Driver malfunction detection method

Abstract

This invention provides a driver abnormality detection method that can accurately identify abnormalities even in complex driving scenarios. [Solution] The system learns driving behavior data of healthy drivers using sparse coding (S1), selects a dictionary from the dictionary candidates that matches the driver's driving behavior schema in the target driving scene, constructs a driving behavior model (S2), calculates a baseline predicted value for features in driving behavior (S3), calculates the degree of deviation from the baseline predicted value distribution of the driving behavior data of healthy drivers and the driving behavior data of the target driver (S4-S6), and determines the abnormality of the target driver by comparing the degree of deviation with a predetermined threshold (S7-S9).

Description

Technical Field

[0001] The present invention relates to a driver abnormality determination method for determining the occurrence of abnormalities such as deterioration of the brain function of a driver during vehicle operation.

Background Art

[0002] In the operation of vehicles such as automobiles, various driving supports are provided according to the state of the driver. In order to appropriately execute such driving support, it is necessary to accurately estimate the state of the driver (for example, the occurrence of abnormalities such as deterioration of brain function) from the driving behavior of the driver. In this regard, various techniques for estimating the state of a target person based on the characteristics of the behavior of the target person have been proposed, not limited to vehicle operation.

[0003] For example, Patent Document 1 (Japanese Patent No. 6734986) proposes an invention that enables efficient analysis by applying sparse modeling in a behavior trend analysis apparatus or method for analyzing the tendency of behavior of a target person with respect to a service provided by a service provider (for example, a paid content distribution service). Further, Patent Document 2 (Japanese Unexamined Patent Application Publication No. 2021-167163) proposes an invention for determining the abnormality of a driver in a driver abnormality determination apparatus that determines the abnormality of a driver based on one or more detection target items and determination conditions adopted according to the driving scene.

Prior Art Documents

Patent Documents

[0004]

Patent Document 1

Patent Document 2

Summary of the Invention

Problems to be Solved by the Invention

[0005] Thus, techniques have existed to estimate a subject's state from their actions (for example, a vehicle driver's driving behavior). However, conventional techniques have not been able to accurately estimate a driver's state in complex driving scenarios (for example, driving at intersections or merging into other lanes). In other words, in complex driving scenarios, it is necessary to estimate the driver's state by evaluating the entire driving behavior, which combines multiple elements (for example, visual behavior and driving operations such as acceleration and steering), as a single organized behavior, but no technology exists that can handle such estimation. For example, in the technology of Patent Document 2, abnormality judgments are made for multiple detection target items, but judgments are made individually for each item, and the organized behavior combining each item is not evaluated.

[0006] This invention was made in consideration of the circumstances described above, and aims to provide a driver abnormality detection method that can accurately determine abnormalities in a vehicle driver by evaluating the organized driving behavior as a whole in complex driving scenarios. [Means for solving the problem]

[0007] To achieve the above objective, the present invention adopts the following solution. That is, as described in claim 1, a driver abnormality determination method for determining an abnormality of a vehicle driver comprises the steps of: generating dictionary candidates by learning the driving behavior data of a healthy driver in a target driving scene by sparse coding; selecting a dictionary from the dictionary candidates that is consistent with the driver's driving behavior schema in the target driving scene; constructing a driving behavior model that calculates predicted values ​​of the features of the driver's driving behavior based on the dictionary; calculating predicted values ​​of the features of the driving behavior of a healthy driver in the target driving scene using the driving behavior model and using these predicted values ​​as reference predicted values; calculating the degree of deviation of the driving behavior data of the healthy driver in the target driving scene from the reference predicted values; calculating the degree of deviation of the driving behavior data of the target driver in the target driving scene from the reference predicted values; and determining an abnormality of the target driver by comparing the degree of deviation of the healthy driver with the degree of deviation of the target driver.

[0008] According to the above solution method, the dictionary (e.g., dictionary 2) in the driving behavior model (e.g., driving behavior model 1) will be consistent with the driver's driving behavior schema in the target driving scene. Therefore, even if the driving behavior in the target driving scene is a complex combination of multiple driving behaviors, the complex (organized) driving behavior can be evaluated as a whole, and the baseline predicted values ​​of the features in the driver's driving behavior can be accurately calculated. Consequently, by comparing the degree of deviation of the target driver's driving behavior data from the baseline predicted values ​​with the degree of deviation of the driving behavior data of a healthy driver (a driver in a normal state), the abnormality of the target driver (e.g., a decline in brain function such as the onset of a brain disease) can be accurately determined.

