Deep learning discrimination method for dangerous lane changing behavior of motor vehicle

Abstract

The invention discloses a deep learning discrimination method for a dangerous lane changing behavior of a motor vehicle, and relates to deep learning dynamic discrimination for the dangerous lane changing behavior of the motor vehicle based on large-scale historical trajectory data and accident data analysis. The method comprises the steps of construction based on a dangerous lane changing behavior historical database, extraction of motor vehicle dangerous lane changing track features and environment features, construction of a dangerous lane changing behavior deep learning model based on multi-dimensional features, dangerous lane changing behavior recognition and alarm based on dynamic tracks and the like. The limitation that only passive post-event management is adopted in a traditional supervision method is broken through, the dangerous lane changing behavior of the motor vehicle is found in time by fully utilizing historical track data characteristics and historical accident experience, the monitoring coverage range of the dangerous lane changing behavior of the road can be remarkably expanded, and related accident risks are reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of national highway motor vehicle driving safety supervision, and particularly relates to a deep learning discrimination method for dangerous lane-changing behavior of a motor vehicle. BACKGROUND

[0002] Dangerous lane-changing behavior of a motor vehicle constitutes a serious safety hazard in road traffic in China. From the perspective of traffic order, random lane-changing can disrupt normal traffic order, cause a 30-40km / h average deceleration of rear vehicles, and increase the probability of traffic congestion by 58%. In terms of accident risk, dangerous lane-changing on urban roads accounts for 23% of all scraping accidents, and on highways, it is more likely to cause multiple vehicles to collide in a chain, with a 37% higher fatality rate than ordinary accidents. Typical cases show that in the collision accidents caused by lane-changing on the Wen-ci section of S231 provincial road, the lane-changing vehicle bears 100% responsibility. At the legal level, Article 44 of the Regulations on the Implementation of the Road Traffic Safety Law clearly stipulates that lane-changing shall not affect other vehicles, and violators will face demerit points and fines.

[0003] Existing monitoring technologies mainly fall into the following categories: fixed video monitoring systems use AI algorithms to identify lane-changing trajectories, with an accuracy rate of up to 89%, but with a 15% night-time misjudgment rate; mobile law enforcement devices use laser ranging technology to accurately capture lane-changing distances, but the 80,000 yuan per unit cost restricts large-scale deployment; vehicle-mounted OBD monitoring can record turn signal usage, but the current installation rate is less than 15% and is easily manually turned off by drivers. Emerging technologies, such as radar and vision fusion systems, reduce the false positive rate to less than 5%, but require dense infrastructure support with equipment every 500 meters. Big data analysis platforms integrate ETC, GPS, and other multi-source data to establish a portrait of vehicle lane-changing behavior, but there is a technical bottleneck of 10-15 meter positioning drift.

[0004] Existing dangerous lane-changing monitoring technologies mainly fall into three categories: fixed video monitoring systems use AI algorithms to identify lane-changing trajectories, with an accuracy rate of up to 89%, but with a 15% night-time misjudgment rate; mobile law enforcement devices use laser ranging technology to accurately capture lane-changing distances, but the 80,000 yuan per unit cost restricts large-scale deployment; vehicle-mounted OBD monitoring can record turn signal usage, but the current installation rate is less than 15% and is easily manually turned off by drivers. Emerging technologies, such as radar and vision fusion systems, reduce the false positive rate to less than 5%, but require dense infrastructure support with equipment every 500 meters. Big data analysis platforms integrate ETC, GPS, and other multi-source data to establish a portrait of vehicle lane-changing behavior, but there is a technical bottleneck of 10-15 meter positioning drift.

