A driver behavior analysis and prediction method based on causal inference

Abstract

The application provides a driver behavior analysis and prediction method based on causal inference, belongs to the field of traffic engineering human factor analysis, and mainly evaluates and understands the driver behavior by constructing four types of driving index evaluation systems, including the headway, the time to collision, the speed and brake amplitude, and the eye tracking information. Then, the causal inference model of double machine learning is adopted to deeply study the driving behavior of the driver in various scenes and the difference thereof, and accurate causal estimation of the driver behavior is provided, so that the driver behavior can be further predicted. The method comprehensively analyzes the relationship between the driving behavior and the driving environment, can be close to the actual situation, improves the driving safety and the driving efficiency, and has high practical value for the fields of intelligent driving and traffic safety research.

Description

Technical Field

[0001] This invention belongs to the field of human factors analysis in traffic engineering, and more specifically relates to a method for analyzing and predicting driver behavior based on causal inference. Background Technology

[0002] In recent years, with the acceleration of urbanization, traffic congestion and frequent traffic accidents have become serious urban problems. The development of autonomous driving (AV) offers an innovative solution. Its core advantage lies in its ability to effectively reduce traffic accidents, optimize traffic flow, and improve road utilization efficiency through advanced perception systems and intelligent algorithms. AVs can analyze traffic conditions in real time, automatically adjust driving routes and speeds, and avoid traffic congestion and potential collisions. Furthermore, autonomous driving technology can reduce driver fatigue and human error, further improving road safety. With continuous technological advancements and maturity, AVs are expected to become an important component of future urban transportation systems, greatly improving urban traffic conditions and promoting sustainable development. In the near future, AVs will gradually coexist with human vehicles (HVs) in complex traffic scenarios and road environments. Before the widespread application of autonomous driving technology, analyzing and assessing the interaction between AVs and HVs is of great practical significance. For example, by studying how drivers react to AVs merging into traffic, autonomous driving companies can improve the systems and algorithms of autonomous vehicles based on their driving behavior; traffic planning and management departments can also use this information to improve road designs that are more AV- and HV-friendly. However, existing methods for analyzing the impact of AVs and HVs on driver behavior either remain at a macro level or only conduct relatively simple driving index analyses, without in-depth modeling and mining of driver behavior. These methods are insufficient to effectively meet the requirements of complex and repeatable simulation experiments, especially when dealing with counterfactual data (i.e., data not present in actual experiments) across multiple scenarios. Therefore, a comprehensive framework for driver behavior analysis and modeling is needed to promote the appearance and algorithm design of AVs, thereby improving road traffic efficiency and the social acceptance of AVs, and laying a solid foundation for the harmonious coexistence of AVs and HVs in the future. Summary of the Invention

[0003] The technical problem to be solved by this invention is to propose a driver behavior analysis and prediction method based on causal inference, so as to achieve a comprehensive causal estimation of driver behavior in a mixed AV and HV environment.

[0004] To achieve the above objectives, the present invention employs the following technical solution: the method comprises:

[0005] A driving performance evaluation system was constructed, which includes four categories of indicators: headway, time to collision, speed and braking amplitude, and eye-tracking information.

[0006] This study applies causal inference theory to driving behavior analysis to investigate the differences in driving behavior of human drivers when encountering AV and HV.

[0007] For a specific driving behavior evaluation index, experimental data of drivers in AV and HV scenarios are collected separately. After data preprocessing, the data is input into a causal inference model based on dual machine learning, which can output a causal estimate of driving behavior.

[0008] In one approach, the headway is the time interval between two consecutive vehicles passing the same point; the time interval is calculated by dividing the spatial distance between the preceding and following vehicles by the speed of the following vehicle, as shown below:

[0009]

[0010] Where X L (t) represents the position of the vehicle ahead at time t, X F (t) represents the position of the following vehicle at time t, V F (t) represents the speed of the car behind at time t.

