CN116631186B - Highway traffic accident risk assessment method and system based on dangerous driving event data - Google Patents

Abstract

The invention discloses a highway traffic accident risk assessment method and system based on dangerous driving event data, wherein the method comprises the following steps: acquiring expressway traffic accident data and four types of dangerous driving event data; searching and counting the frequency of dangerous driving events on road sections before accidents and when no accidents occur and the average value of the maximum acceleration and the maximum speed of vehicles in all the same type of dangerous driving events respectively; constructing a sample data set by taking the dangerous driving event information as an independent variable and taking whether an accident is a dependent variable; processing the data into an image form, constructing and training a deep learning model, and taking into consideration the internal relation of dangerous driving event data in different event types and upstream and downstream space positions, so as to mine the rule of dangerous driving events on a road section before an accident occurs; and finally, acquiring dangerous driving event data of the whole highway section in real time, and evaluating the traffic accident risk in real time according to the trained model.

Description

Highway traffic accident risk assessment method and system based on dangerous driving event data

Technical Field

The invention provides a highway traffic accident risk assessment method and system based on dangerous driving event data, belonging to the technical field of road traffic safety.

Background technique

The traffic safety of highways has always been the focus of government attention. In recent years, the real-time assessment of highway traffic accident risks has become a research hotspot, which is of great significance for carrying out active traffic safety management and reducing the incidence of traffic accidents.

Existing highway traffic accident risk assessment methods usually use traffic flow data collected by fixed detectors (such as ETC gantries, ground ring coils, etc.) to determine accident risks. However, this type of data is limited by the location and density of fixed detectors, and it is difficult to dynamically reflect the traffic operation status at all locations on the entire road section, which in turn affects the accuracy of accident risk assessment. With the advancement of data collection technology, it is becoming easier to obtain data on dangerous driving behaviors of vehicles, and a series of dangerous driving behavior monitoring, risk assessment and early warning methods have emerged. However, these methods all use a single individual vehicle as the research object for risk assessment, and lack methods for traffic accident risk assessment from the perspective of road traffic managers, with the accident risk of the road section as the research object, and based on the data of dangerous driving events of vehicles on the road section.

In terms of accident risk assessment models, most existing methods use statistical analysis or machine learning models to mine the characteristic laws of factors such as road traffic flow before an accident occurs. Some methods also use recurrent neural network (RNN) deep learning models to consider the time correlation of traffic flow data for accident risk modeling. However, these methods all use one-dimensional vector data to represent factors such as road traffic flow before an accident occurs, and it is difficult to simultaneously consider the inherent connections of data in multiple dimensions such as time, space, and influencing factors. The convolutional neural network (CNN) deep learning model can efficiently extract local patterns in image data, comprehensively consider the correlation of data in multiple dimensions for model training and prediction, and the computational cost is much lower than that of RNN.

Since the dangerous driving incident data is collected through the vehicle navigation software, its collection range can cover all locations of the entire highway section and is not affected by weather conditions. Therefore, using the dangerous driving incident data of vehicles on the road and processing it into image form, and inputting it into the CNN deep learning model, it can comprehensively consider the multi-dimensional correlation contained in the data and realize real-time and refined assessment of traffic accident risks on various sections of the highway without installing new equipment.

Summary of the invention

Technical problem to be solved by the present invention: In view of the shortcomings of the prior art, the present invention uses the dangerous driving event data of vehicles on the highway, processes it into image form according to different dangerous driving event types and spatial position relationships, and uses a deep learning model including modules such as convolutional neural network (CNN) to mine the laws of dangerous driving events on the road section before the traffic accident occurs. A highway traffic accident risk assessment method and system based on dangerous driving event data are proposed, aiming to solve the problem that the existing highway traffic accident risk assessment method is limited by the location and density of fixed detectors, making it difficult to finely assess the accident risks at various locations on the highway, and difficult to extract and learn the intrinsic connections between data in different influencing factors and spatial positions, resulting in limited risk assessment accuracy.

The present invention adopts the following technical solutions to solve the above technical problems:

The present invention proposes a highway traffic accident risk assessment method based on dangerous driving event data, comprising the following steps:

S1: Determine the research section and statistical time range, and obtain the following two types of information:

(1) Data on dangerous driving incidents of vehicles on the research section of the expressway within the statistical time range, including the time of occurrence of the dangerous driving incident, the longitude, latitude and up and down directions of the location of the incident, the type of dangerous driving incident (sudden acceleration, sudden deceleration, sudden left lane change and sudden right lane change), and the maximum acceleration and maximum speed of the vehicle during the dangerous driving incident.

(2) Traffic accident data occurring on the research section of the expressway within the statistical time range, including the start and end time of the accident, the longitude, latitude, and up and down directions of the accident location.

S2. For each traffic accident, find the dangerous driving event data of the upstream and downstream sections of the accident site before the accident, take the dangerous driving event data as the independent variable and the accident as the dependent variable, and construct the accident case data set;

S3, searching for dangerous driving event data of upstream and downstream sections when no traffic accidents occurred, taking the dangerous driving event data as the independent variable and the absence of accidents as the dependent variable, constructing a non-accident case data set as a control group, merging the accident case data set and the non-accident case data set to obtain a sample data set;

S4. Standardize the data: process the dangerous driving event data into images according to different dangerous driving event types and spatial position relationships, build a deep learning model for training and parameter adjustment until the model effect is optimal;

S5: Real-time acquisition of dangerous driving incident data of vehicles on all sections of the expressway. Based on S2, the frequency of occurrence of various dangerous driving incidents and the average values of the maximum acceleration and maximum speed of vehicles in dangerous driving incidents are calculated, and the data are input into the deep learning model constructed by S4 to calculate the traffic accident risk assessment value and risk level of each section.

