Quantification method of risk factors of pedestrian-left-turning vehicle conflicts at intersections based on causal machine learning

Abstract

The present invention discloses a quantification method for pedestrian-left-turn vehicle conflict risk factors at intersections based on causal machine learning, specifically: 1: extract pedestrian-left-turn vehicle conflict data; 2: compare pedestrian-left-turn vehicle conflict data with signal control type T type and safety influencing factors; 3: Constructing a directed acyclic graph; 4: Based on the directed acyclic graph, a dual robust learning algorithm was used to construct a causal inference model for different signal control types. Among the safety influencing factors, factors related to pedestrian-left-turn vehicle conflicts were identified as conflict risk factors, and the heterogeneous causal effects of different signal control types on pedestrian-left-turn vehicle conflicts were estimated. This invention effectively identifies and addresses confounding factors, improving the model's robustness and generalization capabilities when used with real-world data.

Description

Technical Field

[0001] The present invention belongs to the field of traffic technology, and in particular relates to a method for quantifying risk factors of pedestrian-left-turning vehicle conflicts at intersections based on causal machine learning. Background Art

[0002] Intersections are key nodes where pedestrians and vehicles meet in urban traffic networks and are also areas where traffic accidents frequently occur. Conflicts between pedestrians and vehicles often stem from temporal overlap, poor spatial design, blind spots, and a lack of signal mechanisms specifically designed for pedestrians or turning vehicles. To mitigate these conflicts, urban planners have implemented strategies such as optimizing signal phases, establishing independent turn signals, improving infrastructure visibility, and strengthening traffic safety education to enhance overall intersection safety. At signalized intersections, collisions and near-collisions between pedestrians and left-turning vehicles are common, resulting in reduced traffic efficiency and health risks. Therefore, inferring the impact of left-turn control modes on pedestrian-vehicle safety and understanding the underlying patterns of other related safety factors are crucial. Therefore, a method for mining the risk factors for pedestrian-left-turning vehicle conflicts at intersections is urgently needed. Summary of the Invention

[0003] Purpose of the invention: In order to solve the problems existing in the above-mentioned prior art, the present invention provides a method for quantifying the risk factors of pedestrian-left-turning vehicle conflicts at intersections based on causal machine learning.

[0004] Technical Solution: The present invention discloses a method for quantifying the risk factors of pedestrian-left-turning vehicle conflicts at intersections based on causal machine learning, which is characterized by:

[0005] Step 1: Extract pedestrian-left-turn vehicle conflict data during the observation period;

[0006] Step 2: Combine pedestrian-left-turn vehicle conflict data with signal control type T type and matching of safety influencing factors;

[0007] Step 3: Set the signal control type to T type As the treatment variable, safety influencing factors as confounding variables, and pedestrian-left-turning vehicle conflict frequency Y as the output variable, a directed acyclic graph is constructed;

[0008] Step 4: Based on a directed acyclic graph, a dual robust learning algorithm is used to construct a causal inference model under different signal control types. Factors related to pedestrian-left-turn vehicle conflicts are extracted from safety influencing factors as conflict risk factors, and the heterogeneous causal effects of different signal control types on pedestrian-left-turn vehicle conflicts are estimated.

[0009] Furthermore, the step 1 is specifically as follows: setting a post-intrusion time threshold, and defining a conflict in the video stream in which the post-intrusion time is less than the post-intrusion time threshold as a pedestrian-left-turning vehicle conflict.

[0010] Furthermore, the signal control type T in step 2 type Includes protected left turn phase, permitted left turn phase, and protected-permitted left turn phase;

[0011] Safety influencing factors include traffic volume attribute X vol , road geometry attributes X road , signal timing attribute X signal And basic attribute X basic ; The traffic volume attribute X vol Including left-turn traffic volume Vol_Veh, pedestrian traffic volume Vol_Ped, and the opposite straight-moving vehicle traffic volume Vol_Opp of left-turn vehicles; road geometric attributes X road Includes: number of left-turn lanes LT, number of import lanes EntL, number of exit lanes ExitL, number of lanes for opposite traffic Thr_Opp, speed limit for left turn Slim and speed limit for opposite traffic Slim_Opp; signal timing attribute X signal Including the minimum green light duration GM, the average left turn green light duration AG, the maximum cycle MaxC of signal control, the minimum cycle MinC, the pedestrian walking time ExiW and the red light flashing time ExiFDW; basic attributes X basic Includes: weekdays, weekends, off-peak hours, and morning and evening peak hours.

