An accident risk assessment method considering heterogeneity of risk-averse behavior

Abstract

The application discloses an accident risk assessment method considering the heterogeneity of risk-avoiding behavior, comprising: collecting trajectory data of a vehicle and preprocessing the trajectory data to enhance the trajectory quality; extracting the risk-avoiding behavior characteristics of a driver; using the relative motion state between vehicles and the risk-avoiding behavior information, adopting a Monte Carlo simulation method to probabilistically model the reaction time, the maximum deceleration and the maximum jerk of the driver, and calculating the probabilistic collision time (PCT); and comparing the calculated PCT value with a preset risk threshold to perform real-time accident risk assessment. The application incorporates the dynamic risk-avoiding behavior of the preceding vehicle and the following vehicle in a real collision scene into a risk assessment model, parameterizes the heterogeneous distribution of the reaction time, the maximum braking deceleration and the jerk, and generates a collision risk curve with a confidence interval. Compared with a traditional time collision index, the application breaks through the homogenization assumption of driving behavior, and can significantly improve the accuracy and timeliness of risk perception in a complex traffic scene.

Description

Technical Field

[0001] This invention belongs to the field of intelligent transportation safety analysis, specifically relating to an accident risk assessment method that considers the heterogeneity of risk avoidance behavior. Background Technology

[0002] Traditional traffic accident monitoring methods mainly rely on fixed monitoring equipment and manual inspections, which suffer from problems such as large blind spots, delayed response, and high false alarm rates. In recent years, accident risk assessment methods based on vehicle trajectory have gradually developed, but they still face the following challenges: 1. Insufficient data accuracy: Radar and video sensors are susceptible to environmental noise interference, leading to drift and fragmentation in trajectory data; 2. Limitations in driving behavior modeling: Relying on simulation or small sample data, failing to utilize emergency braking modes in large-scale real collision trajectories; 3. Existing methods mostly assume homogeneous driver behavior, ignoring the heterogeneity of parameters such as reaction time and braking intensity, making it difficult to accurately depict real risk scenarios; 4. Delayed risk assessment: Traditional time-of-flight collision indicators (such as TTC and MTTC) rely on the assumption of constant speed, failing to reflect dynamic interaction processes such as emergency braking and acceleration avoidance, resulting in insufficient warning timeliness. Summary of the Invention

[0003] Purpose of the invention: To address the shortcomings of the existing technology, the purpose of this invention is to propose an accident risk assessment method that considers the heterogeneity of risk avoidance behavior. By modeling driver risk avoidance behavior and dynamically calculating probabilistic crash time (PCT), this method breaks through the limitations of the homogeneous driving behavior assumptions of traditional models, achieves more accurate and real-time risk warnings, and provides technical support for improving the intelligent level of road safety management.

[0004] Technical solution: To achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0005] In a first aspect, the present invention provides an accident risk assessment method that considers the heterogeneity of risk-avoidance behavior, comprising the following steps:

[0006] Collect vehicle trajectory data and preprocess it to enhance trajectory quality;

[0007] Driver risk avoidance behavior features are extracted based on preprocessed trajectory data. Risk avoidance behavior includes one or more of the vehicle’s emergency braking, acceleration and steering. The extracted features include driver reaction time, maximum deceleration and maximum acceleration.

[0008] Using information on the relative motion state and avoidance behavior between vehicles, the Monte Carlo simulation method is used to probabilistically model the driver's reaction time, maximum deceleration and maximum acceleration, and calculate the probabilistic collision time (PCT).

[0009] Accident risk assessment is performed in real time by comparing the calculated PCT value with the preset risk threshold.

[0010] Furthermore, the preprocessing includes one or more of the following steps:

[0011] By comparing the spatiotemporal gaps of adjacent timestamps with preset thresholds, different vehicle trajectories can be separated.

[0012] Trajectory interpolation and correlation are performed by comparing speed differences;

[0013] Identify and eliminate ghost trails;

[0014] For trajectories with inconsistent classifications, the dominant category is determined by the proportion of their lengths.

[0015] Feature point drift is corrected based on velocity measurements.