Specifically, step S2 is:

S2.1: According to the longitude, latitude, and up and down directions of the dangerous driving incidents and traffic accidents, mark them on the corresponding sections of the highway and establish the upstream and downstream relationships between the sections.

S2.2: Based on the spatiotemporal location of each traffic accident, extract data on four types of dangerous driving events: sudden acceleration, sudden deceleration, sudden left lane change, and sudden right lane change within a spatial range l upstream and downstream of the accident location within a time range t before the accident.

S2.3: For each traffic accident, calculate the following:

(1) Calculate the frequency of occurrence of the four types of dangerous driving events within the time and space range in S2.2, recorded as:

Times = [L_up, L_down, R_up, R_down, A_up, A_down, B_up, B_down]

The first part of each variable represents the type of dangerous driving event, i.e., L, R, A, and B represent the frequency of sudden left lane change, sudden right lane change, sudden acceleration, and sudden deceleration in the time range t before the accident. The second part of each variable represents the location where the dangerous driving event occurred, i.e., up and down represent the road sections within the spatial range l upstream and downstream of the accident location, respectively.

(2) Calculate the average value of the maximum acceleration and maximum speed of the vehicle in the same type of dangerous driving incident within the time and space range in S2.2, and record it as:

Acceleration = [La_up, La_down, Ra_up, Ra_down, Aa_up, Aa_down, Ba_up, Ba_down]

Speed = [Ls_up, Ls_down, Rs_up, Rs_down, As_up, As_down, Bs_up, Bs_down]

Among them, La and Ls in the first part of each variable represent the average values of the maximum acceleration and maximum speed of the vehicle in all sharp left lane change events, and the meanings of the remaining Ra, Rs, Aa, As, Ba, and Bs are similar. The second part of each variable represents the location where the dangerous driving event occurred.

S2.4: Combine the Times, Acceleration and Speed data corresponding to each traffic accident to obtain dangerous driving event data with 24 characteristic variables. Summarize the dangerous driving event data corresponding to all traffic accidents and use them as independent variables, recorded as:

_Xi ＝[Times _i , Acceleration _i , Speed _i ]

Among them, Times _i represents the frequency of occurrence of four dangerous driving events corresponding to the i-th traffic accident, and Acceleration _i and Speed _i have the same meaning.

For each accident case data, a label y = 1 is assigned and it is used as the dependent variable, denoted as Y = [y _i ], y _i = 1. The independent variable and the dependent variable are matched to obtain the accident case data set.

Specifically, step S3 is:

S3.1: For each traffic accident, according to the methods in S2.2, S2.3 and S2.4, extract and calculate the dangerous driving event data within the spatial range of l upstream and downstream of the accident site within the time range of t _{0 before the accident, t 0 ~} 2t ₀ , ..., (n-1)t ₀ ~nt ₀ , and summarize the initial non-accident case data. Where n is a positive integer, which controls the proportion of the control group selection and can be selected according to actual conditions.

In order to avoid the impact of traffic accidents within the above time range, check the time of all non-accident case data and delete the non-accident case data between t time before any traffic accident and t′ time after the accident. Delete the cases with missing information to obtain the final non-accident case data, which is used as the independent variable and recorded as:

X _j = [Times _j , Acceleration _j , Speed _j ].

Among them, Times _j represents the frequency of occurrence of four dangerous driving events corresponding to the jth non-accident case, and Acceleration _j and Speed _j have the same meaning.

For each non-accident case data, the label y = 0 is assigned and used as the dependent variable, denoted as Y = [y _j ], y _j = 0. The independent variable and the dependent variable are matched to obtain the non-accident case data set.

S3.2: Combine the accident case dataset and the non-accident case dataset to obtain a sample dataset, recorded as:

X＝[Times _k ,Acceleration _k ,Speed _k ]

=[L_up _k ,L_down _k ,R_up _k ,R_down _k ,A_up _k ,A_down _k ,B_up _k ,B_down _k ,La_up _k ,La_down _k ,Ra_up _k ,Ra_down _k ,Aa_up _k ,Aa_down _k ,Ba_up _k ,Ba_down _k ,Ls_up _k ,Ls_down _k ,Rs_up _k ,Rs_down _k ,As_up _k ,As_down _k ,Bs_up _k ,Bs_down _k ]

Y＝[y _k ], y _k ＝0 or 1

Specifically, step S4 is:

S4.1: Calculate the mean and standard deviation of each independent variable and standardize all data as follows:

x _norm = (x-μ)/σ

Among them, μ is the mean of the independent variable and σ is the standard deviation of the independent variable.

S4.2: Process each dangerous driving event data into the following four-channel one-dimensional image format. The variables in each channel are as follows:

Channel 1: L_up, La_up, Ls_up, L_down, La_down, Ls_down

Channel 2: R_up, Ra_up, Rs_up, R_down, Ra_down, Rs_down

Channel 3: A_up, Aa_up, As_up, A_down, Aa_down, As_down

Channel 4: B_up, Ba_up, Bs_up, B_down, Ba_down, Bs_down

Among them, the six variables of the same driving behavior type are located in the same channel, and are arranged in the order of upstream and downstream road section relationship (from up to down) and the frequency of dangerous driving events, average maximum acceleration and average maximum speed. The variables of different types of dangerous driving events are located in four different channels. This data format is conducive to the model to extract and learn the intrinsic connection between the data in different types of dangerous driving events and upstream and downstream spatial locations, and then to mine the hidden patterns in the data.

