A short-term warning method for collision risk in work areas based on traffic flow characteristics

Abstract

The present application belongs to the technical field of traffic safety management, and discloses a method for short-term collision risk warning of an operation area based on traffic flow characteristics, comprising the following steps: step 1: identifying vehicles on an upstream section of the operation area, and collecting historical traffic flow characteristics of the upstream section of the operation area; step 2: identifying vehicles on a downstream dangerous section of the operation area, and calculating the historical collision risk of the downstream dangerous section using a collision risk potential field model of the operation area; step 3: determining and verifying the time difference corresponding to the traffic flow characteristics of each upstream section and the collision risk of the downstream dangerous section using a time series correlation analysis, and using the time difference to perform time series alignment on the traffic flow characteristics of the upstream section and the collision risk of the downstream dangerous section; step 4: using the traffic flow characteristics of the upstream section after time series alignment and the collision risk of the downstream dangerous section to train a short-term collision risk warning model; and step 5: using the short-term collision risk warning model to predict the collision risk of vehicles on the road to be evaluated.

Description

Technical Field

[0001] This application belongs to the field of traffic safety management technology, and specifically relates to a short-term warning method for collision risk in an operation area based on traffic flow characteristics. It aims to predict the collision risk in the operation area in advance by real-time monitoring and analysis of traffic flow data to improve road safety. Background Art

[0002] Highway work zones are special areas established to ensure vehicle safety during construction work. Their primary applications involve routine maintenance and renovation and expansion projects. In China, as many highway sections enter their mid-to-late service life and traffic volume continues to grow, the demand for routine maintenance and renovation and expansion projects continues to rise. This increase not only increases the frequency of routine maintenance but also requires renovation and expansion projects to increase the highway's carrying capacity. To ensure smooth traffic flow, highway renovation and expansion projects often adopt a "construction-while-opening" model. This results in the establishment of work zones, which alters the existing traffic environment and significantly increases the risk of vehicle collisions within these sections. With the increase in highway traffic volume, safety issues in work zones have become increasingly serious. Traditional safety measures that rely on static signs and warnings are no longer able to adapt to the rapidly changing traffic environment. Drivers need time to adjust their vehicle speed and trajectory to avoid danger in emergencies, making static safety measures inadequate for risk management in work zones.

[0003] Predicting collision risks in work zones is a crucial foundation for proactive safety control. Existing research primarily focuses on static risk assessment and lacks effective dynamic early warning mechanisms. Therefore, leveraging real-time traffic flow data for short-term collision risk warnings has become a key approach to improving safety in work zones.

[0004] Based on this, a short-term warning method for collision risk in work areas based on traffic flow characteristics is proposed to solve the above problems. Summary of the Invention

[0005] The purpose of this application is to solve the problems of the existing technology and provide a short-term warning method for collision risk in the work area based on traffic flow characteristics. The method can monitor the traffic flow characteristics of the upstream section in real time, predict the collision risk, build an effective short-term warning mechanism, and release warning information in a timely manner to reduce the occurrence rate of collision accidents in the dangerous section downstream of the work area.

[0006] In order to solve the technical problem, the technical solution of this application is: a short-term warning method for collision risk in a work area based on traffic flow characteristics, comprising the following steps:

[0007] Step 1: Identify vehicles on the upstream section of the work area and collect historical traffic flow characteristics of the upstream section of the work area;

[0008] Step 2: Identify vehicles on the dangerous road section downstream of the work area and calculate the historical collision risk of the downstream dangerous road section using the work area collision risk potential field model;

[0009] Step 3: Use temporal correlation analysis to determine and verify the time difference between the traffic flow characteristics of each upstream road segment and the collision risk of the downstream dangerous road segment. Use the time difference to align the traffic flow characteristics of the upstream road segment with the collision risk of the downstream dangerous road segment.

[0010] Step 4: Use the traffic flow characteristics of the upstream section after time series alignment and the collision risk of the downstream dangerous section to train the collision risk short-term warning model;

[0011] Step 5: Use the collision risk short-term warning model to predict the collision risk of the road vehicle to be assessed.

[0012] Preferably, the step 1 is specifically as follows: deploying traffic flow monitoring equipment on the upstream section of the work area, identifying vehicles on the upstream section of the work area, collecting historical traffic flow data, the traffic flow data including traffic volume, maximum speed, average speed, minimum headway and average headway, using a sliding time window to divide the work area scene, and extracting historical traffic flow characteristics under each scene from the time series data; the work area includes an upstream section, an intermediate section and a downstream dangerous section, the upstream section is the starting point of the warning area, and is used to warn downstream dangers.

[0013] Preferably, step 2 is specifically as follows: deploying traffic flow monitoring equipment on the dangerous section downstream of the work area, identifying vehicles on the dangerous section downstream of the work area, collecting historical vehicle trajectory data of the dangerous section downstream of the work area, the vehicle trajectory data including speed, position, vehicle type and road slope, and calculating the historical collision risk of the dangerous section downstream based on the vehicle trajectory data using the collision risk potential field model of the work area.

[0014] Preferably, the collision risk potential field model of the operation area is specifically as follows: taking the target vehicle as the field source, by constructing the risk field force F , collision potential field strength E and risk loading Q , converting the spatial position of the target vehicle and the interaction object into an equivalent time distance , combining the vehicle type and road slope to form a vehicle-road combination function , and then calculate the risk load based on the relative speed Q Finally, the collision risk of the downstream dangerous section is quantified through the extreme field force formula, realizing the dynamic assessment of the collision risk of road vehicles and the positioning of dangerous areas.