[0016] Furthermore, the extraction of driver avoidance behavior features based on preprocessed trajectory data includes: determining the start time and duration of the avoidance behavior; extracting avoidance behavior features; wherein the start time of the vehicle's avoidance behavior is determined by speed difference analysis or wavelet energy peak detection, and the duration is from the start time to the collision time or the time when the distance between vehicles reaches the minimum value and the risk is eliminated; the extraction of driver avoidance behavior features includes: calculating the reaction time based on the propagation time difference of wavelet energy peaks; calculating the acceleration curve using the central difference method and extracting the maximum deceleration; performing linear fitting on the deceleration curve and extracting the maximum slope as the jerk.

[0017] Furthermore, the probabilistic collision time (PCT) calculation method includes:

[0018] Real-time collection of vehicle trajectory data, including longitudinal speed, acceleration, relative distance to the vehicle in front, and deceleration of the vehicle in front;

[0019] Reaction time, maximum deceleration, and maximum jerk are randomly selected from the target scene distribution;

[0020] Solving for the minimum collision time (MCT):

[0021] MCT = argmin{s net (t)=s LV (t)-s FV (t)+gap(t0)=0},t∈[0,T 3,FV (t)]

[0022] In the formula, t0 represents the initial time, and s net (t) is the workshop clearance function, s LV (t) and s FV (t) represents the travel distances of the front and rear vehicles, respectively, gap(t0) is the initial clearance between the vehicles, and T 3,FV(t) is the safe stopping time of the following vehicle calculated based on its current initial velocity and the assumed deceleration curve;

[0023] Through multiple Monte Carlo simulations, the probability density function of PCT is generated, and the confidence interval [PCT] at the preset confidence level is output. lower PCT upper [ ] represents the conservative and optimistic estimates of the collision time.

[0024] Furthermore, when calculating the minimum collision time (MCT), the braking process of the preceding and following vehicles is defined as comprising three stages:

[0025] The braking phase of the preceding vehicle includes an acceleration phase and a constant deceleration phase. The acceleration phase involves the preceding vehicle linearly increasing its deceleration from an initial acceleration to the maximum acceleration phase until it reaches the maximum braking deceleration. The constant deceleration phase involves maintaining a constant maximum braking deceleration level until the preceding vehicle stops once it reaches the maximum braking deceleration. The maximum acceleration and maximum braking deceleration are randomly selected based on an empirical probability distribution function.

[0026] Rear vehicle reaction phase: After detecting the braking of the preceding vehicle, the following vehicle undergoes a reaction time, during which it maintains its initial acceleration; the reaction time is randomly selected based on an empirical probability distribution function;

[0027] The braking phase of the following vehicle includes an acceleration phase and a constant deceleration phase. In the acceleration phase, the following vehicle linearly increases its acceleration from the initial acceleration to the maximum acceleration until it reaches the maximum braking deceleration. In the constant deceleration phase, when the following vehicle reaches the maximum braking deceleration, it remains constant until it stops. The maximum acceleration and maximum braking deceleration are randomly selected based on an empirical probability distribution function.

[0028] Furthermore, through multiple Monte Carlo simulations, the probabilistic collision time was obtained. In the formula Let std(MCT) be the mean of MCT, and std(MCT) be the standard deviation of MCT.

[0029] Furthermore, the accident risk assessment is conducted according to the following formula:

[0030]

[0031] In the formula, T th1 T th2 It is a preset risk threshold, determined through ROC curve analysis, using different PCT values. upper and PCT lower To determine accident risk and calculate different PCT upper and PCT lowerAccident prediction performance is assessed by taking the PCT corresponding to the point where the Youden index is maximized. upper and PCT lower As the optimal threshold point.

[0032] Furthermore, based on the PCT confidence interval and collision risk level, audible and visual alarms and rescue notifications are triggered, and a user interface is provided to display the vehicle's real-time location and trajectory, accident risk curve, and extract hazard avoidance behavior parameters to analyze the causes of the accident.

[0033] In a second aspect, the present invention provides a computer system including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the accident risk assessment method that considers the heterogeneity of risk avoidance behavior.

[0034] Thirdly, the present invention provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the accident risk assessment method that considers the heterogeneity of risk avoidance behavior.