S4.3: Establish a deep learning model. Convolutional layer, pooling layer, flattening layer, fully connected layer, and output layer are constructed in sequence. The four-channel one-dimensional image data obtained in S4.2 is first input into the convolutional layer as the input data of the deep learning model.

The convolution layer (Conv1D layer) has an input dimension of I ₁ and an output dimension of O ₁ , and has m ₁ convolution kernels of size n _1. These convolution kernels perform one-dimensional convolution operations along the direction of the variable arrangement in the input data channel (for example, from L_up to Ls_down) with a step size of s _1. The activation function of the convolution layer is relu. The output of this module is used as the input of the pooling layer.

The input dimension of the pooling layer (AveragePooling1D layer) is I ₂ , the output dimension is O ₂ , and a one-dimensional average pooling operation with a range of n _{2 is performed with a step size of s 2. The} output of this module is used as the input of the flattening layer.

The input dimension of the flatten layer is I ₃ and the output dimension is O ₃ . It converts the multi-channel one-dimensional image data output by the pooling layer into a one-dimensional vector form, and uses a random drop method with a ratio of p to discard data to avoid overfitting. The output of this module is used as the input of the fully connected layer.

The input dimension of the fully connected layer is I ₄ , the output dimension is O ₄ , and it has m ₂ neurons. The activation function of the fully connected layer is relu. At the same time, a random drop method with a ratio of p _drop is used to drop neurons. The output of this module is used as the input of the output layer.

The output layer has 1 neuron. The activation function of the output layer is sigmoid, which converts the output value to the interval (0,1). Therefore, the output data of the entire deep learning model means the probability of a traffic accident occurring on the current research section.

S4.4: Train the established deep learning model based on the sample data set. Using the area under the receiver operating characteristic curve (AUC), sensitivity, and specificity as evaluation indicators, the ten-fold cross-validation method, binary cross entropy loss function, and Nadam optimizer are used to tune the established deep learning model based on the sample data set to obtain the deep learning model with the best effect.

Specifically, step S5 is:

S5.1: Real-time acquisition of four types of dangerous driving events data on the entire highway: sudden acceleration, sudden deceleration, sudden left lane change and sudden right lane change. Based on S2, the frequency of occurrence of the four types of dangerous driving events Times′ within the time range before t at each location on the highway and the average value of the maximum acceleration Acceleration′ and maximum speed Speed′ of the vehicle in all the same types of dangerous driving events are calculated. The input data is summarized.

S5.2: Input the data into the deep learning model constructed in S4 to obtain the probability p∈(0,1) of a traffic accident occurring at each location on the current highway, which is the traffic accident risk assessment value.

When p∈(0,threshold), the traffic accident risk level is low and no intervention is required.

When p∈[threshold,1), the traffic accident risk level is high, and corresponding measures need to be taken to control traffic safety risks.

Among them, threshold∈(0,1).

The present invention also proposes a highway traffic accident risk assessment system based on dangerous driving event data, comprising:

A data acquisition module is used to determine the research section and statistical time range, and to acquire the dangerous driving incident data and traffic accident data on the research section of the expressway within the statistical time range;

The sample construction module is used to find the dangerous driving event data of the upstream and downstream sections of the accident site before each traffic accident, and use the dangerous driving event data as the independent variable and the accident as the dependent variable to construct the accident case data set; then find the dangerous driving event data of the upstream and downstream sections when no traffic accident occurs, use the dangerous driving event data as the independent variable and the accident as the dependent variable to construct a non-accident case data set as a control group, and merge the accident case data set and the non-accident case data set to obtain the sample data set;

The model training module is used to standardize the data, process the data into images according to different types of dangerous driving events and upstream and downstream spatial relationships, and build a deep learning model for training and parameter adjustment until the model effect is optimal;

The road section accident risk assessment module is used to obtain real-time data on dangerous driving incidents of vehicles on all sections of the highway, calculate the frequency of dangerous driving incidents and the average value of the maximum acceleration and maximum speed of vehicles in all similar dangerous driving incidents, input them into the constructed deep learning model, and calculate the traffic accident risk assessment value and risk level of each section.

The present invention adopts the above technical solution, which has the following beneficial technical effects compared with the prior art:

The highway traffic accident risk assessment method based on dangerous driving event data proposed in the present invention is not limited by the location and density of fixed detectors, can cover the entire section of the highway and is not affected by weather conditions. It can perform real-time and refined assessment of traffic accident risks at each section of the highway without installing new equipment.

At the same time, the method proposed in the present invention comprehensively considers the intrinsic connection between dangerous driving event data in terms of different event types and upstream and downstream spatial positions, and evaluates the accident risk of road sections based on two factors: the frequency of dangerous driving events on the road and the vehicle motion parameters in dangerous driving events. This can improve the accuracy of risk assessment and provide new theoretical and technical support for highway traffic management departments to control traffic safety risks.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a flow chart of a highway traffic accident risk assessment method according to the present invention.

FIG. 2 is a schematic diagram of rules for extracting dangerous driving event data before a traffic accident occurs in the present invention.

FIG3 is a schematic diagram of the deep learning model framework in the present invention.

Detailed ways

The technical solutions in the embodiments of the present invention will be described in detail and completely in conjunction with the accompanying drawings. In addition, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, other embodiments obtained by ordinary technicians in this field without creative work are all within the scope of protection of the present invention.

FIG1 is a flow chart of a highway traffic accident risk assessment method based on dangerous driving event data of the present invention. The method can be applied to highway traffic accident risk assessment with traffic accident data and dangerous driving event data including vehicle motion parameters. The specific steps are as follows:

S1: Determine the research section and statistical time range, and obtain the following two types of information:

(1) Data on dangerous driving incidents of vehicles on the research section of the expressway within the statistical time range, including the time of occurrence of the dangerous driving incident, the longitude, latitude and up and down directions of the location of the incident, the type of dangerous driving incident (sudden acceleration, sudden deceleration, sudden left lane change and sudden right lane change), and the maximum acceleration and maximum speed of the vehicle during the dangerous driving incident.