[0015] Preferably, the collision risk potential field model of the working area is:

[0016] ;

[0017] ;

[0018] ;

[0019] ;

[0020] ;

[0021] ;

[0022] ;

[0023] ;

[0024] ;

[0025] ;

[0026] ;

[0027] ;

[0028] ;

[0029] Where:

[0030] F V For extreme field forces;

[0031] is the collision potential field strength;

[0032] The target vehicle and the interaction object i The risk field forces between them;

[0033] For interactive objects i risk loading;

[0034] is the position of the virtual collision point of the interaction object in the target vehicle coordinate system, unit: (m,m);

[0035] is the vehicle-road combination function;

[0036] is the equivalent time distance, unit: s -1 ;

[0037] The coordinates of the interactive object in the road Frenet coordinate system, unit: (m,m);

[0038] is the coordinate of the target vehicle in the Frenet coordinate system of the road, unit: (m,m);

[0039] The vehicle's direction of travel and s The clockwise angle of the axis, unit: degree;

[0040] is an equivalent parameter;

[0041] M is the vehicle attribute parameter;

[0042] I is the road attribute parameter;

[0043] v is the target vehicle speed, unit: m / s;

[0044] v 0 is the speed of the interactive object, unit: m / s;

[0045] is the equivalent angle between the interaction object and the target vehicle, unit: degree;

[0046] SSD is the stopping sight distance, unit: m;

[0047] G is the longitudinal slope rate.

[0048] Preferably, step 3 is specifically as follows: the time series correlation analysis includes cross-correlation analysis, Pearson correlation coefficient analysis and Granger causality test, and the cross-correlation analysis is used to determine a reasonable time difference, that is, the length of time that can be used for early warning; then, based on the time difference, the Pearson correlation coefficient analysis is used to quantitatively evaluate the correlation between the traffic flow characteristics of the upstream section and the collision risk of the downstream dangerous section; finally, the Granger causality test is used to verify the correlation to test whether the traffic flow characteristics of the upstream section statistically significantly affect the change in the collision risk of the downstream dangerous section. If the test result shows that there is a causal relationship, it means that the time difference determined by the cross-correlation analysis is reasonable, and based on this, the traffic flow characteristics of the upstream section and the collision risk of the downstream dangerous section are time-series aligned to provide reliable time-series input for the subsequent construction of a short-term warning model for collision risk.

[0049] Preferably, the collision risk short-term warning model in step 4 includes 9 layers, namely the first Conv1D layer, the second Conv1D layer, the MaxPooling1D pooling layer, the bidirectional LSTM layer, the SeqSelfAttention attention mechanism layer, the Flatten flattening layer, the Dense fully connected layer, the Dropout layer and the Dense output layer; the first Conv1D layer and the second Conv1D layer are 1D CNN convolution layers, which are used to extract the traffic flow characteristics of the upstream section, and the dimension is reduced by the MaxPooling1D pooling layer to input the bidirectional LSTM layer to perform context-related learning on the extracted upstream section traffic flow characteristics, and then the SeqSelfAttention attention mechanism layer is used to enhance the model's attention to important features, followed by the Flatten flattening layer and the Dense fully connected layer to flatten the multidimensional information into a one-dimensional nonlinear transformation; at the same time, the Dropout layer is used to randomly close neurons to prevent overfitting; finally, the Dense output layer is applied to output the binary classification results.

[0050] Preferably, the collision risk short-term warning model determines 6 parameters as model tuning objects: filters1: the number of filters of the first Conv1D layer; kernel_size1: the convolution kernel size of the first Conv1D layer; filters2: the number of filters of the second Conv1D layer; kernel_size2: the convolution kernel size of the second Conv1D layer; lstm_units: the number of units of the bidirectional LSTM layer; dropout_rate: the dropout ratio of the Dropout layer.

[0051] Compared with the prior art, the advantages of this application are:

[0052] (1) This application starts from the goal of short-term warning of collision risk in the operation area. First, the temporal correlation between the traffic flow characteristics of the upstream section of the operation area and the collision risk of the downstream dangerous section is analyzed. By comprehensively utilizing cross-correlation analysis, Pearson correlation coefficient analysis and Granger causality test, it is determined that the traffic flow characteristics of the upstream section of the operation area have a significant impact on the collision risk of the downstream dangerous section after a time period of Δt. On this basis, a short-term warning model for collision risk is constructed using the traffic flow characteristics of the upstream section and the collision risk of the downstream dangerous section. A short-term warning method for collision risk in the operation area based on traffic flow characteristics is proposed. That is, the collision risk of the downstream dangerous section after a time period of Δt is predicted by the traffic flow characteristics of the upstream section, and an early warning is issued to realize an effective dynamic warning mechanism.

[0053] (2) The work area collision risk potential field model disclosed in this application takes the target vehicle as the potential field source, combines the wedge potential field distribution and the Yukawa potential field function, and uses the equivalent time distance Effectively solve the problem of potential field anisotropy, making the quantification of the impact of distance on risk more realistic; comprehensively consider the key influencing factors and combine the vehicle-road combination function Incorporating vehicle attributes (such as vehicle type) and road attributes (such as slope), the risk load is calculated based on the relative speed between the interacting object and the target vehicle. Q At the same time, using extreme field force as a quantitative indicator, it focuses on the front and side interaction objects and collision possibilities of the target vehicle, eliminating interference from rear objects, significantly improving the accuracy and pertinence of quantifying collision risks in the work area. It can also adapt to multiple scenarios such as straight driving, turning, different vehicle models, and slopes, and comprehensively integrates multi-dimensional parameters with actual scene characteristics to achieve dynamic assessment of road vehicle collision risks and locate dangerous areas.

