using the data in these tables. The table shown in FIG. 6 associates accident scene IDs with various physical parameters of the accident. The various physical parameters of the accident include, for example, parameters indicating the vehicle involved, the relative speed of the vehicle at the time of the accident, the collision direction of the vehicle at the time of the accident, and the collision position of the vehicle at the time of the accident. The table shown in FIG. 7 associates vehicle types with vehicle images.

[0098] When the tables shown in Figures 6 and 7 are stored in the image DB 33, the image generation model 34 generates a simulation image using various physical parameters of an accident similar to the image captured by the in-vehicle camera 24 from near-miss information.

[0099] Specifically, the image generation model 34 estimates the movement and damage state of the target vehicle using various physical parameters in the accident, and generates an image of the target vehicle in the accident using CG.The image generation model 34 then generates a simulation image by superimposing this accident image on the scenery of the near miss (an actually photographed image or a scenery image included in the map data).

[0100] In this case, the image generation model 34 may be configured to read image data of the vehicle model of the target vehicle from the table shown in Fig. 7 and use the data for CG. The color of the vehicle may be included in the vehicle model, or may be colored into the vehicle image.

[0101] The image generation model 34 may also be configured by an image information generation AI model 34A (see FIG. 9) that generates the above-mentioned simulation image from an image of a near miss. The image information generation AI model 34A will be described later with reference to FIGS. 9 to 15.

[0102] Furthermore, the video generation model 34 (video information generation AI model 34A) generates and outputs a simulation video of the accident scene including at least one of information on the driver's own injury situation, images of personal injury, property damage situation, and images of property damage. This allows the accident estimation device 1 to more realistically understand the danger of an accident that can occur from a near-miss situation, thereby contributing to improving the effectiveness of traffic safety education.

[0103] However, depending on the content of the accident scene simulation video that drivers are shown, the effectiveness of traffic safety education may not be sufficiently improved. For example, if the content of the simulation video is too extreme, drivers with a low tolerance for extreme content may turn away from the simulation video and fail to understand the importance of the simulation video.

[0104] Therefore, the video generation model 34 processes the content of the simulation video to be viewed next by the same viewer according to the viewer's reaction to the simulation video of the accident scene. Alternatively, the content of the simulation video can be processed by providing a separate image processing unit that processes images according to the viewer's reaction (tolerance, described later) at the subsequent stage of the video information generation unit 31.

[0105] When a driver is viewing a video for the first time, the video information generation unit 31 causes the driver to view a simulation video with the degree of extremity set to a default value. In this case, the video information generation unit 31 generates and causes the driver to view a simulation video that is, for example, CG-only video rather than live-action video, and cuts out scenes of contact with people. The video information generation unit 31 also generates and causes the driver to view a simulation video in which the model and color of the vehicle 2 that appears are modified to be different from those of the vehicle that appeared in the near-miss situation. In other words, the video information generation unit 31 adjusts the degree of reproduction (similarity) of the generated image with respect to the actual image, thereby adjusting the intensity of the stimulation of the generated image (the extremity of the image).

[0106] Then, the video information generation unit 31 acquires an image of the viewer who has viewed the simulation video from the information output device 15, and determines the viewer's tolerance for extremeness using the tolerance determination unit 35. The video information generation unit 31 may acquire the viewer's heart rate from the information output device 15 and determine the tolerance based on the heart rate as well.

[0107] If the tolerance is below the threshold, for example, if it is recognized by image that the viewer is not paying attention to the simulation video, the video information generation unit 31 reduces the level of extremity of the next simulation video to be viewed. If the tolerance is above the threshold, for example, if it is recognized by image that the viewer is paying attention to the simulation video, the video information generation AI model 34A increases the level of extremity of the next simulation video to be viewed.

[0108] When decreasing the degree of extremism, the video information generating unit 31 performs image processing such as gradually decreasing the resolution of the video or performing blurring. When increasing the degree of extremism, the video information generating unit 31 performs image processing such as increasing the resolution of the video. Other methods that can be applied include adjusting the ratio of real images in the image (when decreasing the degree of extremism, the ratio of real images is decreased and the ratio of CG is increased), adjusting the ratio of video to still images (when decreasing the degree of extremity, the ratio of video is decreased and the ratio of still images is increased), etc.