(2) Traffic accident data occurring on the research section of the expressway within the statistical time range, including the start and end time of the accident, the longitude, latitude, and up and down directions of the accident location.

S2: For each traffic accident, find the dangerous driving event data of the upstream and downstream sections of the accident site before the accident. Use the dangerous driving event data as the independent variable and the accident as the dependent variable to construct the accident case data set. The specific steps are as follows:

S2.1: According to the longitude, latitude, and up and down directions of the dangerous driving incidents and traffic accidents, mark them on the corresponding sections of the highway and establish the upstream and downstream relationships between the sections.

S2.2: As shown in Figure 2, based on the spatiotemporal location of each traffic accident, extract the data of four dangerous driving events: sudden acceleration, sudden deceleration, sudden left lane change and sudden right lane change within a range of 1 meter upstream and downstream of the accident location within t minutes before the accident, i.e. the part circled by the oval dotted box in Figure 2. Usually t≤30,l≤1000.

S2.3: For each traffic accident, calculate the following:

(1) Calculate the frequency of occurrence of four types of dangerous driving events within the time and space range shown in Figure 2, recorded as:

Times = [L_up, L_down, R_up, R_down, A_up, A_down, B_up, B_down]

The first part of each variable represents the type of dangerous driving event, i.e., L, R, A, and B represent the frequency of sudden left lane change, sudden right lane change, sudden acceleration, and sudden deceleration in the time range before the accident. The second part of each variable represents the location where the dangerous driving event occurred, i.e., up and down represent the road sections within the spatial range upstream and downstream of the accident location, respectively.

(2) Calculate the average value of the maximum acceleration and maximum speed of the vehicle in the same type of dangerous driving incident within the time and space range shown in Figure 2, and record it as:

Acceleration = [La_up, La_down, Ra_up, Ra_down, Aa_up, Aa_down, Ba_up, Ba_down]

Speed = [Ls_up, Ls_down, Rs_up, Rs_down, As_up, As_down, Bs_up, Bs_down]

Among them, La and Ls in the first part of each variable represent the average values of the maximum acceleration and maximum speed of the vehicle in all sharp left lane change events, and the meanings of the remaining Ra, Rs, Aa, As, Ba, and Bs are similar. The second part of each variable represents the location where the dangerous driving event occurred.

S2.4: Combine the Times, Acceleration and Speed data corresponding to each traffic accident to obtain dangerous driving event data with 24 characteristic variables. Summarize the dangerous driving event data corresponding to all traffic accidents and use them as independent variables, recorded as:

_Xi ＝[Times _i , Acceleration _i , Speed _i ]

Among them, Times _i represents the frequency of occurrence of four dangerous driving events corresponding to the i-th traffic accident, and Acceleration _i and Speed _i have the same meaning.

For each accident case data, a label y = 1 is assigned and it is used as the dependent variable, denoted as Y = [y _i ], y _i = 1. The independent variable and the dependent variable are matched to obtain the accident case data set.

S3: Find the dangerous driving event data of the upstream and downstream sections when no traffic accidents occurred. Take the dangerous driving event data as the independent variable and the non-accident data as the dependent variable to construct a non-accident case data set as the control group. Merge the accident case data set and the non-accident case data set to obtain a sample data set. The specific steps are as follows:

S3.1: For each traffic accident, according to the methods in S2.2, S2.3 and S2.4, extract and calculate the dangerous driving event data within the spatial range of l upstream and downstream of the accident site within the time range of t ₀ before the accident, t ₀ ~2t ₀ , ..., (n-1)t ₀ ~nt ₀ , and summarize the initial non-accident case data. Usually t ₀ ≥1 and t ₀ is a positive integer. n is a positive integer that controls the proportion of the control group selection and can be selected according to actual conditions.

In order to avoid the impact of traffic accidents within the above time range, check the time of all non-accident case data and delete the non-accident case data between t time before any traffic accident and t′ time after the accident. Delete the cases with missing information to obtain the final non-accident case data, which is used as the independent variable and recorded as:

X _j =[Times _j ,Acceleration _j ,Speed _j ].

Among them, Times _j represents the frequency of occurrence of four dangerous driving events corresponding to the jth non-accident case, and Acceleration _j and Speed _j have the same meaning.

For each non-accident case data, the label y = 0 is assigned and used as the dependent variable, denoted as Y = [y _j ], y _j = 0. The independent variable and the dependent variable are matched to obtain the non-accident case data set.

S3.2: Combine the accident case dataset and the non-accident case dataset to obtain a sample dataset, recorded as:

X＝[Times _k ,Acceleration _k ,Speed _k ]

=[L_up _k ,L_down _k ,R_up _k ,R_down _k ,A_up _k ,A_down _k ,B_up _k ,B_down _k ,La_up _k ,La_down _k ,Ra_up _k ,Ra_down _k ,Aa_up _k ,Aa_down _k ,Ba_up _k ,Ba_down _k ,Ls_up _k ,Ls_down _k ,Rs_up _k ,Rs_down _k ,As_up _k ,As_down _k ,Bs_up _k ,Bs_down _k ]

Y＝[y _k ], y _k ＝0 or 1

S4: Standardize the data. Process the dangerous driving event data into images according to different dangerous driving event types and spatial position relationships, build a deep learning model for training and parameter adjustment until the model effect is optimal. The specific steps are as follows:

S4.1: Calculate the mean and standard deviation of each independent variable and standardize all data as follows:

x _norm = (x-μ)/σ

Among them, μ is the mean of the independent variable and σ is the standard deviation of the independent variable.

