A dynamic calibration vehicle collision risk real-time prediction model construction method, platform, medium and terminal

Abstract

The application discloses a kind of dynamic calibration vehicle collision risk real-time prediction model construction method, platform, medium and terminal, wherein the method obtains the vehicle trajectory data of road section or node by sensing device, and extracts characteristic parameters;Calculate the expansion distance collision time of each vehicle and the nearest preceding vehicle at each time, determine the collision risk of each vehicle and the nearest preceding vehicle and extract the collision risk influencing factors;Collision risk and characteristic parameters are sampled in different ways to obtain multiple discrete data sets;Risk prediction model is constructed and model parameters are dynamically calibrated;The prediction effect of each model is evaluated using the area under the ROC curve, and the optimal sampling method and learning rate of the prediction are obtained, to determine the final dynamic vehicle risk prediction model. The problem that historical data estimation of vehicle accident risk cannot adapt to the traffic environment and the model cannot be corrected in real time with data calibration is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of traffic safety, and in particular to a dynamic calibration vehicle collision risk real-time prediction model construction method, platform, medium and terminal. BACKGROUND

[0002] With the increasing traffic flow, the frequency of vehicle collision continues to rise, and the road system has serious traffic safety problems. Among them, the road safety control section or node as the main accident black spot of the road, its safety problem also attracts much attention. Due to the limited space and complex lane configuration, the driver needs to make many quick decisions, which may lead to the driver's quick change of vehicle speed and sudden lane change operation, greatly increasing the collision risk of the road safety control section or node.

[0003] The prior art mainly predicts the collision risk of the road safety control section or node based on large sample high frequency offline data. This kind of method is based on a large amount of continuous data, uses offline static model for prediction to obtain high-precision prediction results. That is, all video information in a certain space-time is extracted and analyzed offline. The data obtained by this method is large, which can reach sub-second level (i.e. not reaching the speed of seconds), which is approximately continuous data. When the data is large sample continuous data, the offline static model is estimated, and the prediction accuracy is undoubtedly higher. However, this kind of method has obvious limitations in safety management practice: first, the high-precision model estimated is only applicable to large sample data corresponding to it. When the sample data changes, the portability of the model may not meet the requirements, which may result in poor prediction performance of the estimated model for new sample data, and the model cannot be used. Secondly, the offline static estimation method cannot help the real-time risk prediction and safety control of the road safety control section or node. In the actual running state of the vehicle, the whole process of video image recognition, data acquisition, data cleaning, data analysis, model estimation cannot be completed in sub-second, and a large amount of continuous data cannot be obtained in real time to estimate the vehicle conflict and statically estimate the model. Finally, estimating the vehicle accident risk with historical data cannot adapt to the rapid changes of traffic environment and traffic flow, especially for toll stations in the operation mode transition stage. Realizing the correction of the accident risk prediction model with data calibration helps to capture the dynamic changes of vehicle safety and accurately and effectively complete the safety monitoring task of each development stage.

[0004] Therefore, there is an urgent need for a road safety control section or node vehicle collision risk real-time prediction construction method that can reflect the spatio-temporal variation characteristics of the traffic condition, and the model can be real-time self-adaptive correction with data calibration. SUMMARY

[0005] In view of the deficiencies of the prior art, the present application provides a dynamic calibration vehicle collision risk real-time prediction model construction method, platform, medium and terminal, which solves the problems that the historical data estimation of vehicle accident risk cannot adapt to the rapid changes of traffic environment and traffic flow, and the model cannot be real-time self-adaptive correction with data calibration.

[0006] In a first aspect, the present application provides a vehicle collision risk real-time prediction model construction method for road safety management and control section or node, comprising:

[0007] S1: obtaining vehicle trajectory data of the safety management and control section or node through a perception device, and extracting vehicle feature parameters and traffic flow feature parameters;

[0008] S2: based on the vehicle feature parameters and the traffic flow feature parameters, calculating the expansion distance collision time of each vehicle and the nearest preceding vehicle at each time, determining the collision risk of each vehicle and the nearest preceding vehicle and extracting the collision risk influencing factors;

[0009] S3: interval sampling of the vehicle feature parameters, traffic flow feature parameters and collision risk obtained from all vehicle trajectory data in multiple different ways to obtain corresponding multiple discrete data sets;

[0010] S4: taking the collision risk influencing factors in the discrete data set as the input and the collision risk of each vehicle and the nearest preceding vehicle as the output, constructing a vehicle collision risk prediction model based on the multiple discrete data sets obtained in step S3, and dynamically calibrating the model parameters to obtain multiple vehicle collision risk real-time prediction models;

[0011] S5: using the area of the ROC curve to evaluate the prediction effect of each vehicle collision risk real-time prediction model to obtain the optimal sampling method and learning rate suitable for safety prediction of the safety management and control section or node, and determine the final vehicle collision risk real-time prediction model.