[0109] Furthermore, when increasing the degree of extremism, the video information generation AI model 34A makes the simulation image (video or still image) more realistic by allowing the viewer to watch extreme scenes, such as scenes of contact with people, without cutting them out, or by making the model and color of the vehicle 2 that appears the same as the vehicle that appeared in the near-miss situation.

[0110] As a result, the accident estimation device 1 can contribute to improving traffic safety education by providing simulation images of accident scenes that can be reliably viewed by various drivers with different tolerances for extreme situations according to the driver's tolerance.

[0111] Next, a description will be given of the operation of the damage information generation unit 32. The damage information generation unit 32 generates and outputs damage information regarding damage caused by an accident when an actual accident occurs from a near-miss situation, based on the accident scene ID of the accident scene input from the accident estimation model 13 and the accident DB 12.

[0112] As a result, the accident estimation device 1 can help drivers involved in near-misses recognize the damage they will suffer if an actual accident occurs due to a near-miss situation, and encourage safe driving, thereby contributing to improving the effectiveness of traffic safety education.

[0113] As shown in Fig. 8, the accident DB 12 stores a table in which accident scene IDs are associated with damage (personal injury and property damage compensation, repairs, and insurance) information. The damage (personal injury and property damage compensation, repairs, and insurance) information includes information such as the liability ratio in the event of the accident, the amount of damage suffered by the driver, and changes in insurance grade and insurance premiums after insurance is applied to the accident. In other words, this damage (repair and insurance) information can be used to calculate insurance premiums, and can be considered insurance premium calculation information for calculating insurance premiums.

[0114] For example, when an accident scene ID is input from the accident estimation model 13, the damage information generation unit 32 searches and reads out the damage information associated with the accident scene ID from the accident DB 12, and outputs the read out damage information as accident information.

[0115] By providing such damage information to the driver, the accident estimation device 1 can help the driver recognize the damage situation after the accident, such as the increased self-payment required for vehicle repair costs and insurance, and encourage safe driving, thereby contributing to improving the effectiveness of traffic safety education.

[0116] Furthermore, the information stored in the accident DB 12 shown in FIG. 8 is generated based on cases of accidents that have actually occurred in the past. In other words, the damage information generation unit 32 generates accident information based on cases that have occurred in the past. This allows the damage information generation unit 32 to generate more realistic accident information for accidents that occur from near-miss situations and provide it to the driver. Note that while the table shown in FIG. 8 stores damage information in association with accident scene IDs, the damage information may also be stored in association with data that characterizes the accident scene, such as physical characteristic data of the parties involved and related vehicles.

[0117] The damage information generation unit 32 may also be configured by a damage information generation AI model 32A (see FIG. 9) that generates the above-mentioned damage information from an image of a near miss. Next, with reference to FIGS. 9 to 15, the accident information generation unit 14 that generates a simulation video of an accident occurring from a near miss situation and the above-mentioned damage information using AI will be described.

[0118] 9, the accident information generation unit 14 includes a video information generation unit 31 and a damage information generation unit 32. The video information generation unit 31 includes a video information generation AI model 34A. The damage information generation unit 32 includes a damage information generation AI model 32A.

[0119] The image information generation AI model 34A is an AI model that has been machine-learned to estimate an accident scene that occurs from the situation of a near-miss when near-miss information is input. For the machine learning of the image information generation AI model 34A, for example, a learning data set such as that shown in FIG. 10 is used.

[0120] As shown in Figure 10, the learning dataset of the video information generation AI model 34A is a dataset in which a plurality of near-miss images serving as input data are each associated one-to-one with a plurality of accident images serving as correct answer data.

[0121] 11 , the video information generation AI model 34A sequentially estimates simulation videos of likely accident scenes from the near-miss information of input data that is sequentially input, and outputs the estimated results. The video information generation unit 31 causes the video information generation AI model 34A to perform machine learning by updating the weighting coefficients of each layer in the DNN using processing such as backpropagation so as to reduce the error between the estimated results of the simulation videos output from the video information generation AI model 34A and the correct values ​​of the learning data corresponding to the input values.

[0122] This allows the video information generation unit 31 to generate a video information generation AI model 34A that can estimate, from the input near-miss information, a simulation video of an accident scene that is likely to occur in the near-miss situation corresponding to the near-miss information.

[0123] As shown in Figure 12, when an actually captured near-miss image is input, the machine-learned video information generation AI model 34A estimates and outputs a simulation video of an accident that occurs from the near-miss situation.