[0009] A preferred embodiment based on the above solution method is as described in claim 2 and subsequent claims of the patent. That is, the dictionary may be constructed by combining bases corresponding to each of the driving behavior schemas (corresponding to claim 2). In this case, since each base of the dictionary corresponds to each of the driving behavior schemas, the dictionary will be appropriately aligned with the driving behavior schemas.

[0010] The driving behaviors represented by each base of the dictionary may be identified by classifying the features of the driving behaviors into the classes of driving in the center of the lane width, acceleration, braking, steering, gaze range, and gaze direction. If the driving behaviors represented by each identified base correspond to each of the driving behavior schemas, it may be determined that each base corresponds to each of the driving behavior schemas (corresponding to claim 3). In this case, the driving behaviors represented by each base of the dictionary can be accurately identified based on the content suggested by each class, so that a dictionary consistent with the driving behavior schema can be appropriately selected.

[0011] The aforementioned driving behavior schema may consist of a deceleration schema, a surroundings confirmation schema, an acceleration schema, a steering schema, and a speed maintenance schema (corresponding to claim 4). In this case, since the driving behavior schema is composed of appropriate elements, complex driving behaviors can be accurately evaluated.

[0012] The step of determining an abnormality in the target driver may include the steps of: calculating the distribution of deviations from the standard predicted values ​​of the driving behavior data of the healthy driver in the target driving scene; calculating the deviation amount, which is the difference between the deviation distribution of the healthy driver and the deviation distribution of the target driver, using KL divergence; and determining that the target driver is abnormal if the deviation amount is greater than or equal to a predetermined threshold (corresponding to claim 5). In this case, driver abnormality is determined based on the deviation amount, which is the difference between the deviation distributions of the healthy driver and the target driver, so that accurate determination can be made using appropriately quantified criteria.

[0013] The driving behavior data may consist of features that show a significant difference between a healthy driver and a driver in an abnormal state in the driving behavior of the target driving scene (corresponding to claim 6). In this case, the features in the driving behavior data can be made suitable for determining driver abnormalities.

[0014] The target driving scene is a driving scene at an intersection or when merging into another lane, and the driving behavior data may consist of features related to the driver's gaze behavior and driving operations (corresponding to claim 6). In this case, the driving behavior data consists of features related to gaze behavior and driving operations, which tend to show particularly large differences between a healthy driver and a driver in an abnormal state (for example, a driver with impaired brain function) when driving at an intersection or when merging, so that accurate driver abnormality can be determined. [Effects of the Invention]

[0015] According to the present invention, a driving behavior model having a dictionary consistent with the driving behavior schema in the target driving scene calculates a baseline predicted value for features in the driver's driving behavior. Based on the degree of deviation from this baseline predicted value (e.g., deviation distribution), an abnormality determination is made for the target driver. Therefore, even in driving scenes where complex driving behaviors are performed by systematically combining multiple driving behaviors, accurate abnormality determination (detection of brain function decline) based on human mechanisms can be performed. [Brief explanation of the drawing]

[0016] [Figure 1] A diagram illustrating the schematic of driver abnormality detection in an embodiment of the present invention. [Figure 2] A diagram illustrating the basics of sparse coding. [Figure 3] A graph showing the rationale for feature selection in relation to the left hemisphere of the brain. [Figure 4] A graph showing the rationale for feature selection in relation to the right hemisphere of the brain. [Figure 5]A diagram showing an overview of dictionary selection, and a diagram showing the case where the dictionary is appropriate. [Figure 6] A diagram showing an overview of dictionary selection, and a diagram showing the case where the dictionary is inappropriate. [Figure 7] A graph showing the effectiveness of driver abnormality determination of the present invention. [Figure 8] A flowchart showing an example of a driver abnormality determination procedure in the present invention.

Embodiments for Carrying Out the Invention

[0017] Hereinafter, embodiments of the present invention will be described based on the accompanying drawings. FIG. 1 shows an overview of the driver abnormality determination method in the embodiments of the present invention. As shown in the figure, in the driver abnormality determination method of the present embodiment, a driving behavior model 1 of the vehicle driver is used. The driving behavior model 1 is a computer model (program) that calculates a predicted value regarding the driving behavior of the driver in the target driving scene based on the driving behavior data in the target driving scene (the driving scene where abnormality determination is performed).