[0005] Therefore, there is an urgent need for a motor vehicle dangerous lane changing behavior deep learning discrimination method based on trajectory data mining, which can timely discover the dangerous lane changing behavior of motor vehicles by using historical trajectory data features and historical accident experience, significantly improve the monitoring coverage of dangerous lane changing behavior on the road, and reduce the risk of related accidents. SUMMARY

[0006] The present application aims to overcome the shortcomings of the prior art, and provides a motor vehicle dangerous lane changing behavior deep learning discrimination method based on trajectory data mining, which can timely discover the dangerous lane changing behavior of motor vehicles by using historical trajectory data features and historical accident experience, significantly improve the monitoring coverage of dangerous lane changing behavior on the road, and reduce the risk of related accidents.

[0007] To achieve the above-mentioned application purposes, the technical solutions adopted by the present application are as follows:

[0008] The present application provides a motor vehicle dangerous lane changing behavior deep learning discrimination method, which comprises the following steps:

[0009] S1, construction based on dangerous lane changing behavior historical database: by accessing the event record information of the road traffic accident processing information system, a historical data set containing dangerous lane changing events is established as the basis for subsequent model training;

[0010] S2, extraction of motor vehicle dangerous lane changing trajectory features and environmental features: collect vehicle trajectory data of the accident, calculate the lane changing index according to the vertical distance between the trajectory point and the road alignment, and perform spatiotemporal continuous image processing on the vehicle trajectory data; identify whether the vehicle has a speeding behavior, calculate the corresponding risk speeding index, and construct a risk speeding event multi-dimensional feature vector from multiple angles;

[0011] S3, construction of a dangerous lane changing behavior deep learning model based on multi-dimensional features: using the multi-dimensional feature vector constructed above, a vehicle speeding risk double-task deep learning prediction model is trained, which realizes the synchronous output of speeding behavior classification and risk index regression by jointly processing structured traffic features and spatiotemporal matrix data;

[0012] S4, dangerous lane changing behavior recognition and alarm based on dynamic trajectory: input the real-time acquired vehicle trajectory data and related environmental information into the trained deep learning model, and the deep learning model analyzes and discriminates the real-time data, and when a dangerous lane changing behavior is identified, an alarm is triggered in real time to notify the relevant regulatory departments or take intervention measures.

[0013] Preferably, the S2 specifically comprises the following steps:

[0014] S2-1. Extraction of Dangerous Lane Change Events Based on Trajectory Spatiotemporal Matching Analysis: Collect vehicle trajectory data within a specific time period before and after the accident and within a 300-meter spatial range of the accident point to construct a neighborhood trajectory database of the accident location; calculate the lane change index Sft based on the vertical distance between the trajectory point and the road alignment; if the lane change index Sft is greater than a preset threshold of 3 meters, it is determined to be a dangerous lane change accident; transform the vehicle trajectories around the accident point into a discretized spatiotemporal matrix M(t) to simulate the continuous dynamic image of vehicle distribution before the accident, and achieve continuous representation of vehicle motion state by completing the neighborhood vehicle trajectories second by second, i.e., generate spatiotemporally continuous image data;

[0015] S2-2, Risk of Speeding Event Identification and Risk of Speeding Index Calculation: Identify whether the vehicle has significant speeding behavior before the accident occurs, and calculate the corresponding "Risk of Speeding Index", which comprehensively considers the severity of the accident and the degree of relative speeding;

[0016] S2-3, Construction of Multidimensional Feature Vectors for Risky Speeding Events: Traverse the set of risky speeding events, and for the current event, conduct event feature analysis and multidimensional vector construction from five perspectives: static traffic features, dynamic traffic features, event environmental influencing factors, speeding trajectory attributes and behavioral features, and construction of multidimensional feature vectors for risky speeding events.

[0017] Preferably, the specific steps of S2-1 are as follows:

[0018] S2-1-1. For the set of vehicle trajectories A_Cars(A,C,G) that can match accident records, where A is the list of corresponding accident records, C is the list of vehicles that match the accident, and G is the set of matching vehicle trajectory data, iterate through each accident record a, and extract the corresponding matching vehicle trajectory data g(a,c) from the matching vehicle c of a within a range of 5 minutes before and after the accident record time.