[0124] Furthermore, the video information generation AI model 34A may be configured to add image processing parameters corresponding to the viewer's reaction (the aforementioned tolerance) to the simulation image generation data, and perform image processing according to the parameters. In this case, for example, viewer reaction data may be added to the input data of the learning data, and the correct answer data may be image data corresponding to the viewer's reaction.

[0125] As a result, when near-miss information and viewer images are input, the video information generation AI model 34A can process the content of the simulation video to be viewed next by the same viewer depending on the viewer's reaction to the simulation video of the accident scene.In addition, it is also possible to realize this by providing a separate image processing unit at the subsequent stage of the video information generation AI model 34A that processes images depending on the viewer's reaction (the tolerance level mentioned above).

[0126] The damage information generation AI model 32A is an AI model that has been machine-learned to estimate damage information related to damages caused by an accident when near-miss information is input and an actual accident occurs based on the near-miss situation. For example, a learning dataset such as that shown in FIG. 13 is used for the machine learning of the damage information generation AI model 32A.

[0127] As shown in Figure 13, the learning dataset of the damage information generation AI model 32A is a dataset in which a plurality of near-miss images serving as input data are each in one-to-one correspondence with a plurality of pieces of damage information serving as correct answer data.

[0128] 14, the damage information generation AI model 32A sequentially estimates damage information related to damage caused by an accident when an actual accident occurs from near-miss information of input data that is sequentially input, and outputs the estimated result. The damage information generation unit 32 causes the damage information generation AI model 32A to perform machine learning by updating the weighting coefficients of each layer in the DNN using processing such as backpropagation so as to reduce the error between the estimated result of damage information output from the damage information generation AI model 32A and the correct value of the learning data corresponding to the input value.

[0129] This allows the damage information generation unit 32 to generate a damage information generation AI model 32A that can estimate damage information regarding damage caused by an accident if an actual accident occurs, from the input near-miss information.

[0130] As shown in Figure 15, when an actual near-miss image is input, the machine-learned damage information generation AI model 32A estimates and outputs damage information regarding the damage caused by the accident if an actual accident occurs based on the near-miss situation.

[0131] 16 and 17, a process executed by the controller 10 of the accident estimation device 1 will be described. Figures 16 and 17 are flowcharts showing an example of a process executed by the controller 10 of the accident estimation device 1 according to the embodiment. This process is repeatedly executed at appropriate intervals after the accident estimation device 1 is started up (the power switch is turned on).

[0132] 16, the controller 10 determines whether or not near-miss information has been received from the vehicle 2 (step S101). If the controller 10 determines that near-miss information has not been received (step S101, No), the controller 10 repeats the determination process of step S101 until near-miss information is received.

[0133] When the controller 10 determines that it has received near-miss information (Yes at step S101), it uses the accident estimation model 13 to estimate an accident scene that is likely to occur in the near-miss situation (step S102).

[0134] Next, the controller 10 executes an accident information generation process (step S103). Through this accident information generation process, the controller 10 generates damage information and a simulation video for the estimated accident scene.

[0135] Thereafter, the controller 10 determines whether it is time to output accident information (step S104). When the accident estimation device 1 is installed in the vehicle 2, the output timing of the accident information is, for example, the timing when the vehicle 2 stops or the timing when the vehicle 2 completes traveling one trip (for example, when the vehicle stops after arriving at the destination). In this case, the accident estimation device 1 may determine whether the vehicle 2 has stopped or completed traveling one trip based on vehicle speed data from the vehicle (determined by a vehicle speed of 0 for a longer period than a stop determination threshold), destination arrival signal data from the navigation device (determined by a match between the destination data and the current location), etc.

[0136] Furthermore, when the accident estimation device 1 is installed in a location other than the vehicle 2, the timing of outputting the accident information is the timing when a user of the accident estimation device 1 performs an operation to request output of the accident information. In this case, the accident estimation device 1 is configured to acquire, from an operation device that accepts user operations, information indicating that an operation to request output of the accident information has been performed.

[0137] If the controller 10 determines that it is not time to output the accident information (step S104, No), it repeats the determination process of step S104 until it is time to output the accident information. If the controller 10 determines that it is time to output the accident information (step S104, Yes), it outputs the accident information including the damage information and the accident simulation video (step S105).