[0018] Here, the driving behavior data is data consisting of feature amounts (for example, amounts characterizing driving behavior such as the movement of the driver's line of sight, accelerator operation, steering wheel operation, etc.) accompanying the driving behavior of the driver in the target driving scene, and is acquired by various detection means (for example, cameras, accelerator sensors, steering angle sensors, etc.), and after being subjected to calculation processing as necessary, is stored in the storage means.

[0019] In the present embodiment, the feature amounts in the driving behavior of one or a plurality of healthy drivers (drivers in a normal state who are healthy persons without any sign of brain function decline) are acquired in time series and are used as the driving behavior data 4 of the healthy drivers. The driving behavior model 1 will calculate a predicted value (reference predicted value) of the feature amounts in the driving behavior of the healthy drivers based on the driving behavior data 4 of the healthy drivers.

[0020] As features of driving behavior, several features are selected that are considered to show particularly large differences (clear differences) between healthy individuals and individuals with impaired brain function in the target driving scene, in order to accurately determine (detect) the occurrence of driver abnormalities (occurrence of cognitive decline) in the target driving scene.

[0021] In this embodiment, when the driving scene is at an intersection or merging into another lane, elements of driving behavior related to visual behavior and driving operations are adopted as feature quantities. Specifically, the following are adopted as feature quantities: "average amount of offset from the center of the lane," "average amount of vehicle acceleration," "deviation of vehicle acceleration," "average amount of accelerator pedal depression," "average frequency of accelerator pedal use," "deviation of frequency of accelerator pedal use," "average amount of brake pedal depression," "deviation of brake pedal depression," "average amount of steering operation," "average range of gaze," "deviation of range of gaze," "average yaw of gaze," and "average yaw of gaze." Details of the selection of feature quantities in this embodiment will be described later with reference to Figures 3 and 4.

[0022] Driving behavior model 1 is generated by sparse coding using motor behavior data 4 from healthy drivers as training data. Here, sparse coding is a machine learning technique that reproduces signals (input data) as a combination of a small number of basis points (linear sum), and is a technique that can learn the relationships between data from a relatively small amount of data.

[0023] As shown in Figure 2, in sparse coding, the input data Y is decomposed into a lexicon matrix D and a sparse matrix X (or its product). Here, the lexicon matrix D is the main basis set in the training data (basis d1 to d6 in the example in Figure 2), and the sparse matrix X is a filter applied to the basis. In sparse coding, the lexicon matrix and sparse matrix are updated iteratively multiple times to minimize the difference between the input signal Y and the output signal DX, and finally a solution for dictionary D and sparse representation X is obtained.

[0024] This invention models complex driving behaviors (the coordination of visual behavior and driving operations) in complex driving scenarios (for example, driving at intersections and merging) into a driving behavior model 1 that conforms to human mechanisms (the driving behavior schema described later), and performs driver abnormality detection based on this driving behavior model 1. Therefore, it employs sparse coding, which utilizes the sparsity of the data to express the relationships between data using only essential information.

[0025] Driving behavior model 1 includes a dictionary (matrix) 2 and weights (vectors) 3 generated by sparse coding. In driving behavior model 1, dictionary 2 is selected from dictionary candidates obtained through learning by sparse coding, and those that have a basis consistent with the driver's driving behavior schema in the target driving scene are selected.

[0026] Here, "schema" is a concept used in cognitive psychology (schema theory) to explain human cognitive behavioral processes, and it refers to organized patterns in thought and behavior. It is believed that when humans take action, they recall related schemas from their brains and activate the necessary schemas at the appropriate time to achieve a series of actions.

[0027] In vehicle driving, a series of driving actions are thought to be realized through a combination of various driving behavior patterns (i.e., driving behavior schemas). For example, when driving at an intersection, a series of driving actions are performed, such as "decelerating before the intersection, checking left and right, and turning the steering wheel." In this invention, the idea of ​​schema theory is used in the generation of dictionaries in sparse coding, based on the observation that the execution of driving actions through a combination of multiple driving behavior schemas is similar to the representation of signals through a combination of multiple bases in sparse coding.