S4.2: Process each dangerous driving event data into a four-channel one-dimensional image format, as shown in the input layer of Figure 3 (due to space limitations, some variable names are not marked). The variables in each channel are as follows:

Channel 1: L_up, La_up, Ls_up, L_down, La_down, Ls_down

Channel 2: R_up, Ra_up, Rs_up, R_down, Ra_down, Rs_down

Channel 3: A_up, Aa_up, As_up, A_down, Aa_down, As_down

Channel 4: B_up, Ba_up, Bs_up, B_down, Ba_down, Bs_down

Among them, the six variables of the same driving behavior type are located in the same channel, and are arranged in the order of upstream and downstream road section relationship (from up to down) and the frequency of dangerous driving events, average maximum acceleration and average maximum speed. The variables of different types of dangerous driving events are located in four different channels. This data format is conducive to the model to extract and learn the intrinsic connection between the data in different types of dangerous driving events and upstream and downstream spatial locations, and then to mine the hidden patterns in the data.

S4.3: As shown in Figure 3, the convolution layer, pooling layer, flattening layer, fully connected layer, and output layer are constructed in sequence. The four-channel one-dimensional image data obtained in S4.2 is first input into the convolution layer as the input data of the deep learning model.

The convolution layer (Conv1D layer) has an input dimension of I ₁ and an output dimension of O ₁ , and has m ₁ convolution kernels of size n _1. These convolution kernels perform one-dimensional convolution operations along the direction of the variable arrangement in the input data channel (the direction of the input layer arrow in Figure 3) with a step size of s _1. The activation function of the convolution layer is relu. The output of this module is used as the input of the pooling layer.

The pooling layer (AveragePooling1D layer) has an input dimension of I ₂ and an output dimension of O ₂ , and performs a one-dimensional average pooling operation with a range of n ₂ with a step size of s ₂ to reduce the data dimension and avoid overfitting. The output of this module is used as the input of the flattening layer.

The input dimension of the flatten layer is I ₃ and the output dimension is O ₃ . It converts the multi-channel one-dimensional image data output by the pooling layer into a one-dimensional vector form, and uses a random drop method with a ratio of p to discard data to avoid overfitting. The output of this module is used as the input of the fully connected layer.

The input dimension of the fully connected layer is I ₄ , the output dimension is O ₄ , and it has m ₂ neurons. The activation function of the fully connected layer is relu. At the same time, a random drop method with a ratio of p _drop is used to drop neurons. The output of this module is used as the input of the output layer.

The output layer has 1 neuron. The activation function of the output layer is sigmoid, which converts the output value to the interval (0,1). Therefore, the output data of the entire deep learning model means the probability of a traffic accident occurring on the current research section.

S4.4: Train the established deep learning model based on the sample data set. The area under the receiver operating characteristic curve (AUC), sensitivity and specificity are used as evaluation indicators to evaluate the effect of the deep learning model. Specifically, the Youden index method is used to calculate the Youden index under each classification threshold, and the calculation formula is as follows:

Youden＝Sensitivity+Specificity-1

Among them, TP represents the number of accidents that are both actual and predicted, FN represents the number of accidents that are actually predicted to be non-accidents, TN represents the number of non-accidents that are actually predicted to be, and FP represents the number of non-accidents that are actually predicted to be accidents. The threshold corresponding to the maximum Youden index is selected as the classification threshold of the model, and the sensitivity and specificity of the model under this threshold are used as model evaluation indicators. The closer the values of the model's AUC, sensitivity, and specificity are to 1, the better the model's prediction performance is.

Based on the above three evaluation indicators, the ten-fold cross validation method, binary cross entropy loss function and Nadam optimizer are used to tune the established deep learning model based on the sample data set to obtain the deep learning model with the best effect.

S5: Real-time acquisition of dangerous driving incident data of vehicles on all sections of the highway, inputting it into the deep learning model constructed in S4, and calculating the traffic accident risk assessment value and risk level of each section. The specific steps are as follows:

S5.1: Obtain the data of four dangerous driving events, namely, sudden acceleration, sudden deceleration, sudden left lane change and sudden right lane change, on all sections of the highway in real time, and establish the upstream and downstream relationship between sections. According to the calculation method in S2, obtain the frequency of occurrence of the four dangerous driving events Times′ within the time range before t at each position of the highway, as well as the average value of the maximum acceleration Acceleration′ and the maximum speed Speed′ of the vehicle in all the same type of dangerous driving events, and summarize them to obtain the input data X of the model.

S5.2: Input X into the deep learning model constructed in S4 to obtain the probability p∈(0,1) of a traffic accident occurring at each location on the current highway, which is the traffic accident risk assessment value.

When p∈(0,threshold), the traffic accident risk level is low and no intervention is required.

When p∈[threshold,1), the traffic accident risk level is high, and corresponding measures need to be taken to control safety risks.

Among them, threshold∈(0,1).

Specific case

In order to demonstrate the applicability of the highway traffic accident risk assessment method based on dangerous driving event data provided by the embodiment of the present invention in real scenarios, the following specific case is given for further explanation:

Taking the Liyang-Changxing section of China's G25 Expressway as an example, the data collection time range is from September 26, 2020 to October 2, 2020. The dangerous driving incident data of vehicles on this section are collected, and the time interval for data collection is 1 second. The data fields include the time of occurrence of the dangerous driving incident, the longitude, latitude and uplink and downlink directions of the location of the incident, the type of dangerous driving incident (sudden acceleration, sudden deceleration, sudden left lane change and sudden right lane change), and the motion parameters of the vehicle in the dangerous driving incident (maximum acceleration and maximum speed). The traffic accident data on this section within the time range are collected, and the data fields include the start time and end time of the accident, the longitude, latitude and uplink and downlink directions of the location of the accident. After sorting, a total of 371 traffic accidents were recorded on the research section within the statistical scope.