[0138] Thereafter, the controller 10 determines whether or not the video of the viewer of the simulation video has been input (step S106). If the controller 10 determines that the video of the viewer has not been input (step S106, No), the controller 10 repeats the determination process of step S106 until the video of the viewer is input.

[0139] If the controller 10 determines that the viewer's video has been input (Yes at step S106), it determines the viewer's tolerance level for the extremeness of the simulation video from the input video (step S107).

[0140] When determining the viewer's tolerance using heart rate information, step S106 determines the input of the viewer's heart rate, and step S107 determines the tolerance based on the viewer's heart rate.

[0141] Then, the controller 10 changes the degree of extremism of the accident simulation video to be output next in accordance with the determined tolerance (step S108), and ends the process. Note that if the tolerance determined in step S107 is the same as the tolerance for the extremism of the previous simulation video, the controller 10 does not change the degree of extremism.

[0142] Next, the details of the accident information generation process in step S103 will be described with reference to Fig. 17. When the accident information generation process starts, the controller 10 generates damage information based on an accident scene estimated based on near-miss information (step S201).

[0143] The controller 10 generates information indicating the estimated percentage of fault between the driver and the other party in the accident scene, information indicating at least one of the personal injury status and property damage status, and information indicating future changes (progressions) in insurance premiums due to the occurrence of the accident. In addition, images such as images of personal injury status and property damage status, which can be generated based on the accident simulation image described below (by cutting out images of people or damaged parts of the vehicle), may be generated as damage information.

[0144] Furthermore, the controller 10 reads the viewer's tolerance level for the extremeness of the simulation video previously viewed, and generates a simulation image (video or still image) of an accident of an extreme level according to the tolerance level using the image generation model 34 (step S202), and then ends the process. In other words, this ends the accident information generation process of step S103. Thereafter, step S104 shown in FIG. 16 is executed.

[0145] 4. Application Examples of the Accident Estimation Device The accident estimation device 1 according to the embodiment can be applied to various devices. Application examples of the accident estimation device will be described below.

[0146] [4-1. Example of application to an insurance premium review device] The insurance premium review device is used to create insurance premium calculation standards and calculate insurance premiums for each individual. Currently, insurance premiums are calculated based on data related to accidents that have occurred. Therefore, insurance premiums are calculated based only on actual data, and do not include information on potential accidents. On the other hand, since near misses are potential accidents, when applied to an insurance premium review device, the accident estimation device 1 can estimate the damage from potential accidents based on the occurrence of near misses and the magnitude (damage) of the accidents that lead to them, and reflect this in insurance premiums.

[0147] In other words, for example, an insurance premium calculation standard (for example, data on near miss occurrences over an appropriate period of time is added to the parameters) will be created based on the total amount of damages caused by actual accidents and the amount of damages caused by accidents resulting from near misses (multiplied by an appropriate coefficient (big data is required (data collected over an appropriate period of time))), and the insurance premium for each individual will be calculated based on their data (data on near miss occurrences over an appropriate period of time will be used to calculate the insurance premium).

[0148] [4-2. Example of Application to an In-Vehicle Device] The accident estimation device 1 may be mounted on an in-vehicle device such as a car navigation device or a drive recorder. In this case, when a near-miss occurs, the in-vehicle device estimates an accident that is likely to occur in the near-miss situation based on the near-miss situation, and notifies the vehicle occupants of accident information related to the accident.

[0149] For example, the in-vehicle device notifies the occupant of accident information when the vehicle stops or when one trip ends. This allows the in-vehicle device to notify the occupant of accident information about an accident that is likely to occur as a result of a near miss, when the vehicle stops or when one trip ends, in the event of a near miss occurring, without much time having passed since the near miss. Therefore, by providing the occupant with accident information about an accident that is likely to occur as a result of a near miss that the occupant recently experienced, the in-vehicle device can encourage safe driving and contribute to improving the effectiveness of traffic safety education.

[0150] [4-3. Application Example to Terminal Device] The accident estimation device 1 may be mounted on a terminal device installed in an institution that provides traffic safety education, such as a driving school or a driver's license testing center. In this case, the terminal device estimates an accident that is likely to occur in a near-miss situation based on the near-miss situation, and outputs traffic safety education teaching material information corresponding to the accident.