[0019] S2-1-2. Iterate through each trajectory point of g(a,c) sequentially, and obtain the vehicle speed closest to the time of the accident record as the zero time t_accident_real, which is the precise time of the accident. Five minutes before this time is the corresponding adjacent time period t_seg, and the latitude and longitude coordinates c_accident_real corresponding to this time are the precise location of the accident.

[0020] S2-1-3, Further extract the list of trajectory points within 300 meters of c_accident_real: Tr_pre(a,c)={p1,p2,...,p n}, p n That is, c_accident_real point. The lane change index Sft(a,c) is calculated based on the status of each trajectory point in the list and the spatial relationship of the adjacent road network.

[0021] After traversing A through S2-1-1 to S2-1-3, the accurate occurrence time set A_T of the dangerous lane-changing accident can be obtained, and the corresponding occurrence time period set A_T_Seg and the position set A_C are obtained;

[0022] S2-1-4, traverse each vehicle c in the total vehicle list C_total and the corresponding total vehicle trajectory data list G_total, and obtain the trajectory point subset g_S(c,A_m) consistent with the time range of A_T_seg from the trajectory set g(c) of each vehicle c, wherein A_m is the subset of the corresponding accident record list of the vehicle c consistent with the time range of A_T_seg. Assuming that the occurrence time period of a_m in A_T_seg is 2019-08-1814:30:14-14:35:14, and the vehicle c has a trajectory record set g'(c) in this time period, the set is added to g_S(c,A_m), and the accident a_m record is added to A_m;

[0023] S2-1-5, if g_S(c,A_m) is an empty set, jump to the next vehicle, otherwise traverse each accident record a_m in A_m, search the 500-meter range around the accurate occurrence position c_accident_real_m of a_m, if the trajectory data set g'(c) corresponding to the vehicle exists in the range with the same driving direction and the included angle of the forward direction less than 90 degrees, it is considered that the trajectory data set g(c,a_m) in the corresponding range of the vehicle and the accident record a_m have a spatio-temporal matching relationship;

[0024] After traversing T_Cars(C_total,G_total) through S2-1-4 to S2-1-5, the corresponding spatio-temporal matching vehicle set M_C, the matching trajectory set M_G, and the neighborhood spatio-temporal continuous image set STG of all accidents can be obtained;

[0025] The calculation method of the lane-changing index Sft is as follows:

[0026] (1) Traverse each trajectory point p1~p n , calculate the included angle of the forward direction p i →p i+1 and the adjacent road segment of the road network, and record the road segment with an included angle less than 90 degrees as the same direction. Select the matching driving road segment with the smallest included angle of the adjacent road segment in the same direction as the matching driving road segment, and the included angle is β i . The search range of the road segment is 80 meters, and there are several adjacent road segments.

[0027] (2) For the current point p i, further search from the total trajectory database to select a historical passing vehicle same direction trajectory point set HP = {hp1, hp2,... hp m} within a 80-meter range adjacent to the accident point, and calculate p i the vertical distance of each point in the horizontal direction, thereby obtaining the maximum vertical distance dmax1 i in the counterclockwise direction and the maximum vertical distance dmax2 i in the clockwise direction;

[0028] (3) According to the vertical distance of the trajectory point and the road alignment, calculate the lane changing index Sft, and the formula of the lane changing index Sft is as follows:

[0029]

[0030] i.e. the cumulative horizontal lateral displacement from the p1 point to the p n point.

[0031] If Sft(a, c) > 3 meters, it is considered that the vehicle has changed lanes horizontally during the accident occurrence stage, and further identifies the accident as a dangerous lane changing accident.

[0032] Preferably, in S2-1, the trajectory data is continuously spatio-temporally imaged as follows:

[0033] (1) Taking the accident point c_accident_real_m as the center and 500 meters as the radius, a circle is drawn, and the circle is divided vertically up and down within a range of 100 meters in the direction of the vehicle 100 meters before the accident. Finally, the divided area is divided at an interval of 1 meter to obtain a discrete matrix space region Z. For any time t, each cell in the region corresponds to a vector (x, y, m), x and y are coordinates, and m is the vehicle type number corresponding to the trajectory point in the cell at the current time. If there is no trajectory point in the cell, m is 0.