Since the recorded time of traffic accidents is often slightly later than the actual time of occurrence, a short time interval for the aggregation of dangerous driving event data will affect the accuracy of the prediction. In addition, when the aggregation time interval is short, the number of dangerous driving events in each interval is small and the distribution difference in different intervals is large, which is not conducive to modeling. Therefore, this case selects 30 minutes as the time interval for data aggregation. In addition, 250 meters is selected as the spatial interval for the selection of dangerous driving event data, which is convenient for fine extraction of dangerous driving events that have a greater impact on traffic accidents. That is, t = 30, l = 250.

According to step S2 of the specific implementation method, the dangerous driving event data that occurred in the upstream and downstream sections before the accident was extracted and data fusion was performed to obtain 265 accident cases. For each traffic accident, according to step S3 of the specific implementation method, the dangerous driving event data within 250 meters upstream and downstream of the accident site within the range of 0 to 1, 1 to 2, and 2 to 3 hours before the accident (i.e., t ₀ =1) was extracted and calculated as the non-accident case corresponding to the traffic accident, and the initial non-accident case data was obtained by summarizing.

To avoid the influence of accident cases, the time of all non-accident cases was checked, and the non-accident case data with a time between 30 minutes before and 180 minutes after any traffic accident was deleted. Cases with missing information were deleted, and the final 722 non-accident case data were obtained, forming a sample data set with an accident-to-non-accident ratio of approximately 1:3.

According to the step S4 of the specific implementation method, the data is standardized, each dangerous driving event data is processed into a four-channel one-dimensional image form, and a deep learning model including a one-dimensional convolution module is established. The specific parameters of the model are as follows:

The input dimension of the convolutional layer is (N, 6, 4), with 512 convolution kernels of size 3. These convolution kernels perform one-dimensional convolution operations with a stride of 1. Therefore, the output dimension is (N, 4, 512), where N is the number of samples in the training set.

The pooling layer performs a one-dimensional average pooling operation with a range of 3 and a stride of 1. Therefore, the output dimension is (N, 2, 512).

The flattening layer converts the multi-channel one-dimensional image data output by the pooling layer into a one-dimensional vector form, and uses a random dropout method with a ratio of 0.5 to discard data to avoid overfitting. Therefore, the output dimension is (N, 1024).

The fully connected layer has 32 neurons and the activation function is relu. Neurons are dropped randomly with a rate of 0.5. Therefore, the output dimension is (N,32).

The output layer has 1 neuron, and the output dimension is (N, 1), which is a one-dimensional vector. The activation function of the output layer is sigmoid. The output data means the probability of a traffic accident occurring on the current research section.

The ten-fold cross validation method, binary cross entropy loss function and Nadam optimizer were used to train the deep learning model and perform hyperparameter tuning to obtain the deep learning model with the best effect.

The validation set data is used as the real-time data of dangerous driving events of vehicles on the entire section of the highway, and according to the step S5 of the specific implementation method, it is input into the trained deep learning model to calculate the traffic accident risk assessment value. Experiments show that compared with the three benchmark models of logistic regression, support vector machine and artificial neural network model (input data only supports one-dimensional vector form), the deep learning model proposed in the present invention (input data is in image form) has the highest AUC value, and can identify 71.3% of accident situations and 75.1% of non-accident situations under the optimal threshold. It is the only model with a sensitivity and specificity greater than 70%. At the same time, the deep learning model has the highest sensitivity under a small false alarm rate (less than 30%). The experimental results of the specific case prove that the comprehensive performance of the highway accident risk assessment method based on dangerous driving event data proposed in the present invention is the best, and it can extract and learn the inherent connection between dangerous driving event data in different event types and upstream and downstream spatial positions, effectively improve the accuracy of risk assessment, and has a high application value.

The above description is only a specific embodiment of the present invention, which is only used to help understand the principle and core idea of the present invention, and is not used to limit the present invention. Within the idea and principle of the present invention, any modification, equivalent replacement, etc. made by any technician in the field to the technical solution described in the present invention should be included in the protection scope of the present invention.

Claims (3)

1. The highway traffic accident risk assessment method based on the dangerous driving event data is characterized by comprising the following steps of:

S1, determining a research road section and a statistical time range, and acquiring vehicle dangerous driving event related data on the expressway research road section and traffic accident data on the research road section in the statistical time range;

S2, aiming at each traffic accident, searching dangerous driving event data of an upstream road section and a downstream road section of an accident site before the accident occurs, taking the dangerous driving event data as independent variables, and taking the accident as the dependent variables to construct an accident case data set;

S3, searching dangerous driving event data of an upstream road section and a downstream road section when no traffic accident occurs, taking the dangerous driving event data as an independent variable, taking the non-occurring accident as a dependent variable, constructing a non-accident case data set as a comparison group, and combining the accident case data set and the non-accident case data set to obtain a sample data set;

s4, carrying out standardized processing on the data: processing dangerous driving event data into an image form according to different dangerous driving event types and spatial position relations, constructing a deep learning model, and training and parameter adjustment until the model effect is optimal;

s5, acquiring dangerous driving event data of vehicles on all road sections of the expressway in real time, calculating to obtain occurrence frequency of various dangerous driving events and average value of maximum acceleration and maximum speed of the vehicles in the dangerous driving events according to the S2, inputting the average value into a deep learning model constructed in the S4, and calculating to obtain traffic accident risk assessment values and risk levels of all road sections;