[0028] In this embodiment, driving behavior is assumed to consist of a combination of five driving behavior schemas: "deceleration schema (a pattern of driving behavior that decelerates the vehicle)," "surroundings confirmation schema (a pattern of driving behavior that checks the area around the vehicle)," "acceleration schema (a pattern of driving behavior that accelerates the vehicle)," "steering schema (a pattern of driving behavior that turns the direction of travel of the vehicle left or right)," and "speed maintenance schema (a pattern of driving behavior that drives while maintaining the speed of the vehicle)." Then, a dictionary in which each base of the dictionary appropriately corresponds to (is consistent with) each of these five driving behavior schemas is adopted as dictionary 2 in driving behavior model 1. Details of the optimization of the dictionary based on these five driving behavior schemas will be described later with reference to Figures 5 and 6.

[0029] Once dictionary 2 is optimized, we perform regression to determine what weights 3 should be applied to dictionary 2 to reproduce actual driving behavior (driving behavior data 4 of a healthy driver). This constructs driving behavior model 1. Using this driving behavior model 1, we calculate predicted values ​​for features in the driving behavior of a healthy driver and use these as reference predicted values ​​(values ​​of features that a healthy driver should normally possess).

[0030] Thus, while baseline predicted values ​​are calculated from the driving behavior model 1, even healthy individuals do not always behave exactly according to the baseline predicted values. Due to individual differences and variations in various conditions, there are some deviations from the ideal, normal driving behavior. Therefore, the degree of deviation (variation) from the baseline predicted values ​​of the driving behavior data 4 of healthy drivers is calculated and used as the criterion for determining abnormality in the target drivers (subjects subject to abnormality determination). For example, the distribution of deviations from the baseline predicted values ​​of healthy individuals' driving behavior data 5 is calculated.

[0031] In detecting abnormalities in a vehicle driver (the driver being assessed), the system detects characteristic features of the driver's driving behavior while the vehicle is in motion to obtain driving behavior data 6 of the driver being assessed. The system then calculates the degree of deviation of this driving behavior data 6 (measured values ​​of the detected features) from a baseline predicted value. By comparing this degree of deviation from the baseline predicted value of the driver being assessed with the degree of deviation from the baseline predicted value of a healthy driver, the system determines whether the driver being assessed is abnormal (impaired brain function).

[0032] For example, the difference between the distribution of deviations from the baseline predicted value 7 of the driving behavior data 6 of the target driver acquired in time series and the distribution of deviations from the baseline predicted value 5 of the driving behavior data 4 of a healthy driver is calculated using KL divergence (Kulback-Leibler divergence), and the deviation is compressed into one dimension using the Mahalanobis distance to calculate the degree of deviation. If this degree of deviation (the difference in the distribution of deviations from the baseline predicted value between the target driver and the healthy driver) is greater than or equal to a predetermined threshold, it is determined that an abnormality has occurred in the target driver (for example, that brain function has deteriorated due to a stroke, etc.).

[0033] Next, the details of the selection of feature quantities in the driver's driving behavior in this embodiment will be explained with reference to Figures 3 and 4. As mentioned above, the present invention aims to determine (detect) the occurrence of driver abnormalities (occurrence of brain function decline) in a target driving scene, so as feature quantities for driving behavior, feature quantities are adopted that clearly show the difference between a driver in a normal state (healthy person) and a driver in an abnormal state (person with brain function decline) in the target driving scene.

[0034] In this regard, observations of the driving behavior of stroke patients using driving simulators have shown that when the target driving scenario involves driving at intersections or merging, there are significant differences in visual behavior and driving operations (pedal operation and steering operation) between healthy individuals and stroke patients. For example, individuals with impaired brain function tend to drive at low speeds and frequently accelerate and decelerate.

[0035] Furthermore, from a medical perspective, it is known that dysfunction in the right and left hemispheres of the cerebrum has different effects on abnormal driving behavior. For example, dysfunction in the right hemisphere of the brain has a greater impact on attention function, making it difficult for individuals with this dysfunction to allocate their attention during activities. On the other hand, dysfunction in the left hemisphere of the brain has a greater impact on vehicle control, making individuals with this dysfunction more prone to problems with pedal and steering.