[0054] (3) This application proposes an improved short-term collision risk warning model, combining three models to construct 1D CNN + LSTM + Attention. Based on the respective advantages of these three models, 1D CNN's local feature extraction, LSTM's time series data processing capabilities, and Attention's contextual focus, the improved short-term collision risk warning model constructed in this application performs better than all other model combinations on the test set, demonstrating the advantages of this comprehensive approach in generalization ability and comprehensive performance. It can more comprehensively capture and utilize complex patterns and relationships in the data, thereby achieving a better balance in various performance indicators;

[0055] (4) This application improves the traffic safety level in the operation area and provides effective risk management tools for traffic management departments, enabling them to formulate scientific and reasonable traffic management strategies based on real-time traffic flow characteristics and collision risk prediction results. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 A flowchart of a method for short-term warning of collision risk in a work area based on traffic flow characteristics provided in Example 1 of the present application;

[0057] Figure 2 The equivalent time distance provided in Example 1 of this application Calculation diagram;

[0058] Figure 3 A time series change diagram of characteristic variables of small vehicles and large vehicles in each lane of the upstream road section provided in Example 1 of the present application;

[0059] Figure 4 This is a comparison chart of the prediction results and actual values ​​of the short-term warning model for collision risk in the work area provided in Example 1 of the present application. DETAILED DESCRIPTION

[0060] The present application is described in detail below with reference to the accompanying drawings and specific examples, but the present application is not limited to these examples. This application covers any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this application. To provide a thorough understanding of this application, specific details are described in detail in the following examples of this application, but those skilled in the art can fully understand this application without these detailed descriptions.

[0061] This application discloses a method for short-term early warning of collision risk in a work area based on traffic flow characteristics, comprising the following steps:

[0062] Step 1: Identify vehicles on the upstream section of the work area and collect historical traffic flow characteristics of the upstream section of the work area;

[0063] Step 2: Identify vehicles on the dangerous road section downstream of the work area and calculate the historical collision risk of the downstream dangerous road section using the work area collision risk potential field model;

[0064] Step 3: Use temporal correlation analysis to determine and verify the time difference between the traffic flow characteristics of each upstream road segment and the collision risk of the downstream dangerous road segment. Use the time difference to align the traffic flow characteristics of the upstream road segment with the collision risk of the downstream dangerous road segment.

[0065] Step 4: Use the traffic flow characteristics of the upstream section after time series alignment and the collision risk of the downstream dangerous section to train the collision risk short-term warning model;

[0066] Step 5: Use the collision risk short-term warning model to predict the collision risk of the road vehicle to be assessed.

[0067] Preferably, the step 1 is specifically as follows: deploying traffic flow monitoring equipment on the upstream section of the work area, identifying vehicles on the upstream section of the work area, collecting historical traffic flow data, the traffic flow data including traffic volume, maximum speed, average speed, minimum headway and average headway, using a sliding time window to divide the work area scene, and extracting historical traffic flow characteristics under each scene from the time series data; the work area includes an upstream section, an intermediate section and a downstream dangerous section, the upstream section is the starting point of the warning area, and is used to warn downstream dangers.

[0068] Preferably, step 2 is specifically as follows: deploying traffic flow monitoring equipment on the dangerous section downstream of the work area, identifying vehicles on the dangerous section downstream of the work area, collecting historical vehicle trajectory data of the dangerous section downstream of the work area, the vehicle trajectory data including speed, position, vehicle type and road slope, and calculating the historical collision risk of the dangerous section downstream based on the vehicle trajectory data using the collision risk potential field model of the work area.

[0069] Preferably, the collision risk potential field model of the operation area is specifically as follows: taking the target vehicle as the field source, by constructing the risk field forceF , collision potential field strength E and risk loading Q , converting the spatial position of the target vehicle and the interaction object into an equivalent time distance , combining the vehicle type and road slope to form a vehicle-road combination function , and then calculate the risk load based on the relative speed Q Finally, the collision risk of the downstream dangerous section is quantified through the extreme field force formula, realizing the dynamic assessment of the collision risk of road vehicles and the positioning of dangerous areas.

[0070] Preferably, the collision risk potential field model of the working area is:

[0071] ;

[0072] ;

[0073] ;

[0074] ;

[0075] ;

[0076] ;

[0077] ;

[0078] ;

[0079] ;

[0080] ;

[0081] ;

[0082] ;

[0083] ;

[0084] Where:

[0085] F V For extreme field forces;

[0086] is the collision potential field strength;

[0087] The target vehicle and the interaction object i The risk field forces between them;

[0088] For interactive objects i risk loading;

[0089] is the position of the virtual collision point of the interaction object in the target vehicle coordinate system, unit: (m,m);

[0090] is the vehicle-road combination function;

[0091] is the equivalent time distance, unit: s -1 ;

[0092] The coordinates of the interactive object in the road Frenet coordinate system, unit: (m,m);

[0093] is the coordinate of the target vehicle in the Frenet coordinate system of the road, unit: (m,m);

[0094] The vehicle's direction of travel and s The clockwise angle of the axis, unit: degree;

[0095] is an equivalent parameter;

[0096] M is the vehicle attribute parameter;

[0097] I is the road attribute parameter;

[0098] v is the target vehicle speed, unit: m / s;

[0099] v 0 is the speed of the interactive object, unit: m / s;

[0100] is the equivalent angle between the interaction object and the target vehicle, unit: degree;

[0101] SSD is the stopping sight distance, unit: m;

[0102] G is the longitudinal slope rate.