[0151] This allows the information to be provided in the form of highly stimulating actual accidents, including not only actual accidents but also a wide range of accident-related information that can occur in various situations, such as accidents that follow on from near misses.This means that the information provided will have a strong impact on the participants, contributing to improving the effectiveness of traffic safety education.

[0152] For example, the terminal device estimates the type of accident that is likely to occur based on the near miss that the participant has actually experienced, and provides a simulation video of the process from the near miss to the actual accident as educational material information for traffic safety education.

[0153] This allows the terminal device to help drivers understand the horror of accidents that can occur from near-miss situations that the trainees themselves have experienced, thereby contributing to improving the effectiveness of traffic safety education.

[0154] 5. Information Processing System Next, an information processing system according to an embodiment will be described. Fig. 18 is an explanatory diagram of an information processing system 100 according to an embodiment. As shown in Fig. 18, the information processing system 100 includes an accident estimation device 1, an insurance premium consideration device 101, an on-board device 102, a terminal device 103, and a vehicle 2.

[0155] The accident estimation device 1, insurance premium consideration device 101, in-vehicle device 102, terminal device 103, and vehicle 2 are connected to each other so that information can be communicated via a communication network N. The insurance premium consideration device 101, in-vehicle device 102, and terminal device 103 are each equipped with an accident estimation device 1.

[0156] As a result, the information processing system 100 can provide the above-mentioned accident information at any of the locations where the accident estimation device 1 is installed, the location where the insurance premium consideration device 101 is installed, inside the vehicle where the on-board device 102 is installed, and the location where the terminal device 103 is installed. Therefore, the information processing system 100 can contribute to improving the effectiveness of traffic safety education by providing accident information to users in various locations.

[0157] [6. Supplementary Note] The features of the present invention are as follows: (1) An accident inference device comprising a controller that acquires near-miss information related to an occurred near-miss, and outputs accident type information for the occurred near-miss by referring, using the acquired near-miss information, to an accident type database that associates the near-miss information with accident types related to accidents that are feared to occur from the near-miss. (2) The accident inference device described in (1), wherein the controller outputs the virtual accident information for the occurred near-miss by referring, using the output accident type information, to an accident database that associates the accident type information with virtual accident information that indicates accident phenomena that will occur due to the accident. (3) The accident inference device described in (1), wherein the controller uses the near-miss information as input information instead of the accident type database, and outputs the accident type information for the occurred near-miss using an accident type model trained with learning data in which the accident type information is correct data. (4) The accident estimation device according to (2), wherein the controller uses the near-miss information as input information instead of the accident type database, and outputs the accident type information for the near-miss that has occurred using an accident type model trained with learning data in which the accident type information is correct data. (5) The accident estimation device according to (2) or (4), wherein the virtual accident information includes a simulation image of a virtual accident. (6) The accident estimation device according to (2) or (4), wherein the virtual accident information includes virtual damage information regarding damage caused by the virtual accident. (7) The accident estimation device according to (6), wherein the virtual damage information includes information indicating the fault ratios of the first party and the second party in the accident. (8) The accident estimation device according to (6) or (7), wherein the virtual damage information includes information indicating at least one of a personal injury situation, an image of the personal injury situation, an item damage situation, and an image of the item damage situation in the virtual accident. (9) The accident estimation device according to (6), (7), or (8), wherein the hypothetical damage information includes information indicating a change in insurance premiums due to the occurrence of a hypothetical accident.(10) An accident estimation device that acquires near miss information on a near miss that has occurred, and outputs the virtual accident information by inputting the acquired near miss information into a virtual accident model that has been trained with learning data that uses near miss information on the occurrence circumstances of the near miss as input information and virtual accident information on an accident that is feared to occur from the near miss as correct answer information. (11) An insurance premium review device that acquires near miss information on a near miss that has occurred, calculates virtual accident information on an accident that is feared to occur from the near miss based on the near miss information, calculates insurance premium calculation information on calculating an insurance premium based on the virtual accident information, and outputs the calculated insurance premium calculation information. (12) An on-board device that acquires near miss information on a near miss that has occurred, calculates virtual accident information on an accident that is feared to occur from the near miss based on the near miss information, and notifies a vehicle occupant of the virtual accident information. (13) A terminal device that acquires near miss information on an occurred near miss, calculates hypothetical accident information on an accident that is feared to occur from the near miss based on the near miss information, and outputs traffic safety education teaching material information including the hypothetical accident information. (14) The terminal device according to (13), wherein the hypothetical accident information is a simulation video showing the situation of an accident that is feared to occur from the near miss. (15) An information processing system having an accident estimation device and an on-board device mounted on a vehicle, wherein the accident estimation device acquires near-miss information on a near-miss that has occurred from the on-board device, calculates virtual accident information on an accident that is feared to occur from the near-miss based on the near-miss information, and transmits the calculated virtual accident information to the on-board device, and the on-board device collects near-miss information on the near-miss from an on-board sensor when the near-miss occurs, transmits the collected near-miss information to the accident estimation device, receives the virtual accident information transmitted from the accident estimation device, and notifies the received virtual accident information to an occupant of the vehicle.(16) An information processing system having an accident estimation device, an on-board device mounted on a vehicle, and an insurance premium review device, wherein the on-board device collects near-miss information related to a near-miss from an on-board sensor when the near-miss occurs and transmits the collected near-miss information to the accident estimation device, the accident estimation device acquires near-miss information related to the near-miss that has occurred from the on-board device, calculates virtual accident information related to an accident that is feared to occur from the near-miss based on the near-miss information, and transmits the calculated virtual accident information to the insurance premium review device, and the insurance premium review device receives the virtual accident information transmitted from the accident estimation device, calculates insurance premium calculation information related to the calculation of insurance premiums based on the received virtual accident information, and outputs the calculated insurance premium calculation information. (17) An accident inference method, in which a computer acquires near miss information regarding a near miss that has occurred, and outputs accident type information for the near miss that has occurred by referring to an accident type database that associates the near miss information with accident types related to accidents that are feared to occur from the near miss, using the acquired near miss information. (18) An accident inference program, in which a computer executes the steps of: acquiring near miss information regarding a near miss that has occurred, and outputting accident type information for the near miss that has occurred by referring to an accident type database that associates the near miss information with accident types related to accidents that are feared to occur from the near miss, using the acquired near miss information. (19) An AI model generation method, in which a computer uses near miss information regarding the occurrence status of a near miss as input information, and causes an AI (Artificial Intelligence) model to learn using learning information in which the accident types of accidents that are feared to occur from the near miss are correct information.(20) An AI generation program that causes a computer to execute a procedure for training an AI (Artificial Intelligence) model using learning information in which near miss information regarding the occurrence of near misses is used as input information and the accident type of an accident that is feared to occur as a result of the near miss is used as correct answer information.