[0012] Furthermore, in step 3, when constructing the directed acyclic graph, the variable T type Any indirect path to the output variable Y is regarded as a backdoor path, that is, link T type ←X c →Y is the backdoor path, X c A safety factor.

[0013] Furthermore, a dual robust learning algorithm is used to estimate the heterogeneous causal effects of different signal control types.

[0014]

[0015]

[0016] in, represents the expected value, X c is a safety factor, K(X c ) represents the similarity metric, θ t represents the treatment effect, θ t (Xc ) is:

[0017]

[0018] Among them, Y(T type =t) represents the frequency of pedestrian-left-turning vehicle conflicts under signal control type;

[0019] The expression is as follows:

[0020]

[0021] Among them, Y i is the frequency of pedestrian-left-turning vehicle conflicts actually observed, and the observation interval is obtained by dividing the observation time period. g t (X c ), For p t (X c ), t = 0, 1, 2; when t = 0, it indicates that the signal control type is the protection type left turn phase; when t = 1, it indicates that the signal control type is the permission type left turn phase; when t = 2, it indicates that the signal control type is the protection-permission left turn phase; g t (X c ) and p t (X c ) is as follows:

[0022] g t (X c )+ε t =Y(T type =t)

[0023] p t (X c )+δ=Pr(T type =t|X c )

[0024] Among them, ε t and δ are unobserved noises, Pr(T type =t|X c ) refers to the given X c Under the condition of , the probability of signal control type is predicted; based on the directed acyclic graph, the regression model is used to estimate g t (X c ), using the classification model to estimate p t (X c ).

[0025] Furthermore, the method also includes a preliminary screening of safety-related factors after step 2. Specifically, the linear relationship between safety-related factors is quantified using the Pearson correlation coefficient and multicollinearity is removed. Then, the BorutaSHAP method is used to eliminate safety-related factors with low contribution to conflict risk prediction.

[0026] Beneficial effects:

[0027] 1. This paper introduces causal graphs and a dual robust learning algorithm, which can effectively identify and handle confounding factors, thereby avoiding the problem of selection bias and improving the robustness and generalization ability of the model when facing real-world data.

[0028] 2. Unlike previous studies that only partially explain the impact of signal control types on conflict risk, this paper more accurately quantifies the heterogeneous causal effects of different signal control types on pedestrian-left-turn vehicle conflicts, providing a scientific basis for optimizing signal control strategies.

[0029] 3. The present invention takes into account physical factors such as traffic volume, road geometry, and signal timing attributes, and can comprehensively evaluate various factors affecting conflict risk, thereby enhancing the comprehensiveness and practicality of the factor mining model.

[0030] 4. By using the Pearson correlation coefficient and BorutaSHAP method to select important features, screen out key confounding variables, and construct a causal diagram, collinearity characteristics are effectively eliminated, thereby improving the accuracy of factor mining in the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 Flow chart of the method of the present invention.

[0032] Figure 2 This is a framework diagram of the method of the present invention.

[0033] Figure 3 Schematic diagram of Pearson correlation coefficient analysis.

[0034] Figure 4 is a directed acyclic graph.

[0035] Figure 5 This is the result diagram of heterogeneous causal effect estimation. DETAILED DESCRIPTION

[0036] The accompanying drawings, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0037] like Figure 1 , as shown in 2, the present invention specifically includes the following steps:

[0038] S1: The Post-Encroachment Time (PET) is used to define the pedestrian-left-turning vehicle conflict threshold, and the pedestrian-left-turning vehicle conflict data is extracted through video processing technology.

[0039] S2: Data aggregation. Match pedestrian-left-turn vehicle conflict data with signal control types and safety influencing factors.

[0040] S3: Confounding feature screening. Pearson correlation coefficient and BorutaSHAP method were used to eliminate collinearity and confounding factors with low contribution to model prediction from multiple safety influencing factors.

[0041] S4: Construct a causal relationship diagram to identify safety-influencing factors and characteristics for pedestrian-left-turning vehicle conflict modeling.

[0042] S5: A causal inference model under specific signal control types was constructed using the double robust learning (DRL) algorithm to explore the risk factors of pedestrian-left-turning vehicle conflicts and estimate the heterogeneous causal effects (HTE) of different signal control types on the risk of pedestrian-left-turning vehicle conflicts.