[0012] Further, the specific process of S1 is:

[0013] S11: collecting vehicle trajectory sample data of the safety management and control section or node through the perception device arranged at the road safety management and control section or node, and extracting the complete motion trajectory of the vehicle for data preprocessing;

[0014] S12: extracting the vehicle trajectory data of each vehicle after data preprocessing, including the position coordinates O of the vehicle at each time, the vehicle ID, and the time sequence identifier;

[0015] S13: based on the vehicle trajectory data of S12, further processing to obtain the vehicle feature parameters and traffic flow feature parameters of all vehicles; wherein the vehicle feature parameters include but are not limited to vehicle length L i , vehicle width Wi , vehicle speed V, vehicle speed direction V i , vehicle speed change rate V a , vehicle travel distance D, inter-vehicle distance D ij , vehicle type V type ; traffic flow characteristic parameters include but are not limited to average traffic flow F, average traffic density K, headway H.

[0016] Further, the collision risk influencing factors include but are not limited to: vehicle length L i , vehicle width W i , front vehicle speed L v , rear vehicle speed F v , vehicle travel distance D, inter-vehicle distance D ij , vehicle type V type , average traffic density K.

[0017] Further, the specific process of S2 is:

[0018] S21: Based on the vehicle trajectory data, establish a coordinate axis, match each vehicle at time t with its nearest front vehicle, and calculate the instantaneous traffic conflict ETTC between the matched vehicles, i.e. the extended distance collision time; the calculation formula of ETTC is as follows:

[0019]

[0020] Wherein, L i and L j are the lengths of vehicles i and j respectively; D ij is the distance between the center points of the two vehicles; d ij is the distance between the nearest points of the two vehicles; O and V are the two-dimensional coordinates and speed vectors of the vehicle center points.

[0021] S22: According to the discrimination condition of the instantaneous traffic conflict obtained by calculation, discriminate the collision risk of all matched vehicles;

[0022] Wherein, the discrimination condition is:

[0023] When ETTC is less than or equal to the preset threshold ETTC', the collision risk of the matched vehicle is that there is a collision risk;

[0024] When ETTC is greater than the preset threshold ETTC', the collision risk of the matched vehicle is then there is no collision risk.

[0025] Further, the specific process of S3 is:

[0026] S31: Set n different sampling time intervals, and divide the continuous data at equal intervals;

[0027] S32: On the basis of equally spaced data, the selection of sampling points for each interval data is performed;

[0028] The selection of sampling points includes two kinds, one is to select the data with time sequence number 1 in each sampling interval as the sampling data; the other is to randomly select the data with any time sequence number in each sampling interval as the sampling data;

[0029] S33: Based on n different sampling time intervals and two kinds of sampling points, K different sampling methods are obtained, where K=A2 n The vehicle feature parameters, traffic flow feature parameters and collision risk data obtained in all vehicle trajectory data samples are sampled to obtain K discrete data sets.

[0030] Further, S41: based on each discrete data set obtained in S3, a logistic regression model is defined using a preset proportion of data in the data set, i.e. a vehicle collision risk prediction model is constructed to predict the vehicle collision probability under the influence of various factors;

[0031] S42: the remaining data in the data set is input into the vehicle collision risk prediction model in turn for dynamic calibration to obtain multiple vehicle collision risk real-time prediction models, wherein each discrete data set corresponds to a vehicle collision risk real-time prediction model; the specific process of dynamic calibration is as follows:

[0032] S421: the time when the last data defining the logistic regression model is input is set to t-1, and this sample data is set to S t-1 At this time, the model parameters are Based on the current model parameters The sample data S t at time t, the vehicle collision risk data y t at time t, the sample loss function value L(y t , p(y t )) is calculated. Then, according to the loss function value L(y t , p(y t )), the sample S t is calculated. The gradient value of each vehicle collision risk prediction model parameter is calculated. The loss function is cross-entropy loss, and its formula is:

[0033] L(y t , p(y t )) = -y t log(p(y t )) - (1-y t )log(1-p(y t ))

[0034] wherein, p(y t ) is the probability of collision occurring under the sample S t ;

[0035] The gradient value calculation formula is:

[0036]

[0037] wherein, is the gradient value of the parameter , j = 1, 2, …, n, and n is the number of independent variables;

[0038] S422: calculating the model parameters at time t based on the gradient descent algorithm , that is, completing a dynamic parameter calibration, and the calculation formula is:

[0039]

[0040] wherein, is the logistic regression model parameter at time t, and a is the learning rate;

[0041] S423: judging whether the remaining data in the data set is sequentially traversed or not Input: if yes, ending the calibration; if no, t + 1 and returning to S421.

[0042] Further, the specific process of S5 is:

[0043] S51: constructing an ROC curve based on the true positive rate and false positive rate determined by comparing the prediction result of the vehicle collision risk real-time prediction model with the actual collision risk result calculated in S2; based on the vehicle collision risk real-time prediction model dynamically calibrated at each time, calculating the area under the ROC curve to obtain a series of AUC values of the results of each vehicle collision risk real-time prediction model respectively constructed by multiple discrete data sets;

[0044] S52: comparing the AUC value and the stability of each vehicle collision risk real-time prediction model:

[0045] S521: judging whether the AUC value is higher than a preset AUC threshold: if yes, it indicates that the prediction performance of the model is good; if no, it indicates that the prediction performance of the model is poor;

[0046] S522: inputting continuous multiple sample data into the vehicle collision risk real-time prediction model to judge whether the change of the AUC value is less than a preset standard threshold: if yes, it indicates that the stability of the vehicle collision risk real-time prediction model is good; if no, it indicates that the stability of the vehicle collision risk real-time prediction model is poor;

[0047] The interval sampling mode corresponding to the vehicle collision risk real-time prediction model with good prediction performance and stability is the optimal interval sampling mode;

[0048] S53: Sensitivity analysis is performed on the learning rate in the online gradient descent algorithm according to the loss function value and the AUC value, and an optimal learning rate is determined.

[0049] S54: A vehicle risk prediction model suitable for the road safety control section or node is obtained by combining the optimal interval sampling method and the optimal learning rate.

[0050] In a second aspect, the present application provides a road risk monitoring and early warning platform, characterized in that the platform is deployed with the steps of the method described above.

[0051] In a third aspect, the present application provides an electronic terminal comprising a processor and a memory, wherein the memory stores a computer program, and the processor invokes the computer program to execute the steps of the method described above.

[0052] In a fourth aspect, the present application provides a readable storage medium storing a computer program, wherein the computer program is invoked by a processor to execute the steps of the method described above.

[0053] Advantages

[0054] The present application provides a dynamic calibration vehicle collision risk real-time prediction model construction method, platform, medium and terminal, wherein the method establishes a collision risk prediction model based on an online gradient descent algorithm, realizes dynamic calibration of parameters with increasing sample size, reflects the state of real traffic flow, improves prediction accuracy, and breaks through the limitation that traditional methods cannot adapt to rapid changes in traffic environment and traffic flow; the method does not require large sample data, can process continuous data stream in real time, improves prediction efficiency, realizes real-time risk prediction and safety control, and reduces the data processing pressure of the safety monitoring system; the method can calibrate model parameters in real time, improve model prediction accuracy, and is suitable for real-time prediction scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0055] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0056] Figure 1 is a flowchart of a dynamic calibration vehicle collision risk real-time prediction model construction method provided by the embodiment of the present application. DETAILED DESCRIPTION

[0057] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be described in detail below. Obviously, the described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other implementation methods obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0058] Example 1

[0059] like Figure 1 As shown, this embodiment provides a method for constructing a dynamically calibrated real-time vehicle collision risk prediction model, including:

[0060] S1: Obtain vehicle trajectory data of safety-controlled road sections or nodes through sensing devices, and extract vehicle characteristic parameters and traffic flow characteristic parameters.