[0011] In one approach, the impending collision time is defined as "the remaining time between the two vehicles before a collision occurs, if one does happen," and is described as follows:

[0012]

[0013] Where X L (t) represents the position of the vehicle ahead at time t, X F (t) represents the position of the following vehicle at time t, L represents the length of the preceding vehicle, and V F (t) represents the speed of the following vehicle at time t, V L (t) represents the speed of the vehicle in front at time t;

[0014] The headway is mainly used to measure following distance, that is, to use the collision time to indicate the collision risk and safety level in following situations.

[0015] In one approach, the learning task of the causal inference model based on dual machine learning is decomposed into two stages:

[0016] (1) Decompose the conditional estimation of the residuals into two prediction sub-tasks: the result and the processing;

[0017] (2) Estimates of ATE and CATE are obtained through the final prediction model with the residuals of the results and processing.

[0018] In one approach, the specific steps of the causal inference model based on dual machine learning are as follows:

[0019] Y=θ(X)·T+g(X)+∈#(1)

[0020] T=f(X)+η#(2)

[0021] E[∈|X]=0; E[η|X]=0; E[η·∈|

[0022] In equation (1), Y is the outcome, representing the characteristics of the vehicle driven by the driver, including speed and distance from AV or HV. θ(X) is the treatment effect for each X, representing the causal effect of treatment T on outcome Y, where T is the treatment method, i.e., vehicle type, which is AV or HV in a T-junction scenario. g(X) is a function used to control for covariates and confounding factors. ∈ represents the random error, indicating the randomness in Y that the model cannot explain.

[0023] In equation (2), T is the treatment mentioned above, f(X) is a model representing how T depends on the covariate X, and η is a random error similar to ∈ in equation (1);

[0024] In formula (3), given X, the expected value of the error terms ∈ and η is zero, which means that X contains all relevant information affecting Y and T; X is the background data of the driver;

[0025] The goal of this process is to estimate the constant marginal CATE, i.e., θ(X); the idea for estimating θ(X) is as follows, and the equation is rewritten as:

[0026] YE[Y|X]=θ(X)·(TE[TIX])+∈#(4)

[0027] In equation (4), E[Y|X] and E[T|X] are two conditional expectation functions, both of which are nonparametric regression tasks;

[0028] Therefore, by estimating according to q(X)=E[Y|X] and f(X)=E[T|X], the residuals are calculated as follows:

[0029]

[0030] It is then related by the following equation:

[0031]

[0032] Subsequently, since (E[η·∈|X]=0), estimating θ(X) is the final regression problem, that is, regressing Y on X and T (although it is a linear model in T), i.e.:

[0033]

[0034] Thus, we obtain the value of CATE (i.e., θ(X)).

[0035] In one embodiment, the speed and braking amplitude are indicators used to measure the driver's driving behavior and represent their driving state; the driving state of the vehicle driven by the driver is clearly represented by its speed.

[0036] A driver's braking behavior reflects the actions a driver takes when approaching an object that is too close to them in the vehicle or surrounding area.

[0037] In one embodiment, the eye-tracking information includes: gaze direction, convergence / divergence, eye opening degree, pupil diameter, pupil position, position of the focal object, type and name of the focal object, distance of the focal object from the eyeball, and duration of gaze on the focal object. Beneficial effects of this invention:

[0038] By constructing a driving indicator evaluation system and applying causal inference theory to analyze driving behavior, we can conduct in-depth research on the differences in driving behavior between human drivers encountering automated driving (AV) and manual driving (HV). Furthermore, we can analyze specific driving behavior evaluation indicators to predict driver behavior, thereby enhancing traffic safety and improving driving efficiency. Employing a dual machine learning causal inference model, we can more accurately estimate the causal effects of driving behavior, with high model fitting accuracy and more reliable prediction results. In addition, this method considers multiple factors such as eye-tracking information, providing a more comprehensive analysis of driver behavior and enabling more accurate driver behavior prediction. Attached Figure Description

[0039] Figure 1 This is a flowchart of the method of the present invention;

[0040] Figure 2 This is a diagram of the driving indicator evaluation system of the present invention;

[0041] Figure 3 This is a flowchart of the causal effect estimation method based on dual machine learning of the present invention;

[0042] Figure 4 This is a flowchart illustrating the construction process of the heterogeneity analysis decision tree of this invention. Detailed Implementation

[0043] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Typical embodiments of the invention are shown in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete.