[0036] In light of these findings, in this embodiment, the following features related to visual behavior and driving operations were selected as candidate features: "average amount of offset from the center of the lane," "average amount of vehicle acceleration," "deviation of vehicle acceleration," "average amount of accelerator pedal depression," "average amount of accelerator pedal use," "deviation of accelerator pedal use," "average amount of brake pedal depression," "deviation of brake pedal depression," "average amount of steering operation," "average range of gaze," "deviation of range of gaze," "average yaw of gaze," and "average yaw of gaze."

[0037] For these feature candidates, driving behavior data from individuals with impaired brain function (both left hemisphere and right hemisphere) was acquired using a driving simulator and compared with driving behavior data from healthy individuals to score the validity of the feature candidates. Specifically, a baseline predicted value was calculated from driving behavior model 1, which was generated using the above feature candidates as features. Scores were assigned to the deviation of the driving behavior data of individuals with impaired brain function from the baseline predicted value at each measurement point, and the sum of these scores was used as the importance of each feature. The validity of the selected feature candidates was then verified based on the magnitude of their importance.

[0038] Figures 3 and 4 show the results of validating the candidate features. Figure 3 shows the results for individuals with impaired left hemisphere function, and Figure 4 shows the results for individuals with impaired right hemisphere function. As can be seen from the graphs of the results, all of the selected candidate features have high importance, and it is considered that these candidate features can clearly capture the difference between healthy individuals and individuals with impaired brain function. Therefore, it is reasonable to set the selected candidate features as features, and in this embodiment, the candidate features are set as features.

[0039] Next, the details of dictionary selection (optimization) in this embodiment will be explained with reference to Figures 5 and 6. As described above, in constructing the driving behavior model 1 in the present invention, dictionary candidates are obtained by learning the driving behavior data of healthy drivers using sparse coding. Subsequently, for each dictionary candidate, it is checked whether each base is consistent with the driving behavior schema in the target driving scene, and the dictionary whose base is consistent with the driving behavior schema is adopted as dictionary 2 of the driving behavior model 1.

[0040] To explain in more detail, in dictionary optimization, the features of driving behavior are first classified into classes consisting of similar features. Specifically, as shown in the middle section of Figures 5 and 6, the "average amount of offset from the center of the lane" is classified into the "driving in the center of the lane (center line)" class. In addition, the "average amount of vehicle acceleration," the "deviation of vehicle acceleration," the "average amount of accelerator pedal depression," the "average frequency of accelerator pedal use," and the "deviation of accelerator pedal use" are classified into the "accelerator" class. In addition, the "average amount of brake pedal depression" and the "deviation of brake pedal depression" are classified into the "brake" class. In addition, the "average amount of steering operation" is classified into the "steering" class. In addition, the "average amount of gaze range" and the "deviation of gaze range" are classified into the "gaze range" class. In addition, the "average amount of gaze yaw" and the "average amount of gaze yaw" are classified into the "gaze direction" class.

[0041] Next, based on the numerical values ​​of the parts corresponding to each class in each base, the content and magnitude of the driving behavior indicated by the numerical values ​​of the parts belonging to each class in each base are identified. Subsequently, it is checked whether the driving behavior in each base suggested by the identified content and magnitude of the driving behavior is consistent with the driving behavior schema. Candidate dictionaries that are inconsistent with the driving behavior schema are rejected, and dictionaries that are consistent with the driving behavior schema are adopted as dictionary 2 of driving behavior model 1.

[0042] The following provides a detailed explanation. In Figures 5 and 6, the upper section shows examples of dictionary candidates obtained through learning using sparse coding. In each dictionary, parts that are considered to represent characteristic driving behaviors (numerical values ​​with large absolute values) are enclosed in a thick border. The table in the lower section of the figures shows what kind of driving behavior the numerical values ​​of the parts corresponding to the features belonging to each class in bases 0 to 4 of the dictionary suggest.

[0043] For example, considering the dictionary in Figure 5, the portion corresponding to the offset, which is a feature belonging to the class of driving in the center of the lane width, has large absolute values ​​of 0.4 and 0.3 for base 0 and base 1. Therefore, base 0 and base 1 correspond to driving behaviors that adjust their position relative to the lane's center line. On the other hand, for driving behaviors corresponding to bases 2 to 4, the absolute value of the portion corresponding to the offset is small, and no position adjustment relative to the center line is performed.