The data acquired in step S1 includes:

(1) Dangerous driving event related data of vehicles on an expressway, comprising: the occurrence time of dangerous driving events, the longitude, latitude and uplink and downlink directions of the occurrence place, the type of the dangerous driving events, and the maximum acceleration and maximum speed of the vehicle in the dangerous driving events; dangerous driving event types include: rapid acceleration, rapid deceleration, rapid left lane change and rapid right lane change;

(2) Highway traffic accident data comprising: the starting time and the ending time of the accident, the longitude and the latitude of the accident place and the uplink and downlink directions;

The step S2 specifically comprises the following sub-steps:

S201, marking dangerous driving events and traffic accident positions on corresponding road sections of the expressway according to longitude, latitude and uplink and downlink directions, and establishing an uplink and downlink relation between the road sections;

S202, according to the space-time position of each traffic accident, four dangerous driving event data of rapid acceleration, rapid deceleration, rapid left lane change and rapid right lane change in the space range of the upstream and downstream of the accident site in the time range of t before the accident occurs are extracted;

S203, for each traffic accident, respectively calculating the following contents:

(1) Calculating occurrence frequencies of four dangerous driving events in the space-time range in the step S202, and marking as:

Times＝[L_up,L_down,R_up,R_down,A_up,A_down,B_up,B_down]

Wherein a first part of each variable represents a dangerous driving event type, namely L, R, A and B respectively represent the frequency of occurrence of sudden left lane change, sudden right lane change, sudden acceleration and sudden deceleration driving events in a time range t before an accident occurs, and a second part of each variable represents a dangerous driving event occurrence place, namely up and down respectively represent road sections in a space range of an upstream and a downstream of the accident place;

(2) Calculating the average value of the maximum acceleration and the maximum speed of the vehicle in the same type of dangerous driving event in the space-time range in the step S202, and marking as:

Acceleration＝[La_up,La_down,Ra_up,Ra_down,Aa_up,Aa_down,Ba_up,Ba_down]

Speed＝[Ls_up,Ls_down,Rs_up,Rs_down,As_up,As_down,Bs_up,Bs_down]

Wherein La, ls in the first portion of each variable represent average values of maximum acceleration and maximum velocity of the vehicle in all rapid left lane change events, ra, rs represent average values of maximum acceleration and maximum velocity of the vehicle in all rapid right lane change events, aa, as represent average values of maximum acceleration and maximum velocity of the vehicle in all rapid acceleration events, ba, bs represent average values of maximum acceleration and maximum velocity of the vehicle in all rapid deceleration events, respectively;

S204: combining the time, acception and Speed data corresponding to each traffic accident to obtain dangerous driving event data with 24 characteristic variables, summarizing the dangerous driving event data corresponding to all the traffic accidents, and recording the dangerous driving event data as independent variables as: x _i＝[Times_i,Acceleration_i,Speed_i ], wherein time _i represents the occurrence frequency of four dangerous driving events corresponding to the ith traffic accident, and acception _i and Speed _i represent the average value of the maximum Acceleration and the maximum Speed of the vehicle in the four dangerous driving events corresponding to the ith traffic accident respectively;

Each accident case data is given a label y=1, the label y=1 is taken as a dependent variable, the dependent variable is marked as Y= [ Y _i],y_i =1, and the independent variable and the dependent variable are matched to obtain an accident case data set;

the step S3 specifically comprises the following sub-steps:

s301: for each traffic accident, extracting and calculating dangerous driving event data in the space range of the upstream and downstream of the accident site in the time range of t ₀、t₀～2t₀、…、(n-1)t₀～nt₀ before the accident occurs, and summarizing to obtain initial non-accident case data, wherein n is a positive integer for controlling the proportion selected by a control group;

Checking the time of all initial non-accident cases, and deleting non-accident case data with the time between t time before any traffic accident and t' time after the accident; deleting the case with information missing to obtain final non-accident case data, and recording the final non-accident case data as an independent variable: x _j＝[Times_j,Acceleration_j,Speed_j ], wherein time _j represents the occurrence frequency of four dangerous driving events corresponding to the jth non-accident case, and acceletion _j and Speed _j represent the average value of the maximum Acceleration and the maximum Speed of the vehicle in the four dangerous driving events corresponding to the jth non-accident case, respectively;

Each piece of non-accident case data is given a label y=0, the label y=0 is taken as a dependent variable, the dependent variable is marked as Y= [ Y _j],y_j =0, and the independent variable and the dependent variable are matched to obtain a non-accident case data set;

S302, combining the accident case data set and the non-accident case data set to obtain a sample data set, and marking as:

X＝[Times_k,Acceleration_k,Speed_k]

＝[L_up_k,L_down_k,R_up_k,R_down_k,A_up_k,A_down_k,B_up_k,B_down_k,La_up_k,La_down_k,Ra_up_k,Ra_down_k,Aa_up_k,Aa_down_k,Ba_up_k,Ba_down_k,Ls_up_k,Ls_down_k,Rs_up_k,Rs_down_k,As_up_k,As_down_k,Bs_up_k,Bs_down_k]

Y= [ Y _k],y_k =0 or 1;

The step S4 specifically comprises the following sub-steps:

s401, carrying out standardization treatment on data according to x _norm = (x-mu)/sigma, wherein mu is an average value, and sigma is a standard deviation;