[0035] Beneficial Effects: This invention provides an accident risk assessment method considering the heterogeneity of avoidance behavior. It extracts characteristic parameters of the driver's avoidance behavior before the collision by analyzing the trajectories of moving vehicles before the accident. Based on this, a Monte Carlo simulation framework considering the avoidance behavior of multiple vehicles in a platoon is constructed, and the Probabilistic Collision Time (PCT) is innovatively proposed as a new safety assessment indicator. This accident risk assessment method considering the heterogeneity of avoidance behavior can quantify and assess the level of accident risk. It is not only applicable to highway scenarios but can also be extended to complex environments such as urban intersections and tunnels, providing core risk perception capabilities for intelligent transportation systems and offering a reliable and practical means to improve the intelligence and real-time performance of traffic management. Compared with the prior art, the present invention has the following advantages: (1) It breaks through the homogeneity assumption of the traditional SSM and considers the heterogeneity of the risk avoidance behavior of multiple vehicles in the queue for the first time, which can dynamically reflect the impact of the risk avoidance behavior of following vehicles on the accident risk; (2) It generates the PCT confidence interval through Monte Carlo simulation to quantify the uncertainty of collision time. The experiment shows that in 260 accident scenarios on the Chengdu Ring Expressway, the accident risk detection recall rate reached 82.6% and the false alarm rate was 12.8%. Compared with TTC, the recall rate increased by 3.8% and the false alarm rate decreased by 3.5%. It can be seen that the present invention significantly improves the accuracy and timeliness of risk perception in complex traffic scenarios. Attached Figure Description

[0036] Figure 1 This is a flowchart illustrating the overall process of an embodiment of the present invention.

[0037] Figure 2 This is a layout diagram of accident data collection stations on the Chengdu Ring Expressway G4202.

[0038] Figure 3 A schematic diagram of the system equipment deployment and detection site for an accident.

[0039] Figure 4 This is a schematic diagram of the trajectory before the accident.

[0040] Figure 5 A schematic diagram illustrating the physical meaning of the MCT index.

[0041] Figure 6 This is a schematic diagram of parameters for risk-averse behavior.

[0042] Figure 7 The PCT probability density distribution generated for Monte Carlo simulation.

[0043] Figure 8 This is a schematic diagram of the PCT accident risk curve.

[0044] Figure 9 This is a comparison chart of ROC curves (PCT vs. TTC / MTTC / TTCD / THW). Detailed Implementation

[0045] The technical solution of the present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0046] like Figure 1 As shown in the figure, an accident risk assessment method considering the heterogeneity of risk avoidance behavior disclosed in an embodiment of the present invention includes the following steps:

[0047] Step S10: Collect vehicle trajectory data and preprocess it to enhance trajectory quality. This mainly includes:

[0048] Step S11: Utilize cameras, millimeter-wave radar, and edge computing devices deployed within the traffic monitoring area to collect vehicle trajectory data in real time, including vehicle position, speed, acceleration, and inter-vehicle distance. The data collection design mainly involves:

[0049] (1) Multimodal sensor collaborative deployment design. Camera array design: Select high-resolution (1080p or 4K) cameras with wide dynamic range (WDR) and night vision capabilities. Based on the traffic flow and terrain characteristics of the monitored area, adopt a stereo vision layout (such as a combination of fisheye cameras and telephoto cameras) to eliminate blind spot coverage. Millimeter-wave radar deployment: Use 77GHz band millimeter-wave radar with a horizontal scanning angle ≥120°, a detection distance ≥200m, and a vertical scanning angle ≥20°. Align the radar with the camera's spatial position and achieve spatiotemporal synchronization between the radar point cloud and video images through a calibration board. Multi-sensor fusion layout: Based on the principle of radar as the main sensor and video as the auxiliary sensor, deploy radar-camera joint perception units at key lane nodes (such as ramps and curves). The radar is responsible for accurate measurement of motion parameters (speed, acceleration), and the camera assists in target classification (vehicles / pedestrians) and trajectory continuity verification.

[0050] (2) Edge Computing Fusion Processing Unit Design. In terms of hardware architecture, each radar-camera joint sensing unit is equipped with an NVIDIA Jetson AGX Xavier edge computing device, featuring a built-in multi-core CPU (8-core ARM v8.2) and GPU (512-core Volta architecture), supporting parallel access of radar point clouds (RS232 protocol) and video streams (RTSP protocol). For time synchronization, radar and camera timestamps are aligned using PTP (Precise Time Protocol), with an error of <1ms. For spatial calibration, a transformation matrix between the radar polar coordinate system and the camera image coordinate system is established based on the checkerboard calibration method to achieve target position mapping.