[0044] Unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0045] like Figure 1 As shown, a method for analyzing and predicting driver behavior based on causal inference includes:

[0046] S1. Construct a driving indicator evaluation system, including four types of indicators: vehicle headway, impending collision time, speed and braking amplitude, and eye tracking information.

[0047] like Figure 2 As shown, to measure a driver's driving behavior, a driving indicator evaluation system needs to be constructed. This system includes four types of indicators: Time Headway, Time To Collision (TTC), Speed ​​and Braking Amplitude, and Eye Tracking Information. The first two indicators apply to situations where the driver's vehicle interacts with other vehicles, while the latter two indicators focus on either the driver's vehicle or the driver themselves. Time Headway measures the following distance between the driver's vehicle and other vehicles; TTC indicates the collision risk and safety level in following situations; Speed ​​and Braking Amplitude directly represent the state of the driver's vehicle, i.e., the vehicle's speed and the magnitude of braking; Eye Tracking Information reflects the driver's eye movement distribution patterns, including gaze direction, convergence and divergence, eye opening, pupil diameter, pupil position, focal object position, type and name of the focal object, distance of the focal object from the eyeball, and duration of gaze on the focal object. Through these eye movement data indicators, deeper behavioral factors of the driver can be uncovered.

[0048] S101, Calculate Time Headway

[0049] Time headway is defined as the time interval between two consecutive vehicles passing the same point. It provides important guidance for road design and management, and it depends on traffic flow composition and driving behavior. The time interval is calculated as the spatial distance between the preceding and following vehicles divided by the speed of the following vehicle, as shown below:

[0050]

[0051] Where X L (t) represents the position of the vehicle ahead at time t, X F (t) represents the position of the following vehicle at time t, V F (t) represents the speed of the car behind at time t.

[0052] S102. Calculate the time to collision (TTC).

[0053] Time-to-Traffic (TTC) is a common indicator used to represent safety, defined as "the time remaining between two vehicles before a collision occurs if one is to happen." It reflects the level of safety and collision risk during driving. Mathematically, it is described as follows:

[0054]

[0055] Where X L (t) represents the position of the vehicle ahead at time t, X F (t) represents the position of the following vehicle at time t, L represents the length of the preceding vehicle, and V F (t) represents the speed of the following vehicle at time t, V L (t) represents the speed of the vehicle in front at time t.

[0056] In this invention, the headway is mainly used to measure the following distance, while the TTC is used to indicate the collision risk and safety level in following situations.

[0057] S103, Speed ​​and Braking Amplitude

[0058] To measure a driver's driving behavior, indicators that directly represent their driving state are needed. The driving state of the vehicle being driven is clearly indicated by its speed. Furthermore, a driver's braking behavior reflects the actions taken when approaching objects that are too close to the vehicle or surrounding area, thereby minimizing the possibility of a collision. This type of indicator focuses on the vehicle itself being driven.

[0059] S104, Eye Tracking Information

[0060] Compared to traditional driving behavior evaluation indicators, incorporating eye-tracking data related to driving habits is crucial, as it can uncover deeper behavioral factors. Eye-tracking information includes: gaze direction, convergence / divergence, eye opening, pupil diameter, pupil position, focal object position, type and name of the focal object, distance of the focal object from the eyeball, and duration of gaze on the focal object. This eye-tracking data is relatively rich, and in actual analysis, it is not necessary to use all of it; adjustments can be made flexibly according to the specific situation. A specific use case is given below.

[0061] To analyze the behavioral patterns of drivers' eye information rules, five indicators were used: the type and name of the focal object, the distance of the focal object from the eyeball, the longest fixation time among all focal objects, the number of types of focal objects, and the pupil position of the focal point. The first indicator directly reflects the type of focal object; the second indicator reflects the distance pattern of the object being gazed at; based on the first indicator, a fourth indicator can be derived, which, together with the third indicator, reflects the driver's level of concentration. Specifically, the larger the third indicator and the smaller the fourth indicator, the more focused the driver is while driving; the fifth indicator reflects the distribution pattern of the eyeball from a macroscopic perspective.