[0158] Further advantages and modifications will readily occur to those skilled in the art. Therefore, the invention in its broader aspects is not limited to the specific details and representative embodiments shown and described above. Accordingly, various modifications may be made without departing from the spirit or scope of the general inventive concept as defined by the appended claims and their equivalents.

[0159] REFERENCE SIGNS LIST 1 Accident estimation device 2 Vehicle 10 Controller 11 Communication unit 12 Accident DB 13 Accident estimation model 13A Accident estimation AI model 14 Accident information generation unit 15 Information output device 20 Drive recorder 21 Acceleration sensor 22 Steering angle sensor 23 Emotion sensor 24 In-vehicle camera 25 In-vehicle microphone 26 Flash memory 27 Controller 28 Communication unit 29 Near-miss detection unit 31 Video information generation unit 32 Damage information generation unit 32A Damage information generation AI model 33 Video DB 34 Video generation model 34A Video information generation AI model 100 Information processing system 101 Insurance premium review device 102 In-vehicle device 103 Terminal device N Communication network

Claims

1. An accident estimation device having a controller that acquires near miss information regarding a near miss that has occurred, and outputs accident type information for the near miss that has occurred by referring to an accident type database that associates the near miss information with accident types related to accidents that are feared to occur from the near miss.

2. The accident estimation device according to claim 1, wherein the controller outputs the virtual accident information for the near miss that has occurred by referring to an accident database in which the output accident type information is associated with virtual accident information indicating accident events that may occur as a result of the accident.

3. The accident estimation device described in claim 1, wherein the controller uses the near-miss information as input information instead of the accident type database, and outputs the accident type information for the near-miss that has occurred using an accident type model trained with learning data in which the accident type information is correct data.

4. The accident estimation device described in claim 2, wherein the controller uses the near miss information as input information instead of the accident type database, and outputs the accident type information for the near miss that has occurred using an accident type model trained with learning data in which the accident type information is correct data.