[0051] Step S12: Trajectory Quality Enhancement: Using spatiotemporal threshold segmentation, velocity difference correlation, and ghost trajectory filtering algorithms, sensor noise and trajectory breaks are eliminated, outputting continuous trajectory segments. Specifically, this includes:

[0052] Step S121: Data segmentation method based on spatiotemporal interval threshold, wherein the formula for the spatiotemporal interval threshold is as follows:

[0053]

[0054] In the formula: T d S is the time interval threshold (in seconds). d L is the spatial interval threshold (unit: meters). max v represents the length of the radar detection area (in meters). max This represents the vehicle's maximum speed (in meters per second). By comparing the spatiotemporal gaps between adjacent timestamps with a preset threshold, effective separation of different vehicle trajectories under the same tracking target is achieved, correcting tracking errors.

[0055] Step S122: Trajectory interpolation and association method, wherein the velocity difference threshold formula for trajectory association is as follows:

[0056]

[0057] In the formula: x k,0 and t k,0 Let x be the longitudinal position and time of the end of the current k-th vehicle trajectory. k+1,1 and t k+1,1 v represents the longitudinal position and timestamp of the beginning of the candidate trajectory for the (k+1)th vehicle. k,0 Let SD be the velocity at the end of the current k-th vehicle's trajectory. k For speeds less than 17 km / h, linear interpolation is used to connect segmented trajectories of the same vehicle to correct target tracking loss errors.

[0058] Step S123: Ghost trajectory filtering method, wherein the trajectory length threshold determination formula is as follows:

[0059] L = |x last -x first |

[0060] In the formula: x last and x first These are the longitudinal coordinates of the beginning and end positions of the trajectory. If the trajectory length is less than 22m, it is judged as a ghost trajectory and removed, eliminating other non-vehicle detection targets.

[0061] Step S124: Classification Correction and Drift Correction Method

[0062] (1) Dominant Classification Correction: For trajectories with inconsistent classifications, the dominant category is determined based on the length proportion.

[0063]

[0064] In the formula: c i Let c be the classification result for the i-th timestamp, c be the classification result for trucks and cars, n be the total number of vehicle type classifications, and ΔL be the classification result for the i-th timestamp. i δ represents the length of the trajectory segment, and δ is the indicator function for different vehicle models.

[0065] (2) Cumulative drift correction: Correcting feature point drift based on velocity measurements.

[0066]

[0067] Feature point drift DF i The calculation is as follows:

[0068]

[0069] In the formula: Δx i Let Δx be the longitudinal displacement increment at the i-th time stamp. i =x i +xi-1 x i and x i-1 v represents the detection vertical position of the i-th and i-1th timestamps, respectively. i and v i-1 Δt represents the detection speed at timestamps i and i-1, respectively. i This is the time increment for the i-th timestamp.

[0070] Step S20: Extract driver avoidance behavior features based on preprocessed trajectory data. Avoidance behaviors include emergency braking, acceleration, and steering of the vehicle. The extracted features include driver reaction time, maximum deceleration, and maximum acceleration.

[0071] This step analyzes the vehicle's emergency braking, acceleration, and steering avoidance behaviors to determine the maximum deceleration, maximum acceleration, and reaction time of these avoidance behaviors. Specifically, it includes the following steps:

[0072] Step S21: Determining the Initiation Time of Avoidance Behavior: Avoidance behavior is defined as the actions taken by the driver, such as braking and steering, to avoid accident risks. It is necessary to identify the driver's avoidance behavior process from the full sample trajectory to facilitate the subsequent extraction of avoidance behavior parameters. Based on the moving vehicle trajectory from millimeter-wave radar, the initiation time of the corresponding vehicle's avoidance behavior is determined through speed difference analysis and wavelet energy peak detection. Specifically, as follows:

[0073] (1) Speed ​​difference threshold method: When the longitudinal speed difference ΔV between the following vehicle and the vehicle in front is v Fv -v LV When the historical maximum value is reached, it is marked as the start time t0 of the risk aversion behavior.

[0074] (2) Wavelet Energy Peak Method: Perform Mexican hat wavelet transform on the velocity curve, extract the wavelet energy spectrum, and mark the energy peak as the starting time t0 of the risk-avoidance behavior. The formula is expressed as:

[0075]

[0076] Where W(s,t) are wavelet transform coefficients, s is the scale parameter, and t is the time parameter.