S402, processing each dangerous driving event data into a four-channel one-dimensional image form, wherein variables in each channel are as follows:

channels 1:L_up, la_up, ls_up, L_down, la_down, ls_down

Channel 2 R_up, ra_up, rs_up, R_down, ra_down, rs_down

Channels 3: A_up, aa_up, as_up, A_down, aa_down, as_down

Channels 4:B_up, ba_up, bs_up, B_down, ba_down, bs_down

The method comprises the steps that 6 variables of the same driving behavior type are located in the same channel, and are arranged according to the relation between an upstream road section and a downstream road section and the sequence of occurrence frequency, average maximum acceleration and average maximum speed of dangerous driving events, and the variables of different types of dangerous driving events are located in 4 different channels;

S403, establishing a deep learning model: sequentially constructing a convolution layer, a pooling layer, a flattening layer, a full-connection layer and an output layer, and inputting the four-channel one-dimensional image data obtained in the step S402 into the convolution layer as input data of a deep learning model; the convolution layer is provided with m ₁ convolution kernels with the size of n ₁, the convolution kernels carry out one-dimensional convolution operation along the variable arrangement direction in an input data channel in the step length of s ₁, the activation function is relu, and the output of the convolution layer is used as the input of the pooling layer;

The pooling layer performs one-dimensional average pooling operation with the range of n ₂ by the step length of s ₂, and the output of the pooling layer is used as the input of the flattening layer;

The flattening layer converts the multi-channel one-dimensional image data output by the pooling layer into a one-dimensional vector form, a random discarding method with the ratio of p is adopted to avoid overfitting, and the output of the flattening layer is used as the input of the full-connection layer;

The full-connection layer is provided with m ₂ neurons, the activation function is relu, the random discarding method with the ratio of p is adopted to avoid overfitting, the random discarding method with the ratio of p _drop is adopted, and the output of the full-connection layer is used as the input of the output layer;

The output of the full-connection layer is provided with 1 neuron, the activation function of the output layer is sigmoid, and the meaning of the output data is the probability of traffic accidents in the current research road section;

S404, training and optimizing the established deep learning model based on the sample data set: taking the area, sensitivity and specificity of the test subject under the working characteristic curve as evaluation indexes, and adopting a ten-fold cross validation method, a binary cross entropy loss function and Nadam optimizer for optimization to obtain a deep learning model with optimal effect;

The step S5 specifically comprises the following steps:

s501, obtaining data of four dangerous driving events of rapid Acceleration, rapid deceleration, rapid left lane change and rapid right lane change of a vehicle on a whole highway section in real time, calculating the occurrence frequency Times ' of the four dangerous driving events in a time range of t before each position of the highway according to the step S2, and collecting average values of the maximum Acceleration accelization ' and the maximum Speed ' of the vehicle in all the same type of dangerous driving events to obtain input data of a model;

S502: inputting data into the deep learning model constructed in the step S4, and calculating the probability p E (0, 1) of the occurrence of the traffic accident at each position of the current expressway, namely, the traffic accident risk assessment value;

When p is E (0, threshold), the traffic accident risk level is low risk, and no intervention is needed;

When p is epsilon (threshold, 1), the traffic accident risk level is high risk, and corresponding measures are needed to be adopted to manage and control the safety risk;

Wherein threshold E (0, 1).

2. A highway traffic accident risk assessment system based on the method of claim 1, comprising:

The data acquisition module is used for determining a research road section and a statistical time range and acquiring dangerous driving event data on the expressway research road section and traffic accident data on the research road section within the statistical time range;

The sample construction module is used for searching dangerous driving event data of an upstream road section and a downstream road section of an accident site before the accident occurs according to each traffic accident, taking the dangerous driving event data as independent variables and taking the accident as the dependent variables, and constructing an accident case data set; searching dangerous driving event data of an upstream road section and a downstream road section when no traffic accident occurs, taking the dangerous driving event data as an independent variable, taking the non-occurring accident as a dependent variable, constructing a non-accident case data set as a comparison group, and combining the accident case data set and the non-accident case data set to obtain a sample data set;

The model training module is used for carrying out standardized processing on the data, processing the data into an image form according to different dangerous driving event types and upstream and downstream spatial relations, constructing a deep learning model, and carrying out training and parameter adjustment until the model effect is optimal;

The road section accident risk assessment module is used for acquiring dangerous driving event data of vehicles on the whole road section of the expressway in real time, calculating the occurrence frequency of dangerous driving events and the average value of the maximum acceleration and the maximum speed of the vehicles in all the same type of dangerous driving events, inputting the average value into the constructed deep learning model, and calculating to obtain the traffic accident risk assessment value and the risk level of each road section.

3. The system of claim 2, wherein the model training module builds a deep learning model by sequentially building a convolutional layer, a pooling layer, a flattening layer, a fully-connected layer, and an output layer, wherein

The convolution layer is provided with m ₁ convolution kernels with the size of n ₁, the convolution kernels carry out one-dimensional convolution operation along the variable arrangement direction in the input data channel in the step length of s ₁, the activation function is relu, and the output of the convolution layer is used as the input of the pooling layer;

The pooling layer performs one-dimensional average pooling operation with the range of n ₂ by the step length of s ₂, and the output of the pooling layer is used as the input of the flattening layer;

The flattening layer converts the multi-channel one-dimensional image data output by the pooling layer into a one-dimensional vector form, a random discarding method with the ratio of p is adopted to avoid overfitting, and the output of the flattening layer is used as the input of the full-connection layer;

The full-connection layer is provided with m ₂ neurons, the activation function is relu, the random discarding method with the ratio of p is adopted to avoid overfitting, the random discarding method with the ratio of p _drop is adopted, and the output of the flattening layer is used as the input of the output layer;

The output of the full-connection layer is provided with 1 neuron, the activation function of the output layer is sigmoid, and the meaning of the output data is the probability of traffic accidents in the current research road section.