[0103] F V 、 、 Fi 、 QiThe parameters are dimensionless field strength index, relative value of potential field strength, dimensionless force between interaction object i and target vehicle, and abstract concepts like “charge”, which have no specific physical units and are dimensionless characterization parameters. 、 、 M, I, G There are no specific physical units for these parameters.

[0104] Preferably, step 3 is specifically as follows: the time series correlation analysis includes cross-correlation analysis, Pearson correlation coefficient analysis and Granger causality test, and the cross-correlation analysis is used to determine a reasonable time difference, that is, the length of time that can be used for early warning; then, based on the time difference, the Pearson correlation coefficient analysis is used to quantitatively evaluate the correlation between the traffic flow characteristics of the upstream section and the collision risk of the downstream dangerous section; finally, the Granger causality test is used to verify the correlation to test whether the traffic flow characteristics of the upstream section statistically significantly affect the change in the collision risk of the downstream dangerous section. If the test result shows that there is a causal relationship, it means that the time difference determined by the cross-correlation analysis is reasonable, and based on this, the traffic flow characteristics of the upstream section and the collision risk of the downstream dangerous section are time-series aligned to provide reliable time-series input for the subsequent construction of a short-term warning model for collision risk.

[0105] Preferably, the collision risk short-term warning model in step 4 includes 9 layers, namely the first Conv1D layer, the second Conv1D layer, the MaxPooling1D pooling layer, the bidirectional LSTM layer, the SeqSelfAttention attention mechanism layer, the Flatten flattening layer, the Dense fully connected layer, the Dropout layer and the Dense output layer; the first Conv1D layer and the second Conv1D layer are 1D CNN convolution layers, which are used to extract the traffic flow characteristics of the upstream section, and the dimension is reduced by the MaxPooling1D pooling layer to input the bidirectional LSTM layer to perform context-related learning on the extracted upstream section traffic flow characteristics, and then the SeqSelfAttention attention mechanism layer is used to enhance the model's attention to important features, followed by the Flatten flattening layer and the Dense fully connected layer to flatten the multidimensional information into a one-dimensional nonlinear transformation; at the same time, the Dropout layer is used to randomly close neurons to prevent overfitting; finally, the Dense output layer is applied to output the binary classification results.

[0106] Preferably, the collision risk short-term warning model determines 6 parameters as model tuning objects: filters1: the number of filters of the first Conv1D layer; kernel_size1: the convolution kernel size of the first Conv1D layer; filters2: the number of filters of the second Conv1D layer; kernel_size2: the convolution kernel size of the second Conv1D layer; lstm_units: the number of units of the bidirectional LSTM layer; dropout_rate: the dropout ratio of the Dropout layer.

[0107] The collision risk is divided into high risk and low risk.

[0108] Example 1

[0109] like Figure 1 , which is a flow chart of a method for short-term early warning of collision risk in an operation area based on traffic flow characteristics of the present application.

[0110] This application provides a method for short-term warning of collision risk in a work area based on traffic flow characteristics, which includes the following steps:

[0111] Step S1: Identify vehicles on the upstream section of the work area and collect historical traffic flow characteristics of the upstream section of the work area.

[0112] Specifically, traffic flow monitoring equipment is deployed on the upstream section of the work area to identify vehicles on the upstream section of the work area and collect historical traffic flow data, including traffic volume, maximum speed, average speed, minimum headway and average headway. The work area scenes are divided using a sliding time window, and the historical traffic flow characteristics of each scene are extracted from the time series data; providing original data support for subsequent feature extraction and risk prediction.

[0113] Step S2: Identify vehicles on the dangerous road section downstream of the work area, and calculate the historical collision risk of the dangerous road section downstream using the work area collision risk potential field model.

[0114] Traffic flow monitoring equipment is deployed on the dangerous sections downstream of the work area to identify vehicles on the dangerous sections downstream of the work area and collect historical vehicle trajectory data on the sections downstream of the work area. The vehicle trajectory data includes speed, position, vehicle type and road slope. Based on the vehicle trajectory data, the collision risk potential field model of the work area is used to calculate the historical collision risk of the dangerous sections downstream.

[0115] Specifically, a collision risk potential field model for the operation area is adopted. This model takes the target vehicle as the field source. By constructing the risk field force, field strength and risk load, the spatial position of the target vehicle and the interaction object is converted into an equivalent time distance. The vehicle type and road slope are combined to form a vehicle-road combination function, and the risk load of the interaction object is calculated based on the relative speed. Finally, the extreme field force formula is used to comprehensively analyze various parameters to quantify the collision risk of downstream dangerous sections, thereby realizing dynamic assessment of road vehicle collision risks and positioning of dangerous areas.

[0116] The collision risk potential field model of the work area proposed in this application mainly includes the risk field force between the target vehicle and the interaction object. F , the collision potential field strength formed by the target vehicle E , the risk load of the interaction object Q There are three parts. Among them, risk field force F Used to quantify the collision risk between the target vehicle and the interactive object, with the direction pointing from the virtual collision point to the target vehicle. E and risk charge Q Related to risk field forces F Different, extreme field force F V It is used to quantify the collision risk of the target vehicle. The calculation method is shown in Formula 1:

[0117] (1)

[0118] Where, F V For extreme field forces; is the collision potential field strength; Represents the target vehicle and the interaction object i The field forces between For interactive objects i risk load.

[0119] Taking the target vehicle as the field source and conceptualizing the target vehicle as a rectangle, a field strength model with time as the distance indicator is proposed, as shown in Equation (2). The field strength model is composed of the equivalent time distance and vehicle-road combination function It consists of two parts.