[0044] Similarly, by examining the features of the accelerator class, we can determine which driving behaviors within the category of accelerating and decelerating a vehicle each base corresponds to. For example, with base 0, the values ​​corresponding to both accelerator pedal depression and accelerator usage frequency are both -0.4 (negative values), so the corresponding driving behavior is deceleration. With base 2, the values ​​corresponding to accelerator pedal depression and accelerator usage frequency are 0.2 and 0.3 (positive values), respectively, so the corresponding driving behavior is acceleration.

[0045] Furthermore, examining the features of the gaze direction class, base 0 corresponds to a value of 0.2 for the average gaze yaw and -0.3 for the deviation, indicating that it corresponds to a driving behavior where the driver directs their gaze to one side, left or right. On the other hand, base 1 corresponds to a value of 0 for the average gaze yaw and 0.6 for the deviation, indicating that it corresponds to a driving behavior where the driver moves their gaze equally to both sides.

[0046] Thus, once the content of driving behaviors has been identified for each class of each base, we examine whether the combinations of driving behaviors in each base correspond appropriately to the driving behavior schema. This shows that bases 0 to 4 of the dictionary in Figure 5 correspond appropriately one-to-one with the five driving behavior schemas and are consistent with the driving behavior schema.

[0047] For example, the driving behavior of base 1 is a driving behavior in which the driver checks left and right while adjusting their position, so it is consistent with the "surroundings check schema". Similarly, the driving behavior of base 3 is a driving behavior in which the driver turns the steering wheel (performs steering operations) while adjusting the vehicle's speed, so it is consistent with the "steering schema". Likewise, the driving behaviors of bases 0, 2, and 4 are consistent with the "deceleration schema", "acceleration schema", and "speed maintenance schema", respectively. Therefore, the dictionary in Figure 5 is adopted as dictionary 2 in the driving behavior model 1.

[0048] On the other hand, in the dictionary shown in Figure 6, the patterns of driving behavior for each base do not correspond appropriately to the five driving behavior schemas, and bases 0-4 cannot be combined into a driving behavior schema. Therefore, the dictionary in Figure 6 is rejected and will not be adopted as dictionary 2 for driving behavior model 1.

[0049] Next, we will explain the verification results regarding the effectiveness of the driver anomaly detection method according to the present invention. Figure 7 shows the distribution of anomaly degrees (estimated errors) calculated by the driver anomaly determination method of the present invention (distribution of deviations from the reference predicted value) for healthy individuals, mildly affected patients (individuals with reduced brain function due to mild stroke), and severely affected patients (individuals with reduced brain function due to severe stroke). The graph also shows the anomaly determination threshold T (90% of the chi-squared distribution).

[0050] According to this verification, the false detection rate for healthy individuals was 0%, the correct detection rate for mildly ill patients was 45.9%, and the correct detection rate for severely ill patients was 94.4%. Therefore, the driver abnormality detection method of the present invention is considered to be effective in identifying individuals with impaired brain function.

[0051] Next, an example of the overall flow of the driver abnormality detection method of the present invention will be described according to the flowchart in Figure 8. In the driver abnormality detection, in step S1, motor behavior data of healthy individuals is learned using sparse coding.

[0052] In step S2, a dictionary 2 is selected from the dictionary candidates learned by sparse coding, which has a basis that matches the driving behavior schema in the target driving scene. A weight 3 is then calculated using regression computation to construct a driving behavior model 1. In the subsequent step S3, a baseline predicted value (a predicted value of the features in the driving behavior of a healthy driver) is calculated based on the constructed motor behavior model 1.

[0053] In step S4, the distribution of deviations 5 from the baseline predicted values ​​of feature quantities in the driving behavior data 4 of healthy drivers is calculated. In the following step S5, the distribution 7 of deviations 7 from the baseline predicted values ​​of feature quantities in the driving behavior data 6 of the drivers to be evaluated (subjects) is calculated.

[0054] In step S6, the degree of deviance is calculated from the difference in the deviance distribution between healthy drivers and drivers under evaluation (calculated using KL divergence and compressed into one dimension using Mahalanobis distance).