[0062] like Figure 3 As shown in Figure S2, the causal inference theory is applied to driving behavior analysis to study the differences in driving behavior of human drivers when encountering AV and HV.

[0063] For two variables, correlation does not necessarily imply causation; however, causation usually manifests as statistical correlation. Causal inference is such a statistical theory and method used to assess causal effects and discover causal relationships between variables using data. For decades, it has been an important research topic in many fields, including statistics, computer science, education, public policy, and economics. In this invention, based on the aforementioned driving indicator evaluation system, causal inference theory is applied to driving behavior analysis to comprehensively study the differences in driving behavior of human drivers when encountering AV and HV scenarios. Specifically, for a specific driving behavior evaluation indicator, experimental data from drivers in AV and HV scenarios are collected separately. After certain data preprocessing, the data is input into the causal inference model based on dual machine learning used in this invention, which can then output a causal estimate of driving behavior.

[0064] The basic concepts and symbols of causal inference are defined, including units, treatments, observations, and counterfactual results.

[0065] Unit. A unit is the object of atomic research being studied. A unit can be a collection of physical objects, companies, or individuals at a specific point in time, such as a classroom or a market. In this invention, the unit refers to the driver of the test vehicle in the experiment, i.e., the subject.

[0066] Treatment. A treatment refers to an action performed on a unit. In this invention, treatment is the vehicle category, i.e., AV or HV, which interacts with the vehicle driven by the driver. Therefore, treatment is a binary variable, defined as T=1 if the vehicle interacting with the test vehicle is an AV, and T=0 if it is an HV. If T=1, the collected experimental data or sample belongs to the treatment group; otherwise, it belongs to the control group.

[0067] Observed outcome. The observed outcome is the result of the actual treatment taken by the unit, also known as the actual outcome. Y(T=t) represents the actual outcome when treatment T exists.

[0068] A counterfactual outcome is the result of an entity taking an alternative course of action (i.e., the opposite of the actual action). The difference between an observed outcome and a counterfactual outcome is that an observed outcome actually occurs, while a counterfactual outcome does not actually exist.

[0069] Pre-treatment variables are variables that are unaffected by treatment. Also known as background variables, they can be distances between different vehicles, vehicle speeds before treatment, etc.

[0070] Confounders are variables that influence both the treatment and the outcome. The relationship between these variables is implicit, not linear. Confounding variables can interfere with the treatment's effect on the outcome and may lead to erroneous conclusions. For example, girls who eat breakfast every day tend to have lower weight. Eating breakfast seems to help with weight loss, but other factors, such as regular exercise and sufficient sleep, may play a significant role in weight loss. In this case, having a better lifestyle is a common cause of both breakfast and lighter weight, so we can also consider it a confounding factor in the causal relationship between breakfast and lighter weight.

[0071] The evaluation metric for causal effects, treatment effect, can be defined at the group, treatment group, control group, and individual levels.

[0072] At the group level, the treatment effect is termed the Average Treatment Effect (ATE), which is defined as follows:

[0073]

[0074] Where Y(T=1) represents the result of the treatment group at the population level, and Y(T=0) represents the result of the control group at the population level.

[0075] For the treatment group, the treatment effect is named the Average Treatment Effect on the Treated Group (ATT), and it is defined as follows:

[0076]

[0077] in, and These represent the true and counterfactual results for the treatment group, respectively.

[0078] For the control group, the treatment effect is termed the average treatment effect on the control group (ATC), which is defined as follows:

[0079]

[0080] in, and These represent the counterfactual results and the true results of the control group, respectively.

[0081] The basic assumptions of causal inference: There are three main assumptions defined in causal inference models:

[0082] Assumption 1: Stable Unit Treatment Value Assumption (SUTVA). Units do not influence each other. That is, a driver's driving behavior is not affected by other drivers.