[0120] (2)

[0121] The equivalent time distance in the collision risk potential field model of the operation area , vehicle-road combination function η and the risk load of the interacting object Q The calculation methods of the three main parameters are as follows:

[0122] (1) Equivalent time distance

[0123] like Figure 2 As shown in (a), the time required for the target vehicle to reach the designated location from the starting point is called time distance, which can be directly used to measure the collision risk. Using time distance as a quantitative indicator of collision risk, assuming that the vehicle and obstacle A (distance is x , the speed is v ) and obstacle B (distance is y , the speed is v * ) have equal collision risks, and the time required for the vehicle to reach the PF is is the real time distance, the time distance between the vehicle and PL This is the equivalent time distance.

[0124] Based on the speed of the target vehicle v , we can get the equivalent time distance as Assume that there is a virtual velocity along the y-axis of the vehicle v * = v , making , is the equivalent parameter.

[0125] like Figure 2 As shown in (b), in the vehicle coordinate system with the target vehicle as the coordinate center, the driving direction as the x-axis, and the right side perpendicular to the driving direction as the y-axis, any point in the surrounding space All satisfy formula (3), the equivalent time distance All of them can be expressed using formula (4). In order to facilitate application to the road Frenet coordinate system, coordinate transformation is performed according to formula (5), as shown below:

[0126] (3)

[0127] (4)

[0128] (5)

[0129] Where, is the position of the virtual collision point of the interactive object in the target vehicle coordinate system. When considering the actual size of the vehicle, the length and width of the vehicle should be subtracted accordingly. is the equivalent time distance; is the coordinate of the interactive object in the road Frenet coordinate system; is the coordinate of the target vehicle in the Frenet coordinate system of the road; The vehicle's direction of travel and sThe clockwise angle of the axis; is an equivalent parameter; v is the target vehicle speed.

[0130] Equivalent parameters Directly affects the spatial distribution of the potential field, which depends on the degree of influence of the lateral distance and longitudinal distance on the collision risk. and vehicle speed v Direct linear correlation, the relationship can be expressed as Combined with the research on the threshold of accident time (TA), the longitudinal safety distance and vehicle speed v The relationship between can be expressed as Therefore, the equivalent parameter It can be expressed as follows:

[0131] (6)

[0132] (2) Vehicle-road combination function η

[0133] Braking performance is one of the important factors affecting collision risk and is closely related to vehicle and road attributes. This application evaluates different attributes (vehicle attribute parameters M and road attribute parameters I ) on the impact of parking sight distance, and establish a vehicle-road combination function To determine the field strength of the collision risk potential field in the operation area, as shown in formula (7). When the car is driving on a flat road, The standard value of is 1.

[0134] (7)

[0135] Taking small cars as the standard, the model parameters of small cars are , large vehicle model attribute parameters It is the ratio of the parking sight distance of large trucks to that of small passenger cars. Further regression analysis of the parameters yields formula (8), and the coefficient of determination is R 2 The average error (A) is 0.9952, and the root mean square error (RMSE) is 0.0029.

[0136] (8)

[0137] Road slope is one of the important factors affecting vehicle braking. Based on the different levels of SSD calculation method provided by the American Association of State Highway and Transportation Officials, assuming the driver's reaction time is 2.5 seconds and the vehicle braking deceleration is 3.4 m / s 2 , the SSD calculation formula is shown in formula (9). Road attribute parameters I For different road slopes G SSD and flat road ( G =0), as shown in formula (10):

[0138] (9)

[0139] (10)

[0140] (3) Risk load of the interaction object Q

[0141] Relative speed is an important parameter of traditional SSM such as TTC and DRAC because it can better measure the collision risk. In this regard, this application proposes a method to calculate the risk load of the interactive object based on the relative speed between the interactive object and the target vehicle. Q The method is shown in Equations (11) and (12). This application focuses on the possibility of collision, so the properties of the interactive object only include its spatial position and velocity. The speed of the guardrail and traffic cone are both zero. The speed of the surrounding vehicles is based on the scalar value of their actual speed, while the speed of the guardrail and traffic cone are both zero. At the same time, the position of the virtual collision point is used as the spatial position of the interactive object. In addition, the greater the positive speed difference between the target vehicle and the interactive object, the higher the possibility of collision. Q When the speed difference between the vehicle and the interaction object is 0 Q =1, when the speed difference is 5.56m / s (20km / h), Q =2.

[0142] (11)

[0143] (12)

[0144] Where, v 0 The speed of the interactive object; The equivalent angle between the interacting object and the target vehicle; is the position of the virtual collision point of the interactive object in the target vehicle coordinate system, and is calculated The actual length and width of the vehicle are not taken into account.

[0145] After determining the equivalent time distance , vehicle-road combination function , the risk load of the interaction object Q Based on the three main parameters, the collision risk potential field model of the work area can be expressed as Equation (13) by combining Equations (1), (2), (4), (6) to (12).

[0146] (13)

[0147] Where, is the position of the virtual collision point of the interactive object in the target vehicle coordinate system; is the equivalent time distance; is an equivalent parameter; M is the vehicle attribute parameter; I is the road attribute parameter; v is the target vehicle speed; v 0 The speed of the interactive object; The equivalent angle between the interacting object and the target vehicle.

[0148] Step S3: Using time series correlation analysis, determine and verify the time difference between the traffic flow characteristics of each upstream section and the collision risk of the downstream dangerous section, and use the time difference to perform time series alignment on the traffic flow characteristics of the upstream section and the collision risk of the downstream dangerous section.