5. The accident estimation device according to claim 2 or 4, wherein the virtual accident information includes a simulation image of the virtual accident.

6. The accident estimation device according to claim 2 or 4, wherein the virtual accident information includes virtual damage information relating to damage caused by the virtual accident.

7. The accident estimation device according to claim 6, wherein the hypothetical damage information includes information indicating the fault ratio between the party in question and the other party in the accident.

8. The accident estimation device according to claim 6, wherein the virtual damage information includes information indicating at least one of a personal injury situation, an image of the personal injury situation, an item damage situation, and an image of the item damage situation in the virtual accident.

9. The accident estimation device according to claim 6, wherein the hypothetical damage information includes information indicating a change in insurance premiums due to the occurrence of a hypothetical accident.

10. An accident estimation device that acquires near miss information regarding a near miss that has occurred, inputs the acquired near miss information into a virtual accident model that has been trained with learning data that uses near miss information regarding the circumstances under which the near miss occurred as input information and virtual accident information regarding accidents that are feared to occur from the near miss as correct answer information, and outputs the virtual accident information.

11. An insurance premium review device that acquires near miss information regarding near misses that have occurred, calculates hypothetical accident information regarding accidents that are feared to occur from the near misses based on the near miss information, calculates insurance premium calculation information regarding the calculation of insurance premiums based on the hypothetical accident information, and outputs the calculated insurance premium calculation information.

12. An in-vehicle device that acquires near-miss information regarding near-misses that have occurred, calculates hypothetical accident information regarding accidents that are feared to occur as a result of the near-misses based on the near-miss information, and notifies the vehicle occupants of the hypothetical accident information.

13. A terminal device that acquires near miss information regarding near misses that have occurred, calculates hypothetical accident information regarding accidents that are feared to occur as a result of the near misses based on the near miss information, and outputs information on teaching materials for traffic safety education that includes the hypothetical accident information.

14. The terminal device according to claim 13, wherein the virtual accident information is a simulation video showing a situation in which an accident is feared to occur due to a near miss.

15. An information processing system having an accident estimation device and an on-board device mounted on a vehicle, wherein the accident estimation device acquires near-miss information regarding a near-miss that has occurred from the on-board device, calculates hypothetical accident information regarding an accident that is feared to occur from the near-miss based on the near-miss information, and transmits the calculated hypothetical accident information to the on-board device, and the on-board device collects near-miss information regarding the near-miss from an on-board sensor when the near-miss occurs, transmits the collected near-miss information to the accident estimation device, receives the hypothetical accident information transmitted from the accident estimation device, and notifies the received hypothetical accident information to an occupant of the vehicle.

16. An information processing system having an accident estimation device, an on-board device mounted on a vehicle, and an insurance premium review device, wherein the on-board device collects near-miss information related to a near-miss from an on-board sensor when the near-miss occurs, and transmits the collected near-miss information to the accident estimation device, the accident estimation device acquires near-miss information related to an occurred near-miss from the on-board device, calculates hypothetical accident information related to an accident that is feared to occur from the near-miss based on the near-miss information, and transmits the calculated hypothetical accident information to the insurance premium review device, and the insurance premium review device receives the hypothetical accident information transmitted from the accident estimation device, calculates insurance premium calculation information related to the calculation of insurance premiums based on the received hypothetical accident information, and outputs the calculated insurance premium calculation information.

17. A method for estimating accidents, in which a computer acquires near miss information relating to a near miss that has occurred, and outputs accident type information for the near miss that has occurred by referring to an accident type database that associates the near miss information with accident types relating to accidents that are feared to occur from the near miss.

18. An accident estimation program that causes a computer to execute the steps of acquiring near miss information regarding a near miss that has occurred, and outputting accident type information for the near miss that has occurred by referring to an accident type database that associates the near miss information with accident types related to accidents that are feared to occur from the near miss.

19. A method for generating an AI model in which a computer uses near miss information regarding the occurrence of near misses as input information and trains an AI (Artificial Intelligence) model using learning information in which the correct answer information is the type of accident that is feared to occur as a result of the near miss.

20. An AI generation program that uses near miss information regarding the occurrence of near misses as input information and learning information in which the correct answer information is the type of accident that is feared to occur from the near miss, and causes a computer to execute a procedure to train an AI (Artificial Intelligence) model.

Images