[0083] Assumption 2: (Statistical) Ignorability. Given a background variable X, the treatment assignment T is independent of the potential outcomes, mathematically expressed as:

[0084] W||Y(=0), Y(T=1)|X

[0085] Assumption 3: Positivity. For any value of X, the processing allocation is not deterministic, mathematically expressed as:

[0086] P(T=t|X=x)>0,

[0087] That is, for drivers with any characteristics, testing can be conducted in both AV and HV scenarios. In the method mentioned in this invention, all three assumptions above are satisfied.

[0088] Causal effect estimation based on dual machine learning methods addresses the challenge of assessing the potential heterogeneous impacts of AV or HV on driving behavior using traditional statistical methods such as least squares regression. Dual machine learning fundamentally solves this problem. It is a bias-reduced orthogonal method used to estimate the average treatment effect (ATE) between treatment and outcome, and can further estimate the conditional average treatment effect (CATE) based on the high-dimensional observable features of the treatment samples. More importantly, it can further explain the heterogeneous impacts based on decision trees. The following is a description of the method's principles.

[0089] The dual machine learning task is broken down into two stages:

[0090] (1) The conditional estimation of the residuals is decomposed into two prediction subtasks: the result and the processing.

[0091] (2) Estimates of ATE and CATE are obtained through the final prediction model with the residuals of the results and processing.

[0092] The steps are as follows:

[0093] Y=θ(X)·T+g(X)+∈#(1)

[0094] T=f(X)+η#(2)

[0095] E[∈|X]=0; E[η|X]=0; E[η·∈|

[0096] In equation (1), Y is the outcome, which can be a characteristic of the vehicle driven by the driver, such as speed, distance from AV or HV, etc. θ(X) is the treatment effect for each X, representing the causal effect of treatment T on outcome Y. T is the treatment method, i.e., vehicle type, which is AV or HV in a T-junction scenario. g(X) is a function used to control for covariate and confounding factor effects. ∈ is the random error, representing the randomness in Y that the model cannot explain.

[0097] In equation (2), T is the treatment mentioned above. f(X) is a model representing how T depends on the covariate X, and η is a random error similar to ∈ in equation (1).

[0098] In Equation (3), it is assumed that, given X, the expected values ​​of the error terms ∈ and η are zero, which means that X contains all relevant information affecting Y and T. In other words, the model has considered all confounding variables. In this study, X is the driver's background information, such as age and other data.

[0099] The goal is to estimate the constant marginal CATE, i.e., θ(X). The idea for estimating θ(X) is as follows, and the equation is rewritten as:

[0100] YE[Y|X]=θ(X)·(TE[T|X])+∈#(4)

[0101] In equation (4), E[Y|X] and E[T|X] are two conditional expectation functions (both are nonparametric regression tasks).

[0102] Therefore, if we can estimate them according to q(X)=E[Y|X] and f(X)=E[T|X], then we can calculate the residuals:

[0103]

[0104] It is then related by the following equation:

[0105]

[0106] Subsequently, since (E[η·∈|X]=0), estimating θ(X) is the final regression problem, that is, regressing Y on X and T (although it is a linear model in T), i.e.:

[0107]

[0108] At this point, the value of CATE (i.e., θ(X)) has been obtained. In addition, the causal effect estimation method based on dual machine learning can construct a causal tree for further heterogeneity analysis. The construction process of the heterogeneity analysis decision tree is attached. Figure 4 As shown.