[0043] Specifically, S1 defines pedestrian-left-turning vehicle conflicts at intersections using the Post-Encroachment Time (PET). PET represents the time interval between the moment the first road user (vehicle or pedestrian) leaves the potential conflict zone (t1) and the moment the second road user (vehicle or pedestrian) enters the zone (t2). PET is calculated as shown in Equation (1):

[0044] PET=t2-t1 (1)

[0045] The threshold for recording PET is 10 seconds. Through video acquisition technology, conflicts with PET < 5 in the video stream are identified as pedestrian-left-turning vehicle conflicts. This is shown in Equation (2), where 1 indicates that the PET is recorded as a conflict, and 0 indicates that the PET is not recorded as a conflict.

[0046]

[0047] The total number of pedestrian-left-turning vehicle conflicts occurring during the period is counted every 60 minutes.

[0048] Specifically, S2 is performed by matching pedestrian-left-turn vehicle conflict data with signal control types and multidimensional features (i.e., confounding factors and features). Safety influencing factors in heterogeneous causal reasoning are determined and represented by high-dimensional confounding variables (X). The collected confounding variables (X) are aggregated at 60-minute intervals. The confounding variables considered in the present invention include:

[0049] Traffic volume attribute X vol Traffic volume attributes include information about the flow of different types of traffic participants within a specific time period. These attributes include left-turn traffic volume (Vol_Veh) (veh / h) and pedestrian traffic volume (Vol_Ped) (Ped / h). These two attributes are directly related to the occurrence of pedestrian-left-turning vehicle conflicts. The traffic volume of oncoming through vehicles (Vol_Opp) (veh / h) is also related to left-turning vehicles.

[0050] Road geometry attributes X road . Road geometry attributes describe the design layout of the intersection. Intersection geometry plays a key role in traffic signal design, especially when evaluating left-turn phases. In cases where left-turn traffic volume is high, especially in combination with protected control, it is usually necessary to set up a dedicated left-turn lane. In the present invention, the following variables are taken into account: the number of left-turn lanes LT, the number of entrance lanes EntL and the number of exit lanes ExitL, the number of lanes for opposite traffic Thr_Opp, the speed limit Slim for left turns and the speed limit Slim_Opp for opposite traffic.

[0051] Signal timing attribute X signal Signal timing attributes relate to the configuration of traffic lights. Specific signal timing attributes should be considered, including the maximum cycle (MaxC) and the minimum cycle (MinC). Signal control parameters for left-turning vehicles passing through an intersection include the minimum green light duration (GM) and the average left-turn green light duration (AG). Two typical pedestrian signal timing parameters include the pedestrian walking time (ExiW) and the red light flashing time (ExiFDW).

[0052] Basic AttributesX basic The basic attributes capture the temporal characteristics of the conflict. These variables include the day of the week (DAY) and the time of day (PERIOD) during which the conflict occurs. These variables help mitigate the impact of potential factors such as lighting and holidays. DAY is divided into weekdays (labeled 0) and weekends (labeled 1). PERIOD is divided into off-peak hours (labeled 0) and morning and evening peak hours (labeled 1).

[0053] Specifically, S3 uses the Pearson correlation coefficient to quantify the linear relationship between safety-related factors X and remove multicollinearity. This deterministic interdependence, known as collinearity, complicates data analysis and exacerbates bias as collinearity increases. The Pearson correlation coefficient heat map is shown in the figure below. Figure 3 Then, the BorutaSHAP method was used to eliminate confounding variables with low contribution to conflict risk prediction.

[0054] Specifically, S4 is as follows: The present invention estimates the heterogeneous treatment effects (HTE) of different signal control types based on the Rubin Causal Model (RCM) framework. The HTE is the effect of the signal control type T under the influence of the high-dimensional confounding variable X. type The influence of the frequency Y of pedestrian-left-turning vehicle conflicts is quantified. Since conflicts between left-turning vehicles and pedestrians are mainly related to the signal control type, it is assumed that there are potential safety-related factors such as road alignment and traffic condition attributes (step S2) that may affect the selection of signal control type, and the frequency of pedestrian-left-turning vehicle conflicts is affected by the signal control type. type =t) indicates signal control type T type =Y corresponding to t. The present invention considers three typical signal control types (ie, protection-type left turn phase, permission-type left turn phase and protection-permission left turn phase). The present invention defines T type ∈{0,1,2} is the left turn signal control type: T type =0 is the protection left turn phase; T type =1 is the permissive left turn phase, T type = 2 for the protection-permit left turn phase. The extracted traffic conflicts are matched with the signal control type at the time of the conflict, collected at 60-minute intervals. Then, the safety influencing factor X, filtered in step S3, is introduced. The estimated conditional average treatment effect (CATE) is calculated:

[0055]

[0056] The present invention uses a directed acyclic graph (DAG) to represent the potential causal relationship between high-dimensional observation data, such as Figure 4 As shown. type Defined as the treatment variable in the causal graph (i.e., directed acyclic graph), the safety impact factor after screening in step S3 is defined as X c , X c=X road ∪X signal ∪X vol ∪X basic , the output variable Y is the frequency of pedestrian-left-turning vehicle conflicts (times / hour). type ←X c →Y is the backdoor path, which is the link from the disposal variable T type Any indirect paths to the output variable Y that pass through the confounding variable X c Pass, it can reflect X c By identifying these backdoor paths and blocking them, the confounding variable X can be eliminated. c , thereby more accurately estimating causal effects.

[0057] The S5 is specifically to estimate the heterogeneous causal effect (HTE) of the signal control type. A double robust learning (DRL) algorithm is proposed to estimate the HTE. First, assume that:

[0058] Y(T type =t)=g t (X c )+ε t (4)

[0059] Pr(T type =t|X c )=p t (X c )+δ (5)

[0060] Where g t T type = the predicted value of the signal control class at time t, Pr is the predicted value of X c Under the influence of T type = the probability of t∈{0,1,2}, and ε t and δ is and The unobservable noise, represents the expected value, g t (X c ) is the function g t For X c The output of . It represents the output of c The expected result of processing type t under the conditions. t (X c ) is the function p t For X c The output of . It represents the output of c The predicted value of the probability that an individual will accept treatment t under the condition of type=t-related CATE θ t (X c ) is calculated as follows:

[0061]

[0062] Where g0(X c ) means that at a given X c Under the condition of type =0 is the predicted value of the signal control class.

[0063] Secondly, the DRL algorithm is used to estimate CATE θ t (X c ). This method models pedestrian-left-turning vehicle conflicts in a two-stage framework. In the first stage, two machine learning (ML) models are applied to estimate g based on the directed acyclic graph in step S4-2. t (X c ) and p t (X c ). Use regression model to estimate g t (X c ), and obtain the estimated result Using classification model to estimate p t (X c ), and obtain the estimated result Then, the inverse propensity scoring (IPS) method is used to eliminate bias. By combining the IPS method with the regression model, as long as the estimation results of any of the two models are accurate, the DRL model can obtain unbiased estimation results. Potential output results of the first stage

[0064]

[0065] Among them, Y i is the frequency of pedestrian-left-turning vehicle conflicts actually observed.

[0066] The second stage of the DRL model estimates the heterogeneous causal effect. At this time, the estimation result of the heterogeneous causal effect CATE is As shown in Formula 8:

[0067]

[0068] Among them, the similarity measurement index K(X c ) is achieved by using the target label and X c After obtaining the fitted DRL model, we can further estimate the heterogeneous causal effects of each signal control type.

[0069] like Figure 5 As shown in FIG, this embodiment provides the HTE of the protection-permission left turn phase signal control mode on Vol_Veh, Vol_Ped and Vol_Opp respectively. (1) Figure 5 (a) in Figure 1 uses a univariate RCS (Restricted Cubic Spline) model to assess the linearity of Vol_Veh. ANOVA results indicate that a 3k knot RCS provides a better fit (P < 0.001). Vol_Veh analysis reveals that pedestrian-left-turning vehicle conflicts increase significantly with increasing left-turn volume. This phenomenon is similar to that observed with permissive left-turn phases. Therefore, it can be concluded that the presence of a permissive green phase in the permissive left-turn phase signal control mode contributes to the protection of low Vol_Veh.

[0070] (2) Figure 5 Figure (b) shows the Vol_Ped pedestrian-left-turning vehicle conflict HTE. ANOVA results indicate that the 3-k knot RCS provides a better fit (P < 0.01). Vol_Ped analysis also shows that pedestrian-left-turning vehicle conflicts exhibit a clear inverse U-shaped trend as the number of pedestrians increases. Contrary to common sense, when Vol_Ped exceeds approximately 50 veh / h at the 95% confidence level, pedestrian-left-turning vehicle conflicts significantly decrease. This suggests that these pedestrians may be crossing the intersection in groups, unable to be obscured by the visual obstruction of left-turning drivers, potentially providing more time for pedestrians to cross safely.