[0034] (2) Neighborhood vehicle trajectory completion; for the trajectory set g(c, a_m) of a vehicle c that matches the accident point space during the accident a_m occurrence period, the trajectory is completed second by second; assuming that the first trajectory point p1 corresponds to time t1, and the second trajectory point p2 corresponds to time t2, then the line connecting p1 to p2 is divided at an interval of (t2-t1-1) seconds. For the last point p4, the distance is calculated by the driving speed of p4, and the path point is completed by the extension line from p3 to p4. The continuous trajectory g'(c, a_m) starting from p1 is obtained, and the continuous trajectory and Z are intersected to obtain the cell number corresponding to each continuous trajectory point; the accident a_m occurrence period refers to a 5-minute segment before the accident occurrence time;

[0035] (3) Traverse each neighborhood vehicle c of the accident point, including the accident vehicle, to obtain the corresponding completed continuous trajectory, and select the first trajectory point time t first of the latest vehicle appearing in the occurrence period as the starting time of the generation of the neighborhood vehicle spatiotemporal image;

[0036] (4) Neighborhood vehicle spatiotemporal continuous imaging data generation; generate a regional Z vehicle distribution matrix M(t) of each second from t first to the occurrence time of the accident, which is represented as follows:

[0037]

[0038] Wherein, m(P, Q) is the current trajectory point belonging to the vehicle type number in the corresponding square, and 0 if there is no trajectory point.

[0039] Preferably, the specific steps of S2-2 are as follows:

[0040] S2-2-1, traverse each record speed v accident in the estimated pre-accident vehicle speed set A_V, if the spatiotemporal matching vehicle set M_C(a) of the corresponding accident a is not an empty set, then calculate the neighborhood average speed avg_v(a) of M_V(a) and the root variance var_v(a), if v accident > (avg_v(a) + var_v(a)), then define the accident a as a risk overspeed accident;

[0041] S2-2-2, for the risk overspeed accident a, calculate the risk overspeed index, and the risk overspeed index calculation formula is as follows:

[0042] indi_v(a) = s(a) * avg_v(a) / (v accident- avg_v(a)),

[0043] Wherein, s(a) is the severity parameter of the accident a, and the value of the “simple” accident is 1, the value of the “general” accident is 5, and the value of the “general” accident above is 20; In addition, (v accident- avg_v(a)) is placed in the numerator, which means that when the overspeed behavior occurs, the smaller the relative overspeed degree, the greater the severity of the accident, and the higher the risk overspeed index;

[0044] S2-2-3, obtain the entire risk overspeed event set O_A through S2-2-1 and S2-2-2.

[0045] Preferably, in S2-3, by traversing the risk overspeed event set O_A, the current event a is analyzed from the following five aspects and a multi-dimensional vector is constructed.

[0046] (1) Static traffic features; extract the fixed attributes of the accident road section, including the precise location c_accident_real of the event a, the adjacent road attribute parameters and the neighborhood road line geometry parameters; the adjacent road attribute parameters include parameter level, lane number, design speed, road type, and the road type includes bridge, tunnel, highway and urban road; the adjacent road attribute parameters are obtained by searching the adjacent corresponding GIS map data attributes; the neighborhood road line geometry parameters include slope, curvature and turning radius, which are calculated according to the trajectory position and elevation information;

[0047] (2) Dynamic traffic features; collect traffic flow data during the accident occurrence period, including various statistical features of the event a corresponding matching vehicle set M_C(a) and trajectory set M_G(a), including traffic flow level and vehicle type and corresponding average speed list; the traffic flow level is represented as follows:

[0048] vol(a) = num(M_C(a)) * θ,

[0049] Where num(M_C(a)) is the number of vehicles in M_C(a); θ is the average sampling rate, which is calculated by the ratio of the actual observation point traffic flow and the corresponding matching trajectory vehicle number; the trajectory vehicle number is the number of vehicles corresponding to the space-time matching data of the observation point observation period and the total vehicle trajectory set T_Cars(C_total, G_total);

[0050] (3) Event environmental impact factors; the event environmental impact factors include weather, visibility, temperature, humidity, road construction conditions and other environmental information recorded at the time of the accident event;

[0051] (4) Speeding trajectory attribute and behavior characteristics; the speeding trajectory attribute and behavior characteristics include accident vehicle type, continuous driving time before the accident, driver attributes and observation speed list Tr-V_pre(a, c) within 100-300 meters before the accident; the continuous driving time before the accident is obtained by statistics of the continuous driving trajectory time of the vehicle before the accident, and the interruption driving interval determination time is 30 minutes; the driver attributes include at least age and gender;

[0052] (5) Risk speeding multi-dimensional feature vector construction; integrate the above various features into a comprehensive multi-dimensional vector as the input of the deep learning model.

[0053] Preferably, the deep learning model consists of four modules: a data input layer, a structured feature processing module, a spatiotemporal matrix processing module, a multimodal fusion module, and a dual-task output module. The structured feature processing module integrates multi-source heterogeneous road and traffic information, and includes a static traffic feature encoding unit, a dynamic feature temporal processing unit, an environmental factor encoding unit, and a feature fusion unit. The road and traffic information includes inherent static attributes of the road and dynamic traffic data. The inherent static attributes of the road have a total of 12 dimensions, including road grade, number of lanes, design maximum speed, road type, and geometric parameters describing slope and curve curvature. The dynamic traffic data includes 8 dimensions, mainly covering traffic flow indicators and statistics such as the average speed and speed standard deviation of three main vehicle types. The inherent static attributes of the road first enter the static traffic feature encoding unit through the data input layer. This unit uses a 64-dimensional embedding layer to map the original low-dimensional or categorical features to a high-dimensional dense vector space to capture their potential semantic information. Meanwhile, dynamic traffic data is fed into the dynamic feature temporal processing unit through the data input layer. The core of this unit is a Long Short-Term Memory (LSTM) network with 32 memory units, specifically designed to capture complex sequence patterns and dependencies in traffic flow state over time. Furthermore, the 4-dimensional environmental parameters input through the data input layer are received by the environmental factor encoding unit and undergo nonlinear transformation and feature abstraction through a fully connected layer with 16 neurons. Subsequently, the feature fusion unit concatenates the 64-dimensional static embedded features, the 32-dimensional temporal features extracted by the LSTM, and the 16-dimensional environmental encoding features along the feature dimensions to form a 112-dimensional joint feature vector. Finally, a fully connected layer equipped with a ReLU activation function further compresses and fuses the 112-dimensional joint feature vector, outputting a 128-dimensional, highly structured comprehensive traffic feature representation. The environmental parameters include weather condition codes, visibility levels, ambient temperature, and ambient humidity.

[0054] Preferably, the spatiotemporal matrix processing module is used to process high-dimensional, gridded spatiotemporal traffic state data. The spatiotemporal matrix processing module includes a three-dimensional convolutional unit, a spatiotemporal attention layer, and a feature compression unit. The data input layer receives a spatiotemporal tensor of dimension T×100×100×3, which corresponds to T consecutive time frames, a 100m×100m spatial grid, and spatiotemporal tensors of three vehicle type codes. The spatiotemporal tensor is first input to the three-dimensional convolutional unit, which applies 32 three-dimensional convolutional kernels of size 3×5×5 to simultaneously perform sliding convolution operations in the temporal and spatial dimensions to extract local... The spatiotemporal neighborhood features are used to generate a tensor containing 32 feature channels. The spatiotemporal attention layer has 4 attention heads, which learn and calculate in parallel the inter-frame correlation weights of the feature tensor in the temporal axis and the inter-grid correlation weights in the spatial plane. These weights are used to enhance the features, i.e., weighted aggregation. Finally, the feature compression unit uses global spatial pooling to aggregate the feature tensors, which have been enhanced by 3D convolution and attention and still maintain the spatiotemporal grid structure, into a fixed-length, 256-dimensional global spatiotemporal feature vector. The 3×5×5 3D convolution kernel is a 3-frame depth-oriented and 5x5 grid 3D convolution kernel.