[0149] Specifically, this application aims to predict collision risks in downstream hazardous sections based on traffic flow characteristics in the upstream section of the work area (traffic flow characteristics at the starting section of the warning area). However, the complexity of the correlation between traffic flow characteristics and collision risk is unclear. Furthermore, because traffic flow characteristics are relatively stable, their core characteristic parameters (such as flow rate and speed) generally do not undergo drastic time-varying changes over short periods of time. Therefore, three methods are used: cross-correlation analysis (cross-correlation coefficient), correlation analysis (Pearson correlation coefficient analysis), and causal analysis (Granger causality test) to analyze the time series relationship between traffic flow characteristics in the upstream section and collision risks in the downstream hazardous section. Specifically, ① cross-correlation analysis aims to determine the appropriate hysteresis time (time difference) length, that is, the length of time that can provide early warning; ② Pearson correlation coefficient analysis is used to quantitatively assess the correlation between traffic flow characteristics in the upstream section and collision risks in the downstream hazardous section; and ③ causal analysis is used to determine whether the traffic flow characteristics in the upstream section can be used to predict collision risks in the downstream hazardous section.

[0150] The above three methods are used to test the hypothesis. Specifically, first, the vehicles on the upstream section pass through Δ tMost of the time, they arrive at the downstream dangerous section; according to the time hysteresis analysis, the hysteresis time between the traffic flow of the upstream section and the collision risk of the downstream dangerous section is significantly affected by the vehicle type. For example, in a case in the operation area, the hysteresis time of small and medium-sized vehicles is about 55 seconds, and the hysteresis time of large vehicles is about 75 seconds; when predicting the collision risk, the hysteresis time between vehicle types should be distinguished. Secondly, before and after the vehicles passing the upstream section arrive at the downstream dangerous section, the evolution of the interaction state between the vehicles has a certain regularity; according to the time series correlation analysis, there is a significant correlation between the traffic flow characteristics of the upstream section and the collision risk of the downstream dangerous section, and the determined hysteresis time is confirmed. Finally, according to the traffic flow characteristics of the upstream section of the operation area, it is possible to achieve Δ in advance t The time length provides a short-term warning of the collision risk of the downstream dangerous section; according to the Granger causality test, it is found that the traffic flow characteristics of the upstream are one of the main reasons for the change of the collision risk of the downstream dangerous section, which means that it is highly feasible to use the time series correlation analysis of the traffic flow characteristics of the upstream to predict the collision risk of the downstream dangerous section.

[0151] Step S4: using the traffic flow characteristics of the upstream road section after time series alignment and the collision risk of the downstream dangerous road section to train a short-term collision risk warning model.

[0152] Specifically, based on a real work area case, 1470 groups of samples were obtained by collecting about 2 hours of data. The hysteresis time of small and medium-sized vehicles and large vehicles was analyzed to be 55 seconds (1375 frames) and 75 seconds (1875 frames), respectively, and relevant feature variables were extracted. The data was processed using a 30-second window length and a 5-second window step size, and a short-term warning model was constructed. The training set and test set were divided into a ratio of 9:1. The first 90% of the data was used as the training set, and the last 10% of the data was used as the test set to train the collision risk short-term warning model. In order to better achieve short-term warning of collision risks in work areas, this application compares the performance of several different deep learning models in collision risk prediction, including one-dimensional convolutional neural networks (1D CNN), long short-term memory networks (LSTM), attention mechanisms (Attention) and their combinations. The performance evaluation results of the developed prediction models are shown in Table 1. The following points can be found:

[0153] 1) For short-term collision risk warning models built using a single algorithm, 1D CNN, LSTM, or Attention: First, 1D CNN demonstrated high recall on both the training and test sets, but low precision and F1 score. This means that the 1D CNN model produced a high number of false positives, incorrectly classifying high-risk samples as low-risk. Second, compared to 1D CNN, LSTM demonstrated balanced performance on both the training and test sets, with precision and recall closer together. This indicates that the LSTM's false positive problem is less severe than that of 1D CNN. Finally, compared to 1D CNN and LSTM, the Attention mechanism achieved similar performance on the test and training sets, demonstrating better generalization, but its overall prediction performance was lower than that of 1D CNN and LSTM.

[0154] 2) Regarding the short-term collision risk warning model that adds a 1D CNN to LSTM and Attention: First, we analyze the CL (1D CNN + LSTM) and CA (1D CNN + Attention) models, rather than the LC and AC models, because the 1D CNN's feature extraction capabilities are generally considered fundamental and critical. Therefore, the 1D CNN should be directly connected to the input layer to extract features, and the LSTM and Attention layers should be used to further explore the impact of these features. Second, both the CL and CA models demonstrate high and similar Precision and Recall, demonstrating that the models are able to balance the prediction accuracy of both risk levels. Finally, compared to using only LSTM or Attention, both the CL and CA combined models show significant performance improvements. This demonstrates that adding a 1D CNN to LSTM and Attention can leverage the 1D CNN's local feature extraction capabilities in processing time series data, thereby enhancing the model's understanding of time series data and improving overall performance.

[0155] 3) Regarding the short-term collision risk warning model combining LSTM and Attention: On the one hand, the performance of both AL and LA was less than ideal, with AL in particular showing lower overall performance than models using only LSTM or Attention. This may be because the attention mechanism is applied before the LSTM layer, resulting in the Attention layer making decisions without the context of the time series, which may limit the model's ability to understand and process time series data. On the other hand, LA performed slightly better than AL, possibly because the LSTM, as the first layer, more effectively processed the sequence data, providing more coherent context for the Attention mechanism.

[0156] 4) Combining three models to build a 1D CNN + LSTM + Attention model: Based on the respective strengths of these three models—1D CNN for local feature extraction, LSTM for time series data processing, and Attention for contextual focus—this application constructed the CLA model, which outperformed all other model combinations on the test set, demonstrating the advantages of this combined approach in generalization and overall performance. This demonstrates that combining these three techniques can more comprehensively capture and utilize complex patterns and relationships in the data, achieving a better balance across various performance metrics.