[0061] S11: Vehicle trajectory data of road safety control sections or nodes is collected through sensing devices, including but not limited to RSUs, cameras, millimeter-wave radar, lidar, and integrated radar-visual systems, and preprocessed. In specific implementations, the form of the vehicle trajectory data is closely related to the sensing devices used, including video or image data, signal data, coordinate data, etc., and is not limited thereto. The raw data is then transformed into standard formats such as CSV using image processing, signal processing, and data structuring technologies. This embodiment uses video data as an example to obtain the complete vehicle movement trajectory through image recognition and data structuring technologies.

[0062] S12: Extract the vehicle trajectory data of each vehicle after data preprocessing, including the vehicle's position coordinates O, vehicle ID, and time series identifier at each moment (in this embodiment, the time series identifier is the video frame ID).

[0063] S13: Based on the vehicle trajectory data from S12, further processing is performed to obtain the feature parameters of all vehicles. In this embodiment, coordinate transformation is performed on each frame of the video data to convert image coordinates into world coordinates, and the feature parameters of all vehicles are obtained through calculations between the coordinates. The vehicle feature parameters include vehicle feature parameters and traffic flow feature parameters; wherein the vehicle feature parameters include, but are not limited to, vehicle length L. i Vehicle width W i Vehicle speed V, vehicle speed direction V i Vehicle speed change rate V a Vehicle travel distance D, vehicle-to-vehicle distance D ij Vehicle type V type Traffic flow characteristic parameters include, but are not limited to, average traffic flow F, average traffic density K, and headway H.

[0064] S2: Based on the vehicle characteristic parameters and traffic flow characteristic parameters, the extended distance collision time of each vehicle and the nearest front vehicle at each time is calculated, the collision risk of each vehicle and the nearest front vehicle is determined, and the collision risk influencing factors are extracted.

[0065] S21: Based on the vehicle trajectory data, a coordinate axis is established, each vehicle in each frame t time video is matched with the nearest front vehicle, and the instantaneous traffic conflict ETTC, i.e. the extended distance collision time, between the matched vehicles is calculated. The calculation formula of ETTC is as follows:

[0066]

[0067] Wherein, L i and L j are the lengths of vehicles i and j; D ij is the distance between the centers of the two vehicles; d ij is the distance between the nearest points of the two vehicles; O and V are the two-dimensional coordinates and speed vectors of the vehicle center points.

[0068] S22: According to the discrimination condition of the calculated instantaneous traffic conflict, the collision risk of all matched vehicles is determined.

[0069] Wherein, the discrimination condition is:

[0070] When ETTC is less than or equal to the preset threshold ETTC', the collision risk of the matched vehicle is that there is a collision risk;

[0071] When ETTC is greater than the preset threshold ETTC', the collision risk of the matched vehicle is that there is no collision risk.

[0072] In this embodiment, when ETTC≤3 seconds, there is a collision risk; when ETTC>3 seconds, there is no collision risk.

[0073] Specifically, according to the vehicle interaction principle, the collision risk influencing factors are selected from four aspects of vehicle size and type, vehicle form speed, vehicle headway, and traffic flow fluctuation. The collision risk influencing factors include but are not limited to: vehicle length L i , vehicle width W i , front vehicle speed L v , rear vehicle speed F v , vehicle driving distance D, vehicle distance D ij , vehicle type V type, average traffic density K. The collision risk influencing factors are obtained by the Pearson correlation test, and variables with obvious correlation are removed. The influencing factor indicators are obtained by processing vehicle characteristic parameters and traffic flow characteristic parameters, and specifically, the coordinate position of each vehicle i is determined, the surrounding vehicles and their vehicle types, vehicle speeds, and distances between the two vehicles are determined, and are used as model collision risk influencing factor indicators.

[0074] S3: Intervals of the vehicle characteristic parameters, traffic flow characteristic parameters, and collision risks obtained from all vehicle trajectory data are sampled in multiple different ways to obtain corresponding multiple discrete data sets.

[0075] S31: n different sampling time intervals are set, and the continuous data is divided at equal intervals; in actual implementation, the number of n can be adjusted according to actual conditions, and in the embodiment, the number of n is set to 3: ΔT1, ΔT2, and ΔT3, that is, the sampling time interval can be selected at ΔT1 seconds, ΔT2 seconds, and ΔT3 seconds.