[0077] Step S22: Extract the avoidance process, from the start time t0 to the collision time t crash (or the time t when the vehicle spacing reaches its minimum value and the risk is eliminated) safe The time difference is used as the duration T. evasion Extract the complete vehicle avoidance process for subsequent avoidance behavior parameter estimation:

[0078] T evasion =t crash-t0

[0079] T evas i on =t safe -t0

[0080] Step S23: Quantify the risk aversion intensity, based on the analysis of deceleration and jerk, as follows:

[0081] (1) Maximum deceleration: The maximum deceleration value is extracted from the acceleration curve calculated by the central difference method.

[0082]

[0083] Where Δt = 0.1s is the time window.

[0084] (2) Maximum jerk: Linear fitting is performed on the deceleration curve a, and the maximum slope is extracted as the jerk:

[0085]

[0086] Step S24: Reaction Time Estimation: Calculate the reaction time based on the propagation time difference of wavelet energy peaks, and extract the wavelet energy peak time t of the vehicle ahead (LV) and the following vehicle (FV). FV and t LV Reaction time is defined as:

[0087] τ FV =t FV -t LV

[0088] Step S30: Using information on the relative motion states and avoidance behaviors between vehicles, a Monte Carlo simulation method is employed to probabilistically model the driver's reaction time, maximum deceleration, and maximum acceleration, and to calculate the probabilistic collision time (PCT). Specifically, this includes:

[0089] Step S31: Collect vehicle trajectory data in real time, including longitudinal velocity, acceleration, relative distance to the vehicle in front, and deceleration of the vehicle in front at time t; calibrate data accuracy and eliminate sensor noise by using millimeter-wave radar and video fusion technology.

[0090] Step S32: Construct a dynamic risk parameter calculation framework and generate parameter distribution based on Monte Carlo simulation:

[0091]

[0092] In the formula, P τ P a P j Based on collision dataset D Crash Or close to collision dataset DNear-Crash The estimated empirical probability distribution functions of reaction time, deceleration, and acceleration that conform to Chinese driving style are obtained from the dataset of moving vehicle trajectory data collected in advance on the target road segment. When the collection time is long enough, the corresponding parameters can be obtained using the collision dataset. When the collection time is insufficient, the corresponding parameters can be identified by the driver's avoidance behavior process when approaching a collision.

[0093] Step S33: Probabilistic Collision Time (PCT) Calculation: Based on Monte Carlo simulation, the heterogeneity of avoidance behavior is simulated. This heterogeneity refers to the varying degrees of avoidance behavior taken by different drivers in different scenarios, including the timing, intensity, and rate of the avoidance behavior. A probabilistic collision time distribution is generated based on the distribution of avoidance behavior heterogeneity.

[0094] (1) Parameter sampling: Randomly sample reaction time τ from the target scene distribution. FV Maximum deceleration a min,break Maximum jerk j max The parameters of the leading vehicle are represented by the subscript LV (Leading Vehicle), and the parameters of the following vehicle are represented by the subscript FV (Following Vehicle). The parameter values ​​of the leading and following vehicles are different. The collision risk probability is considered under various relative deceleration motions of the leading and following vehicles through parameter sampling.

[0095] (2) Kinematic modeling: such as Figure 2 As shown, the braking process of the vehicle in front and the vehicle behind is defined as including three stages:

[0096] The braking phase of the vehicle in front includes an acceleration phase and a constant deceleration phase. During the acceleration phase, the vehicle in front increases its deceleration linearly from its initial acceleration to its maximum acceleration until it reaches its maximum braking deceleration. During the constant deceleration phase, when the LV deceleration reaches the maximum braking deceleration, it maintains a constant maximum braking deceleration level until it comes to a stop.

[0097]

[0098] In the formula: T 2,LV Let the vehicle in front have an initial acceleration a at the current moment. t0,LV The time to reach maximum deceleration T 3,LV The time it takes to decelerate to a stop with maximum braking deceleration.

[0099] Rear vehicle reaction phase: After detecting the braking of the preceding vehicle, the following vehicle goes through a reaction time, during which it maintains its initial acceleration.

[0100] The braking phase of the following vehicle includes an acceleration phase and a constant deceleration phase. During the acceleration phase, the following vehicle linearly increases its acceleration from its initial acceleration to its maximum acceleration until it reaches its maximum braking deceleration. During the constant deceleration phase, once the following vehicle reaches its maximum braking deceleration, it maintains a constant deceleration until it comes to a stop.