[0157] Table 1 Prediction model performance evaluation results

[0158]

[0159] S5. Use the collision risk short-term warning model to predict the collision risk of the road vehicle to be assessed.

[0160] Based on the model comparison results in step S4, this application chose to develop a short-term warning model for collision risk in work areas based on 1D CNN + LSTM + Attention. The model architecture is shown in Table 2. The model architecture consists of nine layers. The first two layers are 1D CNN convolutional layers, which extract traffic flow characteristics of the upstream section. Next, a pooling layer reduces the dimensionality and inputs the bidirectional LSTM layer for contextual learning of the extracted features. Furthermore, an attention mechanism is used to enhance the model's focus on important features. This is followed by a flattening layer and a fully connected layer, which flattens the multidimensional information into a one-dimensional nonlinear transformation. A dropout layer is used to randomly shut down neurons to prevent overfitting. Finally, a dense layer is applied to output the binary classification results.

[0161] Table 2 Architecture of the 1D CNN + LSTM + Attention model

[0162]

[0163] Based on the short-term warning model for collision risk in the work area developed above, time series prediction is performed on the last 10% of the data set, with the time series range from 1540710 frames to 1559335 frames. The time series change trend of the characteristic variables of small and large vehicles in each lane of the upstream section is as follows: Figure 3 As shown in the figure, each characteristic variable is a normalized value. The collision risk is predicted based on the developed short-term warning model for the operation area collision risk, as shown in the figure. Figure 4The figure below compares the predictions of the short-term collision risk warning model for work areas with the actual values. The actual risk level and predicted probability of high risk are disclosed for each predicted node. 23 nodes were predicted to have a high risk level, covering all 18 actual high-risk nodes. However, five low-risk nodes were incorrectly classified as high risk. The risk level of 142 of the 147 nodes was accurately predicted, for an overall prediction accuracy of 96.60%. This indicates that the developed short-term collision risk warning model for work areas can effectively distinguish between high-risk and low-risk categories and predict collision risk levels 55 seconds in advance, meeting the requirements for short-term collision risk warnings in work areas.

[0164] This application starts from the goal of short-term warning of collision risk in the work area. First, the temporal correlation between the traffic flow characteristics of the upstream section of the work area and the collision risk of the downstream dangerous section is analyzed. By comprehensively utilizing cross-correlation analysis, Pearson correlation coefficient analysis and Granger causality test, it is determined that the traffic flow characteristics of the upstream section of the work area have a significant impact on the collision risk of the downstream dangerous section after a time period of Δt. On this basis, a short-term warning model for collision risk is constructed using the traffic flow characteristics of the upstream section and the collision risk of the downstream dangerous section. A short-term warning method for collision risk in the work area based on traffic flow characteristics is proposed, that is, the collision risk of the downstream dangerous section after a time period of Δt is predicted by the traffic flow characteristics of the upstream section, and an early warning is issued to realize an effective dynamic warning mechanism.

[0165] The work area collision risk potential field model disclosed in this application uses the target vehicle as the potential field source, combines the wedge potential field distribution and the Yukawa potential field function, and uses the equivalent time distance Effectively solve the problem of potential field anisotropy, making the quantification of the impact of distance on risk more realistic; comprehensively consider the key influencing factors and combine the vehicle-road combination function Incorporating vehicle attributes (such as vehicle type) and road attributes (such as slope), the risk load is calculated based on the relative speed between the interacting object and the target vehicle. Q At the same time, it uses extreme field force as a quantitative indicator, focuses on the front and side interaction objects and collision possibilities of the target vehicle, and eliminates the interference of objects behind it, which greatly improves the accuracy and pertinence of the quantification of collision risks in the working area. It can also adapt to multiple scenarios such as straight driving, turning, different vehicle models and slopes, and comprehensively integrates multi-dimensional parameters and actual scene characteristics to achieve dynamic assessment of road vehicle collision risks and positioning of dangerous areas.

[0166] This application proposes an improved short-term collision risk warning model, combining three models to construct 1D CNN + LSTM + Attention. Based on the respective advantages of these three models, 1D CNN's local feature extraction, LSTM's time series data processing capabilities, and Attention's contextual focus, the improved short-term collision risk warning model constructed in this application performed better than all other model combinations on the test set, demonstrating the advantages of this comprehensive approach in generalization ability and comprehensive performance, and can more comprehensively capture and utilize complex patterns and relationships in the data, thereby achieving a better balance in various performance indicators.

[0167] This application improves the traffic safety level in the operation area and provides effective risk management tools for traffic management departments, enabling them to formulate scientific and reasonable traffic management strategies based on real-time traffic flow characteristics and collision risk prediction results.

[0168] The preferred embodiments of the present application have been described in detail above, but the present application is not limited to the above embodiments. Various changes can be made within the scope of knowledge possessed by ordinary technicians in this field without departing from the purpose of the present application.

[0169] Many other changes and modifications can be made without departing from the concept and scope of the present application. It should be understood that the present application is not limited to the specific embodiments, and the scope of the present application is defined by the appended claims.