[0076] S32: On the basis of the equal-interval divided data, the sampling points in each interval data are selected;

[0077] The sampling point selection includes two kinds, one is to select the data with a time sequence number of 1 in each sampling interval as the sampling data, and in the embodiment, the first frame of data in the sampling interval is selected as the sampling data; the other is to randomly sample the data with any time sequence number in each sampling interval as the sampling data, and in the embodiment, any frame in the sampling interval is selected as the sampling data. In actual implementation, the random sampling method can be selected according to actual needs, and is not limited.

[0078] Preferably, the random selection adopts the random.choice() function in Python;

[0079] S33: Based on the n different sampling time intervals and the two kinds of sampling points, K different sampling methods are obtained, where K=A2 n The vehicle characteristic parameters, traffic flow characteristic parameters, and collision risks obtained from all vehicle trajectory data are sampled to obtain K discrete data sets. The data in the discrete data sets obtained after sampling are sample data. In the embodiment, since the sampling time length is set to 3, specifically, 1 second, 2 seconds, and 3 seconds, there are 6 sampling methods, namely C1, C2, C3, C4, C5, and C6, and finally 6 discrete data sets are obtained.

[0080] S4: The collision risk influencing factors in the discrete data sets are taken as inputs, the collision risks of each vehicle and the nearest preceding vehicle are taken as outputs, the multiple discrete data sets obtained in step S3 are used to construct vehicle collision risk prediction models, and the parameters of the models are dynamically calibrated to obtain multiple vehicle collision risk real-time prediction models.

[0081] S41: Based on each discrete data set obtained in S3, a logistic regression model is defined using a preset proportion of data in the data set, i.e., a vehicle collision risk prediction model is constructed to predict the vehicle collision probability under the influence of various factors. The formula of the vehicle collision risk prediction model is as follows:

[0082] y ~ Bernoulli (P)

[0083]

[0084]

[0085] where y is the prediction index, i.e., the vehicle collision risk, and takes a value of 0 to 1, 0 indicating no collision risk and 1 indicating a 100% collision risk; logit(P) is a logit function that converts the collision probability into a logit; P is the probability of collision; p(y n ) is a probability mass function representing the probability of each value of y; y n is the prediction index of all sample data; n is the number of all samples; X is the influencing factor of the collision risk; θ is the corresponding coefficient of each influencing factor, indicating the average change of logit(p) when X changes by a unit amount; k is the number of independent variables; and ε n is a random error that follows a normal distribution with a mean of zero.

[0086] In this embodiment, the preset proportion in the data set is set to 20%, and in actual implementation, it can be adjusted according to actual needs and is not limited.

[0087] S42: The remaining data in the data set is sequentially input into the vehicle collision risk prediction model for dynamic calibration to obtain a plurality of vehicle collision risk real-time prediction models, wherein each discrete data set corresponds to a vehicle collision risk real-time prediction model; and the specific process of dynamic calibration is as follows:

[0088] S421: The time when the last data of the defined logistic regression model is input is set to t-1, and this sample data is set to S t-1 , and the model parameters are Based on the current model parameters , the sample data S t at time t, the vehicle collision risk data y t at time t, and the sample loss function value L(y t , p(y t )) are calculated. Then, the sample S t is calculated according to the loss function value L(y t , p(y t )) about each vehicle collision risk prediction model parameter gradient value The loss function is the cross-entropy loss, and its formula is:

[0089] L(y t ,p(y t ))=-y t log(p(y t ))-(1-y t )log(1-p(y t ))

[0090] Where p(y t ) is S t The probability of a collision occurring in a given sample;

[0091] The formula for calculating the gradient value is:

[0092]

[0093] in, For parameters The gradient values, j = 1, 2, ..., n, where n is the number of independent variables;

[0094] S422: Calculate model parameters at time t based on gradient descent algorithm This completes one dynamic parameter calibration, and the calculation formula is:

[0095]

[0096] in, Here are the parameters of the logistic regression model at time t; α is the learning rate.

[0097] S423: Determine if the remaining data in the dataset have been traversed sequentially through the input: if yes, end the calibration; if no, increment t by 1 and return to S421.