[0071] (3) Figure 5 Panel (c) shows that the critical conflict increases with increasing oncoming traffic volume, but further decreases when Vol_Opp exceeds approximately 750 veh / h at a 95% confidence level. This indicates that the protected-permitted left-turn phase always operates in the protected phase, when the oncoming traffic volume is higher, which may further reduce the conflict between pedestrians and left-turning traffic.

[0072] It should also be noted that the various specific technical features described in the above specific embodiments can be combined in any appropriate manner without contradiction. To avoid unnecessary repetition, the present invention will not further describe various possible combinations.

Claims

1. A quantification method for pedestrian-left-turning vehicle conflict risk factors at intersections based on causal machine learning, characterized by: Specifically: Step 1: Extract pedestrian-left-turn vehicle conflict data during the observation period; Step 2: Combine pedestrian-left-turn vehicle conflict data with signal control type T type and matching of safety influencing factors; Step 3: Set the signal control type to T type As the treatment variable, safety influencing factors as confounding variables, and pedestrian-left-turning vehicle conflict frequency Y as the output variable, a directed acyclic graph is constructed; Step 4: Based on a directed acyclic graph, a dual robust learning algorithm is used to construct a causal inference model for different signal control types. Factors related to pedestrian-left-turn vehicle conflicts are identified as conflict risk factors among safety influencing factors. The heterogeneous causal effects of different signal control types on pedestrian-left-turn vehicle conflicts are also estimated. Signal control type T in step 2 type Includes protected left turn phase, permitted left turn phase, and protected-permitted left turn phase; Safety influencing factors include traffic volume attribute X vol , road geometry attributes X road , signal timing attribute X signal And basic attribute X basic The traffic volume attribute X vol Including left-turn traffic volume Vol_Veh, pedestrian traffic volume Vol_Ped, and the opposite straight-moving vehicle traffic volume Vol_Opp of left-turn vehicles; road geometric attributes X road Includes: number of left-turn lanes LT, number of import lanes EntL, number of exit lanes ExitL, number of lanes for opposite traffic Thr_Opp, speed limit for left turn Slim and speed limit for opposite traffic Slim_Opp; signal timing attribute X signal Including the minimum green light duration GM, the average left turn green light duration AG, the maximum cycle MaxC of signal control, the minimum cycle MinC, the pedestrian walking time ExiW and the red light flashing time ExiFDW; basic attributes X basic Including: weekdays, weekends, off-peak hours and morning and evening peak hours; In step 3, when constructing the directed acyclic graph, the variable T is processed. type Any indirect path to the output variable Y is regarded as a backdoor path, that is, link T type ←X c →Y is the backdoor path, X c Safety influencing factors; Estimating heterogeneous causal effects of different signal control types using a dual robust learning algorithm in, represents the expected value, X c is a safety factor, K(X c ) represents the similarity metric, θ t represents the treatment effect, θ t (X c ) is: Among them, Y(T type =t) represents the frequency of pedestrian-left-turning vehicle conflicts under signal control type; The expression is as follows: Among them, Y i is the frequency of pedestrian-left-turning vehicle conflicts actually observed, and the observation interval is obtained by dividing the observation time period. g t (X c ), For p t (X c ), t = 0, 1, 2; when t = 0, it indicates that the signal control type is the protection type left turn phase; when t = 1, it indicates that the signal control type is the permission type left turn phase; when t = 2, it indicates that the signal control type is the protection-permission left turn phase; g t (X c ) and p t (X c ) is as follows: g t (X c )+ε t =Y(T type =t) p t (X c )+δ=Pr(T type =t|X c , Among them, ε t and δ are unobserved noises, Pr(T type =t|X c ) refers to the given X c Under the condition of , the probability of signal control type is predicted; based on the directed acyclic graph, the regression model is used to estimate g t (X c ), using the classification model to estimate p t (X c ).

2. The method for quantifying the risk factors of pedestrian-left-turning vehicle conflicts at intersections based on causal machine learning according to claim 1 is characterized in that: The step 1 specifically includes: setting a post-intrusion time threshold, and defining a conflict in the video stream in which the post-intrusion time is less than the post-intrusion time threshold as a pedestrian-left-turning vehicle conflict.

3. The method for quantifying the risk factors of pedestrian-left-turning vehicle conflicts at intersections based on causal machine learning according to claim 1 is characterized in that: The method also includes a preliminary screening of safety-related factors after step 2, specifically using the Pearson correlation coefficient to quantify the linear relationship between safety-related factors and remove multicollinearity features. Then, the BorutaSHAP method is used to eliminate safety-related factors with low contribution to conflict risk prediction.

Images