[0101]

[0102] In the formula: a t0,FV Let τ be the initial deceleration of the following vehicle. FV T is the reaction time of the following vehicle. 1,FV The following vehicle has an initial acceleration a at the current moment. t0,FV The time T to reach maximum deceleration 2,FV The moment when the following vehicle reaches its maximum braking deceleration. T 3,FV The time it takes for the following vehicle to decelerate to a stop at its maximum braking deceleration.

[0103]

[0104] Constructing the vehicle's equations of motion:

[0105]

[0106] Solve for the minimum collision time (MCT):

[0107] MCT = argmin{s net (t)=s LV (t)-s FV (t)+gap(t0)=0},t∈[0,T 3,FV (t)]

[0108] In the formula, s net It is the workshop clearance function, s LV (t) and s FV (t) represents the travel distances of the front and rear vehicles, respectively. gap(t0) is the initial clearance between the vehicles, T 3,FV (t) represents the safe stopping time of the following vehicle, calculated based on its current initial velocity and the assumed deceleration curve.

[0109] (3) Calculation of probability collision risk index: Through multiple (e.g., 100) Monte Carlo simulations, the sampled parameters are made close to the prior distribution function, generating the probability density function (PDF) of PCT and outputting the 95% confidence interval:

[0110]

[0111] In the formula, Let std(MCT) be the mean of MCT, and std(MCT) be the standard deviation of MCT.

[0112] Step S34: Calculate the probabilistic collision time (PCT) and its confidence interval: Based on the Monte Carlo simulation results, generate the 95% confidence interval [PCT] for PCT. lower PCT upper [ ] represents the conservative and optimistic estimates of the collision time.

[0113] Step S35: Definition of accident prediction performance indicators: Define real accident risk labels based on accident scenarios and high-dimensional scenarios. Through the binary classification results of accident risk assessment, use the confusion matrix to accurately calculate recall, false alarm rate (FAR), F1 score, and Youden, evaluate the accident prediction performance, and determine the optimal warning threshold for PCT.

[0114]

[0115] In the formula, TP represents the number of times an accident scenario is judged as an unsafe scenario, FN represents the number of times an accident scenario is mistakenly judged as a safe scenario, TN represents the number of times a safe scenario is correctly identified as a safe scenario, and FP represents the number of times a normal scenario is mistakenly judged as a high-risk scenario.

[0116] Step S40: Compare the calculated PCT value with the preset risk threshold to perform a real-time accident risk assessment, and send early warning information to the traffic management center and relevant terminals through the online feedback module. Specifically, this includes:

[0117] Step S41: Design of dynamic risk assessment algorithm:

[0118]

[0119] In the formula, the threshold T th1 T th2 ROC curve analysis determined that different PCT values ​​were used. upper and PCT lower To determine accident risk and calculate different PCT upper and PCT lower Accident prediction performance is assessed by taking the PCT corresponding to the point where the Youden index is maximized. upper and PCT lower As the optimal threshold point, it can simultaneously maximize the incident detection rate (Recall) and minimize the incident false alarm rate (FAR).

[0120] Step S42: Trigger audible and visual alarms and rescue notifications based on the PCT confidence interval and collision risk level. Adjust the alarm triggering mechanism according to the needs of the traffic management department. For high-risk scenarios such as heavy fog, heavy rain, and construction zones, focus on warnings of medium risk and above. When a high-risk warning is triggered, the traffic management department should formulate corresponding evacuation and emergency rescue strategies after the accident.

[0121] Step S43: Install audible and visual alarm devices at the traffic monitoring center or on relevant vehicles. When the accident risk level exceeds the threshold, an audible and visual alarm will be issued immediately to attract the attention of relevant personnel. The system will send text messages or make phone calls to relevant personnel (such as traffic management personnel, rescue teams, etc.) through preset phone numbers. The text message / phone call content includes detailed information such as the accident type, time of occurrence, vehicles involved in the accident, geographical location, and severity.

[0122] Step S44: To facilitate monitoring and management, the system provides a user interface display module, including a real-time map, an event list, and detailed event information: real-time display of vehicle location and trajectory, accident risk curve, and extraction of avoidance behavior parameters to analyze the cause of the accident, whether it is due to insufficient driver avoidance effort or excessive reaction time.

[0123] The following is a specific example of a risk assessment for a rear-end collision on a highway.