[0098] S5: The ROC curve area is used to evaluate the prediction effect of each vehicle collision risk real-time prediction model, obtain the optimal sampling method and learning rate suitable for safety prediction of specific safety control road sections or nodes, and determine the final dynamic vehicle risk prediction model.

[0099] S51: Construct an ROC curve based on the True Positive Rate (TPR) and False Positive Rate (FPR) determined by comparing the prediction results of the real-time vehicle collision risk prediction model with the actual collision risk results calculated in S2; where the horizontal axis represents the True Positive Rate and the vertical axis represents the False Positive Rate. Based on the dynamically calibrated real-time vehicle collision risk prediction model at each time step, calculate the area under the ROC curve to obtain a series of AUC values ​​for the results of each real-time vehicle collision risk prediction model constructed from multiple discrete datasets;

[0100] S52: compare the AUC value of each vehicle collision risk real-time prediction model;

[0101] S521: determine whether the AUC value is higher than a preset AUC threshold value: if yes, it indicates that the prediction performance of the model is good; if no, it indicates that the prediction performance of the model is poor. In actual implementation, the preset AUC threshold value can be adjusted according to actual needs, and it is not limited. In this embodiment, the preset AUC threshold value is set to 0.9.

[0102] S522: input continuous multiple sample data into the vehicle collision risk real-time prediction model, and determine whether the AUC value change is less than a preset standard threshold value: if yes, it indicates that the vehicle collision risk real-time prediction model has good stability; if no, it indicates that the vehicle collision risk real-time prediction model has poor stability.

[0103] The interval sampling mode corresponding to the vehicle collision risk real-time prediction model with good prediction performance and stability is selected as the optimal interval sampling mode;

[0104] S53: sensitivity analysis is performed on the learning rate a in the online gradient descent algorithm based on the loss function value and the AUC value.

[0105] For sensitivity analysis, this embodiment uses a learning rate of 0.001 to 0.01 for prediction, and compares the prediction effect of the model based on the loss function value and the AUC value. When the loss function value decreases too slowly, the learning rate is increased; when the loss function value oscillates, the learning rate is decreased. The higher the AUC value, the better the learning rate. Based on this, the optimal learning rate a' is determined.

[0106] S54: combine the optimal interval sampling mode and the optimal learning rate to obtain a vehicle risk prediction model suitable for the road safety control section or node.

[0107] Embodiment 2

[0108] The embodiment provides a dynamic calibration vehicle collision risk real-time prediction model construction system, comprising:

[0109] The data acquisition and preprocessing module acquires vehicle trajectory data of the safety control section or node through a perception device, and extracts vehicle feature parameters and traffic flow feature parameters;

[0110] The collision risk module calculates the expansion distance collision time of each vehicle and the nearest preceding vehicle at each time based on the vehicle feature parameters and the traffic flow feature parameters, determines the collision risk of each vehicle and the nearest preceding vehicle, and extracts the collision risk influencing factors;

[0111] Data sampling and discretization module: interval sampling of vehicle feature parameters, traffic flow feature parameters and collision risk obtained from all vehicle trajectory data in different ways to obtain corresponding multiple discrete data sets;

[0112] Model construction module: taking the collision risk influencing factors in the discrete data set as input and the collision risk of each vehicle with the nearest preceding vehicle as output, constructing a vehicle collision risk prediction model based on the multiple discrete data sets obtained in step S3, and dynamically calibrating the model parameters to obtain multiple real-time vehicle collision risk prediction models;

[0113] Model parameter optimization module: using the area of the ROC curve to evaluate the prediction effect of each vehicle collision risk real-time prediction model to obtain the optimal sampling method and learning rate suitable for safety prediction of safety control sections or nodes, and determine the final vehicle collision risk real-time prediction model.

[0114] Example 3

[0115] The present embodiment provides a road risk monitoring and early warning platform, which deploys the steps of the method as described above.

[0116] Example 4

[0117] The present embodiment provides an electronic terminal: including a processor and a memory, the memory stores a computer program, the processor invokes the computer program to execute the steps of the method as described above.

[0118] Example 4

[0119] The present embodiment provides a readable storage medium: stores a computer program, the computer program is invoked by a processor to execute the steps of the method as described above.