[0109] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0110] It should be understood that the above detailed description of the technical solutions of the present invention with reference to preferred embodiments is illustrative and not restrictive. Those skilled in the art can modify the technical solutions described in the embodiments or make equivalent substitutions for some of the technical features based on reading this specification; however, these modifications or substitutions do not cause the essence of the corresponding technical solutions to depart from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A driver behavior analysis and prediction method based on causal inference, characterized in that: The method comprises: A driving index evaluation system is constructed, including four types of indexes, namely, headway, impending collision time, speed and brake amplitude, and eye movement tracking information; The causal inference theory is applied to driving behavior analysis to study the driving behavior differences of human drivers when encountering automatic driving (AV) and manual driving (HV); For a specific driving behavior evaluation index, experimental data of drivers in automatic driving (AV) and manual driving (HV) scenarios are collected respectively, and after data preprocessing, the data are input into a causal inference model based on double machine learning, and the causal estimation of driving behavior can be output; The learning task of the causal inference model based on double machine learning is divided into two stages: (1) The conditional estimation of the residual is divided into two prediction subtasks of results and processing; (2) The average treatment effect (ATE) and the conditional average treatment effect (CATE) are obtained through a final prediction model with the residuals of results and processing; The specific steps of the causal inference model based on double machine learning are as follows: Y=θ(X)·T+g(X)+∈ (1) T=f(X)+η (2) E[∈|X]=0;E[η|X]=0;E[η·∈|X]=0 (3) In equation (1), Y is the result of the characteristics of the vehicle driven by the driver, including speed, distance from automatic driving (AV) or manual driving (HV), θ(X) is the treatment effect of each X, representing the causal effect of treatment T on the result Y, T is the way of processing and analyzing data in different driving environments, that is, each type of vehicle automatic driving (AV) or manual driving (HV) is regarded as a unique processing method, g(X) is a function for controlling the effect of covariates and confounding factors; ∈ is a random error, representing the randomness in Y that cannot be explained by the model; In equation (2), T is the above-mentioned processing method, f(X) is a model representing how T depends on the covariate X, and η is a random error similar to ∈ in equation (1); In formula (3), the expectation values of error terms ∈ and η are zero given X, indicating that X contains all relevant information that affects Y and T; X is the background data of the driver; The purpose of processing is to estimate the constant conditional average treatment effect (CATE), that is, θ(X); the idea of estimating θ(X) is as follows, the equation is rewritten as: Y-E[Y|X]=θ(X)·(T-E[T|X])+∈ (4) In equation (4), E[Y|X] and E[T|X] are two conditional expectation functions, both of which are non-parametric regression tasks; Therefore, according to q(X)=E[Y|X] and f(X)=E[T|X], the residual is calculated: which is related by the following equation: Subsequently, since E[η·∈|X]=0, the estimation of θ(X) is a final regression problem, that is, Y is regressed on X and T, that is: At this point, the value of the conditional average treatment effect (CATE) θ(X) is obtained.

2. The driver behavior analysis and prediction method based on causal inference according to claim 1, characterized in that: The headway time is the time interval between two consecutive vehicles passing the same point, which is calculated by dividing the space interval between the front and rear vehicles by the speed of the rear vehicle, as shown below: where X L (t) represents the position of the preceding vehicle at time t, X F (t) is the position of the following vehicle at time t, V F (t) represents the speed of the following vehicle at time t.

3. The driver behavior analysis and prediction method based on causal inference according to claim 1, characterized in that: The time to collision is defined as the remaining time until a collision between two vehicles occurs if a collision happens, and the formula of the time to collision is expressed as follows: where X L (t) represents the position of the preceding vehicle at time t, X F (t) is the position of the following vehicle at time t, L represents the length of the preceding vehicle, V F (t) represents the speed of the following vehicle at time t, V L (t) represents the speed of the preceding vehicle at time t; The headway time is mainly used to measure the following distance, and the time to collision is used to indicate the collision risk and safety level in the following situation.

4. The driver behavior analysis and prediction method based on causal inference according to claim 1, characterized in that: The speed and brake amplitude are indicators for measuring the driving behavior of the driver and indicating the driving state of the driver; the driving state of the vehicle driven by the driver is clearly indicated by the speed. The brake behavior of the driver reflects the measures taken by the driver when approaching a vehicle or an object in the surrounding area that is too close to them.

5. The driver behavior analysis and prediction method based on causal inference according to claim 1, characterized in that: The eye tracking information includes the gaze direction, vergence, eye opening degree, pupil diameter, pupil position, focal object position, type and name of the focal object, distance between the focal object and the eyeball, and gaze duration of the focal object.

Images