[0124] Specifically, the accident risk assessment method that considers the heterogeneity of risk-avoidance behavior provided in this example has the following main steps:

[0125] Data Acquisition: Eight sets of millimeter-wave radars and linked cameras were deployed on the Chengdu Ring Expressway (see...). Figure 3 The system collected vehicle trajectories at a frequency of 20Hz; a total of 260 real accident scenarios were detected. The authenticity of the rear-end collisions was confirmed through video playback, and radar trajectory data from one hour before the accident to one hour after the accident was extracted. For example, Figure 4 The video shows the scene of the rear-end collision that occurred at 17:08:06 on May 17, 2024, at chainage ZK39+161. The corresponding radar track can be found here. Figure 5 .

[0126] Feature extraction: From all collected pre-accident and accident scenarios, the avoidance behavior curves of the accident vehicle and surrounding following vehicles are extracted. Avoidance behavior parameters, including reaction time, maximum deceleration, and maximum acceleration, are extracted from these curves. (See details...) Figure 6 The distribution of risk aversion behavior parameters is generated based on the kernel density estimation method.

[0127] PCT calculation: Monte Carlo sampling of risk-averse behavior parameters, the distribution of the generated risk-averse behavior parameters is shown in [see...]. Figure 7The PCT probability collision time and 95% confidence interval are calculated based on the corresponding parameters.

[0128] Risk assessment: See the risk curve before the accident. Figure 8 For vehicles involved in an accident, during the approach to a collision, the 95% confidence interval of the PCT gradually narrows and eventually converges, indicating that the collision evolves into an unavoidable event. Meanwhile, the PCT... upper >T th2 It can trigger a medium-risk warning, and the system sends a level-two warning to the monitoring center. The average advance warning time can reach 1.51 seconds, which can effectively reduce the probability of accidents.

[0129] Performance Verification: Based on tests using 260 sets of real-world accident data, the ROC performance curve is shown below. Figure 9 As shown in the figure, the AUC of PCT is 0.92, and the false alarm rate and false negative rate are significantly improved compared with traditional TTC, MTTC, TTCD and THW.

[0130] The embodiments of the present invention have the following effects: Improved accuracy: Through heterogeneous parameter modeling, PCT achieves a recall rate of 82.6% and a false alarm rate of 12.8% for accident scenarios; Strong real-time performance: The edge computing unit realizes local data processing, with a warning latency of <200ms; High scalability: It supports continuous braking analysis of multiple vehicles in the queue, with F1-scores all above 0.91.

[0131] This invention also discloses a computer system, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, it implements the steps of the accident risk assessment method that considers the heterogeneity of risk avoidance behavior.

[0132] This invention also discloses a computer program product, including a computer program that, when executed by a processor, implements the steps of the accident risk assessment method that considers the heterogeneity of risk avoidance behavior.

[0133] The program code used to implement the method of the present invention can be written in any combination of one or more programming languages. This program code can be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the program code causes the steps of the method of the present invention to be performed. The program code can be executed entirely on the machine, partially on the machine, partially on the machine and partially on a remote machine as a standalone software package, or entirely on a remote machine or server. All aspects not detailed in this invention are well-known to those skilled in the art.

Claims

1. An accident risk assessment method considering the heterogeneity of risk-avoidance behavior, characterized in that, Includes the following steps: Collect vehicle trajectory data and preprocess it to enhance trajectory quality; Driver risk avoidance behavior features are extracted based on preprocessed trajectory data. Risk avoidance behavior includes one or more of the vehicle’s emergency braking, acceleration and steering. The extracted features include driver reaction time, maximum deceleration and maximum acceleration. Using information on the relative motion state and avoidance behavior between vehicles, the Monte Carlo simulation method is used to probabilistically model the driver's reaction time, maximum deceleration and maximum acceleration, and calculate the probabilistic collision time (PCT). Accident risk assessment is performed in real time by comparing the calculated PCT value with the preset risk threshold. Methods for calculating Probabilistic Collision Time (PCT) include: Real-time collection of vehicle trajectory data, including longitudinal speed, acceleration, relative distance to the vehicle in front, and deceleration of the vehicle in front; Reaction time, maximum deceleration, and maximum jerk are randomly selected from the target scene distribution; Solving for the minimum collision time (MCT): ； In the formula, Indicates the initial time. It is the workshop clearance function. and These represent the distances traveled by the vehicles in front and behind, respectively. This is the initial workshop clearance. The safe stopping time of the vehicle is calculated based on the current initial velocity and the assumed deceleration curve of the following vehicle. Through multiple Monte Carlo simulations, the probability density function of PCT is generated, and the confidence interval [PCT] at the preset confidence level is output. lower PCT upper [ ] represents the conservative and optimistic estimates of the collision time.