[0120] It should be understood that in the embodiments of the present application, the processor can be a central processing unit (CPU), and the processor can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, and the like. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor. The memory can include a read-only memory and a random access memory, and provide instructions and data for the processor. A portion of the memory can also include a non-volatile random access memory. For example, the memory can also store device type information.

[0121] The readable storage medium is a computer readable storage medium, which can be an internal storage unit of the controller, such as a hard disk or a memory of the controller. The readable storage medium can also be an external storage device of the controller, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, and the like. Further, the readable storage medium can include both the internal storage unit and the external storage device of the controller. The readable storage medium is used to store the computer program and other programs and data required by the controller. The readable storage medium can also be used to temporarily store data that has been output or will be output.

[0122] Based on such understanding, the technical solutions of the present application, essentially or in the contribution to the prior art, or all or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned readable storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various media that can store program codes.

[0123] It can be understood that the same or similar parts in the above-mentioned embodiments can be mutually referenced, and the contents not described in detail in some embodiments can refer to the same or similar contents in other embodiments.

[0124] Although the embodiments of the present application have been shown and described above, it can be understood that the above-mentioned embodiments are exemplary and cannot be understood as limiting the present application, and the ordinary skilled in the art can make changes, modifications, replacements and variations to the above-mentioned embodiments within the scope of the present application.

Claims

1. A method for constructing a dynamic calibration vehicle collision risk real-time prediction model, characterized in that, Comprise: S1: Obtain vehicle trajectory data of a safety management section or node by a sensing device, and extract vehicle feature parameters and traffic flow feature parameters; S2: Based on the vehicle characteristic parameters and traffic flow characteristic parameters, the extended distance collision time of each vehicle and the nearest front vehicle at each time is calculated, the collision risk of each vehicle and the nearest front vehicle is judged, and the collision risk influencing factors are extracted; wherein the collision risk influencing factors include but are not limited to: vehicle length , vehicle width , front vehicle speed L v , rear vehicle speed F v , vehicle driving distance D, vehicle distance , vehicle type , average traffic density K; S3: Intervals of the vehicle feature parameters, traffic flow feature parameters and collision risk obtained from all vehicle trajectory data are sampled in multiple different ways to obtain corresponding multiple discrete data sets; S4: Collision risk influencing factors in the discrete data sets are taken as inputs, and the collision risk of each vehicle with the nearest preceding vehicle is taken as output, and vehicle collision risk prediction models are constructed based on the multiple discrete data sets obtained in step S3, and the parameters of each model are dynamically calibrated to obtain multiple real-time vehicle collision risk prediction models; S41: Based on the discrete data sets obtained in S3, a preset proportion of data in the data set is used to define a logistic regression model, i.e. to construct a vehicle collision risk prediction model, to predict the vehicle collision probability under the influence of multiple factors; S42: The remaining data in the data set is input into the vehicle collision risk prediction model in turn for dynamic calibration, and multiple real-time vehicle collision risk prediction models are obtained, wherein each discrete data set corresponds to a real-time vehicle collision risk prediction model; The specific process of dynamic calibration is: S421: set the time of inputting the last data of the defined logistic regression model to the time t-1, and input the sample data at this time, the model parameters are ; based on the current model parameters , the sample data at the time t , the vehicle collision risk data at the time t , the sample loss function value is calculated; according to the loss function value , the sample gradient value of each vehicle collision risk prediction model parameter ; wherein the loss function is cross-entropy loss, and its formula is: ； wherein is the probability that a collision occurs under the sample The gradient value calculation formula is: ； wherein is a gradient value with respect to a parameter , , is the number of arguments; S422: calculate based on gradient descent algorithm model parameters at the moment , that is, complete a parameter dynamic calibration, and the calculation formula is: ； wherein, is the logistic regression model parameter at time t; is the learning rate; S423: Determine whether the remaining data in the data set is input in turn: if yes, end calibration; if no, t+1 and return to S421; S5: The prediction effect of each real-time vehicle collision risk prediction model is evaluated by using the area of the ROC curve to obtain the optimal sampling method and learning rate suitable for safety prediction of the safety management section or node, and the final real-time vehicle collision risk prediction model is determined.