[0055] In step S7, it is determined whether the degree of deviation is above a predetermined threshold. If it is not above the threshold, the process proceeds to step S8, where it is determined that the driver under evaluation is in a normal state (a healthy state in which no decline in brain function is observed), and the process ends. On the other hand, if the degree of deviation is above the threshold, the process proceeds to step S9, where it is determined that the driver under evaluation is in an abnormal state (a state in which brain function is impaired), and the process ends.

[0056] Although embodiments of the present invention have been described above, the present invention is not limited to the above embodiments, and appropriate modifications can be made within the scope described in the claims. For example, the driving behavior features, feature classification, and driving behavior schema shown in the above embodiments are examples, and the driving behavior features, feature classification, and driving behavior schema can be set differently from those in the above embodiments to suit the target driving scene, etc.

[0057] Furthermore, although the present invention was described in the above embodiments as a method for determining abnormalities in the driver of an actual vehicle such as an automobile, the present invention can also be applied to determining abnormalities in the driver of a device that simulates an actual vehicle (a simulated vehicle), such as a driving simulator. [Industrial applicability]

[0058] This invention can be used to determine the occurrence of abnormalities (decreased brain function) in a driver while operating a vehicle. [Explanation of Symbols]

[0059] 1. Driving Behavior Model 2. Dictionaries in driving behavior models 3. Weights in the driving behavior model 4. Driving behavior data of healthy drivers 5. Distribution of deviations in healthy drivers 6. Driving behavior data of the driver being evaluated. 7. Distribution of deviations of drivers subject to evaluation

Claims

1. In a driver abnormality detection method for determining abnormalities in the vehicle driver, The process involves generating dictionary candidates by learning driving behavior data of healthy drivers in a target driving scene using sparse coding, selecting a dictionary from the dictionary candidates that matches the driver's driving behavior schema in the target driving scene, and constructing a driving behavior model that calculates predicted values ​​for the driver's driving behavior features based on the dictionary. The steps include: calculating predicted values ​​for the characteristics of a healthy driver's driving behavior in the target driving scene using the aforementioned driving behavior model, and using these predicted values ​​as the baseline prediction; A step of calculating the degree of deviation of the driving behavior data of the healthy driver in the target driving scene from the standard predicted value, A step of calculating the degree of deviation of the driving behavior data of the target driver in the target driving scene from the standard predicted value, The steps include determining an abnormality in the target driver by comparing the degree of deviation of the healthy driver with the degree of deviation of the target driver, and A driver abnormality detection method that includes the following features.

2. In the driver abnormality determination method described in claim 1, The aforementioned dictionary is a driver abnormality detection method constructed by combining bases corresponding to each of the aforementioned driving behavior schemas.

3. In the driver abnormality determination method described in claim 2, A driver abnormality detection method that identifies the driving behavior represented by each base of the dictionary by classifying the feature quantities of the aforementioned driving behavior into the classes of driving in the center of the lane width, accelerator, brake, steering, gaze range, and gaze direction, and determines that each base corresponds to each of the aforementioned driving behavior schemas when the identified driving behavior represented by each base corresponds to each of the aforementioned driving behavior schemas.

4. In the driver abnormality determination method described in claim 2, The aforementioned driving behavior schema is a driver abnormality detection method comprising a deceleration schema, a surroundings confirmation schema, an acceleration schema, a steering schema, and a speed maintenance schema.

5. In the driver abnormality determination method described in claim 1, The step of determining an abnormality in the driver to be judged is: The steps include: calculating the distribution of deviations from the standard predicted values ​​of the driving behavior data of the healthy driver in the target driving scene; The steps include: calculating the distribution of deviations from the standard predicted value of the driving behavior data of the target driver in the target driving scene; The steps include: calculating the degree of deviation, which is the difference between the deviation distribution of the healthy driver and the deviation distribution of the driver to be judged, using KL divergence; The step of determining that the target driver is abnormal if the degree of deviation is greater than or equal to a predetermined threshold. A driver abnormality detection method that includes the following features.

6. In the driver abnormality determination method described in claim 1, The aforementioned driving behavior data consists of feature quantities that show a significant difference between a healthy driver and a driver in an abnormal state in the driving behavior of the target driving scene, and is used as a method for determining driver abnormality.

7. In the driver abnormality determination method described in claim 1, The aforementioned driving scenes are driving scenes at intersections or when merging into other lanes. The aforementioned driving behavior data is a driver abnormality detection method consisting of features related to the driver's gaze behavior and driving operations.

Images