2. The accident risk assessment method considering the heterogeneity of risk-avoidance behavior according to claim 1, characterized in that, The preprocessing includes one or more of the following steps: By comparing the spatiotemporal gaps of adjacent timestamps with preset thresholds, different vehicle trajectories can be separated. Trajectory interpolation and correlation are performed by comparing speed differences; Identify and eliminate ghost trails; For trajectories with inconsistent classifications, the dominant category is determined by the proportion of their lengths. Feature point drift is corrected based on velocity measurements.

3. The accident risk assessment method considering the heterogeneity of risk-avoidance behavior according to claim 1, characterized in that, Extracting driver avoidance behavior features based on preprocessed trajectory data includes: determining the start time and duration of the avoidance behavior; extracting avoidance behavior features; wherein the start time of the vehicle's avoidance behavior is determined by speed difference analysis or wavelet energy peak detection, and the duration is from the start time to the collision time or the time when the distance between vehicles reaches the minimum value and the risk is eliminated; the extraction of driver avoidance behavior features includes: calculating the reaction time based on the propagation time difference of wavelet energy peaks; calculating the acceleration curve using the central difference method and extracting the maximum deceleration; linearly fitting the deceleration curve and extracting the maximum slope as the maximum jerk.

4. The accident risk assessment method considering the heterogeneity of risk-avoidance behavior according to claim 1, characterized in that, When calculating the minimum collision time (MCT), the braking process of the preceding and following vehicles is defined as consisting of three stages: The braking phase of the preceding vehicle includes an acceleration phase and a constant deceleration phase. The acceleration phase involves the preceding vehicle linearly increasing its deceleration from an initial acceleration to the maximum acceleration phase until it reaches the maximum braking deceleration. The constant deceleration phase involves maintaining a constant maximum braking deceleration level until the preceding vehicle stops once it reaches the maximum braking deceleration. The maximum acceleration and maximum braking deceleration are randomly selected based on an empirical probability distribution function. Rear vehicle reaction phase: After detecting the braking of the preceding vehicle, the following vehicle goes through a reaction time, during which it maintains its initial acceleration; The reaction time is randomly selected based on an empirical probability distribution function; The braking phase of the following vehicle includes an acceleration phase and a constant deceleration phase. The acceleration phase is in which the following vehicle linearly increases its acceleration from the initial acceleration to the maximum acceleration until it reaches the maximum braking deceleration. The constant deceleration phase is in which the following vehicle maintains a constant deceleration until it stops when it reaches the maximum braking deceleration. The maximum acceleration and maximum braking deceleration are randomly selected based on an empirical probability distribution function.

5. The accident risk assessment method considering the heterogeneity of risk-avoidance behavior according to claim 1, characterized in that, Through multiple Monte Carlo simulations, the probability collision time PCT was obtained. ±1.96⋅std(MCT); where Let std(MCT) be the mean of MCT, and std(MCT) be the standard deviation of MCT.

6. The accident risk assessment method considering the heterogeneity of risk-avoidance behavior according to claim 1, characterized in that, Accident risk assessment is performed according to the following formula: ； In the formula, , It is a preset risk threshold, determined through ROC curve analysis. and To determine accident risk and calculate different and Accident prediction performance is achieved by taking the point corresponding to the maximum Youden index. and As the optimal threshold point.

7. The accident risk assessment method considering the heterogeneity of risk-avoidance behavior according to claim 1, characterized in that, Based on the PCT confidence interval and collision risk level, the system triggers audible and visual alarms and rescue notifications, provides a user interface to display the vehicle's real-time location and trajectory, accident risk curve, and extracts hazard avoidance behavior parameters to analyze the causes of the accident.

8. A computer system comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the computer program is executed by the processor, it implements the steps of an accident risk assessment method that considers the heterogeneity of risk avoidance behavior as described in any one of claims 1-7.

9. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of an accident risk assessment method that considers the heterogeneity of risk avoidance behavior as described in any one of claims 1-7.

Images