Claims

1. A short-term warning method for collision risk in a work area based on traffic flow characteristics, characterized by: The following steps are involved: Step 1: Identify vehicles on the upstream section of the work area and collect historical traffic flow characteristics of the upstream section of the work area; Step 2: Identify vehicles on the dangerous road section downstream of the work area and calculate the historical collision risk of the downstream dangerous road section using the work area collision risk potential field model; The collision risk potential field model of the operation area is specifically as follows: taking the target vehicle as the field source, by constructing the risk field force F , collision potential field strength E and risk loading Q , converting the spatial position of the target vehicle and the interaction object into an equivalent time distance , combining the vehicle type and road slope to form a vehicle-road combination function , and then calculate the risk load based on the relative speed Q , and finally quantify the collision risk of the downstream dangerous road section through the extreme field force formula, realizing the dynamic assessment of the road vehicle collision risk and the positioning of the dangerous area; Step 3: Use temporal correlation analysis to determine and verify the time difference between the traffic flow characteristics of each upstream road segment and the collision risk of the downstream dangerous road segment. Use the time difference to align the traffic flow characteristics of the upstream road segment with the collision risk of the downstream dangerous road segment. Step 4: Use the traffic flow characteristics of the upstream section after time series alignment and the collision risk of the downstream dangerous section to train the collision risk short-term warning model; The collision risk short-term warning model consists of 9 layers, namely the first Conv1D layer, the second Conv1D layer, the MaxPooling1D pooling layer, the bidirectional LSTM layer, the SeqSelfAttention mechanism layer, the Flatten layer, the Dense fully connected layer, the Dropout layer and the Dense output layer; The first and second Conv1D layers are 1D CNN convolutional layers, used to extract traffic flow features of the upstream section. The MaxPooling1D pooling layer reduces the dimensionality and inputs the bidirectional LSTM layer for contextual learning of the extracted traffic flow features of the upstream section. The SeqSelfAttention mechanism layer then increases the model's focus on important features. This is followed by the Flatten layer and the Dense fully connected layer, which flattens the multidimensional information into a one-dimensional nonlinear transformation. The Dropout layer randomly shuts down neurons to prevent overfitting. Finally, the Dense output layer is applied to output the binary classification results. Step 5: Use the collision risk short-term warning model to predict the collision risk of the road vehicle to be assessed.

2. The method for short-term warning of collision risk in a work area based on traffic flow characteristics according to claim 1 is characterized in that: The step 1 specifically includes: deploying traffic flow monitoring equipment on the upstream section of the work area, identifying vehicles on the upstream section of the work area, collecting historical traffic flow data, the traffic flow data including traffic volume, maximum speed, average speed, minimum headway and average headway, using a sliding time window to divide the work area scene, and extracting historical traffic flow characteristics under each scene from the time series data; the work area includes an upstream section, an intermediate section and a downstream dangerous section, and the upstream section is the starting point of the warning area, which is used to warn downstream dangers.

3. The method for short-term warning of collision risk in a work area based on traffic flow characteristics according to claim 2 is characterized in that: Specifically, step 2 includes deploying traffic flow monitoring equipment on a dangerous road section downstream of the work area, identifying vehicles on the dangerous road section downstream of the work area, collecting historical vehicle trajectory data on the dangerous road section downstream of the work area, where the vehicle trajectory data includes speed, position, vehicle type, and road slope, and calculating the historical collision risk of the dangerous road section downstream using a collision risk potential field model of the work area based on the vehicle trajectory data.

4. The method for short-term warning of collision risk in a work area based on traffic flow characteristics according to claim 1 is characterized in that: The collision risk potential field model of the working area is: ； ； ； ； ； ； ； ； ； ； ； ； ； Where: F V For extreme field forces; is the collision potential field strength; The target vehicle and the interaction object i The risk field forces between them; For interactive objects i risk loading; is the position of the virtual collision point of the interaction object in the target vehicle coordinate system, unit: (m,m); is the vehicle-road combination function; is the equivalent time distance, unit: s -1 ; The coordinates of the interactive object in the road Frenet coordinate system, unit: (m,m); is the coordinate of the target vehicle in the Frenet coordinate system of the road, unit: (m,m); The vehicle's direction of travel and s The clockwise angle of the axis, unit: degree; is an equivalent parameter; M is the vehicle attribute parameter; I is the road attribute parameter; v is the target vehicle speed, unit: m / s; v 0 is the speed of the interactive object, unit: m / s; is the equivalent angle between the interaction object and the target vehicle, unit: degree; SSD is the stopping sight distance, unit: m; G is the longitudinal slope rate.

5. The method for short-term warning of collision risk in a work area based on traffic flow characteristics according to claim 1 is characterized in that: The step 3 is specifically as follows: the time series correlation analysis includes cross-correlation analysis, Pearson correlation coefficient analysis and Granger causality test, and the cross-correlation analysis is used to determine a reasonable time difference, that is, the length of time that can be used for early warning; then, based on the time difference, the Pearson correlation coefficient analysis is used to quantitatively evaluate the correlation between the traffic flow characteristics of the upstream section and the collision risk of the downstream dangerous section; finally, the Granger causality test is used to verify the correlation to test whether the traffic flow characteristics of the upstream section statistically significantly affect the change in the collision risk of the downstream dangerous section. If the test result shows that there is a causal relationship, it means that the time difference determined by the cross-correlation analysis is reasonable, and based on this, the traffic flow characteristics of the upstream section and the collision risk of the downstream dangerous section are time-series aligned, providing reliable time series input for the subsequent construction of a short-term collision risk warning model.

6. The method for short-term warning of collision risk in a work area based on traffic flow characteristics according to claim 1 is characterized in that: The collision risk short-term warning model determines 6 parameters as model tuning objects: filters1: the number of filters in the first Conv1D layer; kernel_size1: the convolution kernel size of the first Conv1D layer; filters2: the number of filters in the second Conv1D layer; kernel_size2: the convolution kernel size of the second Conv1D layer; lstm_units: the number of units in the bidirectional LSTM layer; dropout_rate: the dropout rate of the Dropout layer.

Images