2. The dynamic calibration vehicle collision risk real-time prediction model construction method according to claim 1, characterized in that, The specific process of S1 is: S11: Collect vehicle trajectory sample data of the safety management section or node by the sensing device arranged at the safety management section or node, and extract the complete motion trajectory of the vehicle, and perform data preprocessing; S12: Extract the vehicle trajectory data of each vehicle after data preprocessing, including the position coordinates of the vehicle at each time, vehicle ID, and time sequence identifier ; S13: Based on the vehicle trajectory data of S12, further processing obtains the vehicle characteristic parameters and traffic flow characteristic parameters of all vehicles; wherein the vehicle characteristic parameters include but are not limited to vehicle length , vehicle width , vehicle speed , vehicle speed direction , vehicle speed change rate , vehicle driving distance , inter-vehicle distance , vehicle type ; Traffic flow characteristic parameters include, but are not limited to, average traffic flow , average traffic density , headway . 3.The method of claim 1, wherein, The specific process of S2 is: S21: Establish coordinate axes based on vehicle trajectory data, match each vehicle at time t with the nearest preceding vehicle, and calculate the instantaneous traffic conflict ETTC between the matched vehicles, i.e. the extended distance collision time; The calculation formula of ETTC is as follows: ； where, and are the lengths of the vehicles and respectively; D ij is the distance between the center points of the two vehicles; is the distance between the closest points of the two vehicles; and are the two-dimensional coordinates and velocity vectors of the center points of the vehicles; S22: Determine the collision risk of all matched vehicles according to the determination condition of the calculated instantaneous traffic conflict; Wherein, the determination condition is: When ETTC is less than or equal to a preset threshold ETTC', the collision risk of the matched vehicle is collision risk; When ETTC is greater than the preset threshold ETTC', the collision risk of the matched vehicle is no collision risk.

4. The dynamic calibration vehicle crash risk real-time prediction model construction method according to claim 1, characterized in that, The specific process of S3 is: S31: Set n different sampling time intervals, and divide the continuous data at equal intervals; S32: On the basis of equal interval division data, select sampling points for each interval data; Wherein, the sampling point selection includes two kinds, one is to select the data with time sequence number 1 in each sampling interval as sampling data; The second is to randomly sample the data with any time sequence number in each sampling interval as sampling data; S33: K different sampling methods are obtained based on n different sampling time intervals and 2 sampling points, wherein The vehicle feature parameters, traffic flow feature parameters and collision risk data obtained from all vehicle trajectory data are sampled to obtain K discrete data sets.

5. The dynamic calibration vehicle crash risk real-time prediction model construction method according to claim 1, characterized in that, The specific process of S5 is: S51: An ROC curve is constructed based on the true positive rate and false positive rate determined by comparing the prediction result of the vehicle collision risk real-time prediction model with the actual result of the collision risk calculated in S2; Based on the vehicle collision risk real-time prediction model dynamically calibrated at each moment, the area under the ROC curve is calculated to obtain a series of AUC values of the results of each vehicle collision risk real-time prediction model constructed by multiple discrete data sets; S52: Compare the AUC value corresponding to each vehicle collision risk real-time prediction model in terms of numerical size and stability: S521: Determine whether the AUC value is higher than the preset AUC threshold: if yes, it indicates that the prediction performance of the model is good; if no, it indicates that the prediction performance of the model is poor; S522: Input continuous multiple sample data into the vehicle collision risk real-time prediction model to determine whether the change of the AUC value is less than the preset standard threshold: if yes, it indicates that the stability of the vehicle collision risk real-time prediction model is good; if no, it indicates that the stability of the vehicle collision risk real-time prediction model is poor; The interval sampling method corresponding to the vehicle collision risk real-time prediction model with good prediction performance and stability is selected as the optimal interval sampling method; S53: According to the loss function value and the AUC value, the learning rate in the online gradient descent algorithm is analyzed for sensitivity to determine the optimal learning rate; S54: Combine the optimal interval sampling method and the optimal learning rate to obtain a vehicle risk prediction model suitable for road safety management and control sections or nodes.

6. A road risk monitoring and early warning platform, characterized in that, The platform deploys the steps of the method of any one of claims 1-5.

7. An electronic terminal, characterized by: A processor and a memory are included, the memory stores a computer program, and the processor invokes the computer program to execute the steps of the method of any one of claims 1-5.

8. A readable storage medium characterized by: A computer program is stored, and when invoked by a processor, the computer program executes the steps of the method of any one of claims 1-5.

Images