[0055] Preferably, the multimodal fusion module is used to deeply fuse the structured attributes output by the structured feature processing module and the deep features of the spatiotemporal grid output by the spatiotemporal matrix processing module. The multimodal fusion module includes a cross-network unit and a weighted fusion unit. First, the cross-network unit receives a 128-dimensional feature vector from the structured feature processing module. The cross-network unit explicitly calculates the cross product between features through three layers of dedicated cross operations, thereby generating a higher-order combination of features and outputting a 128-dimensional vector containing rich feature interaction information. Subsequently, the weighted fusion unit performs weighted concatenation of the cross-processed 128-dimensional structured features and the 256-dimensional spatiotemporal features output by the spatiotemporal matrix processing module. Finally, a unified 384-dimensional feature vector that fuses all modal information is generated. Here, the 128-dimensional structured features and the 256-dimensional spatiotemporal features are assigned different importance weights, with the 256-dimensional spatiotemporal features having a weight of 2 and the 128-dimensional structured features having a weight of 1.

[0056] Preferably, the dual-task output module simultaneously performs two prediction tasks based on the fused 384-dimensional feature vector, including a classification head and a regression head. The classification head consists of a fully connected layer containing only a single neuron, and a sigmoid activation function is applied to its output to map the neuron's output value to a range of 0 to 1. The output value of the neuron is the predicted probability of the vehicle engaging in "speeding behavior" in the current spatiotemporal environment. In parallel, the regression head also consists of a fully connected layer containing a single neuron, but uses a linear activation function, i.e., without additional transformation. Its output value is directly used as the continuous value of the predicted "risk index" to quantitatively assess the potential traffic safety risk level in the current scenario. The outputs of the classification head and the regression head together constitute the model's comprehensive assessment result of traffic scenario safety.

[0057] Beneficial effects

[0058] Compared with the prior art, the beneficial effects of this invention are as follows: This invention achieves dynamic discrimination of dangerous lane-changing behavior of motor vehicles through a deep learning discrimination method based on trajectory data mining. This method breaks through the limitations of traditional passive post-event management, can proactively and timely detect dangerous lane-changing behavior, significantly improve the monitoring coverage of dangerous lane-changing behavior on roads, and thus effectively reduce the risk of related accidents. Attached Figure Description

[0059] Figure 1 This is a flowchart illustrating a deep learning method for identifying dangerous lane-changing behaviors of motor vehicles according to the present invention.

[0060] Figure 2 This is a schematic diagram illustrating the analysis and determination of vehicle lane-changing behavior based on trajectory data, which is a deep learning method for discriminating dangerous lane-changing behavior of motor vehicles according to the present invention.

[0061] Figure 3 This is a schematic diagram of a specific trajectory point neighborhood historical trajectory data set for a deep learning method for identifying dangerous lane-changing behavior of motor vehicles according to the present invention.

[0062] Figure 4 This is a schematic diagram of the accident location image processing in a deep learning method for identifying dangerous lane-changing behavior of motor vehicles according to the present invention.

[0063] Figure 5 This is a technical roadmap of a deep learning model for a deep learning-based method for identifying dangerous lane-changing behaviors of motor vehicles, as described in this invention. Detailed Implementation

[0064] To make the objectives, technical solutions, and advantages of the present invention clearer and more complete, the present invention will be further described in detail below with reference to the embodiments. Obviously, the embodiments described below are some embodiments of the present invention, but the scope of protection claimed by the present invention is not limited to the specific embodiments below.

[0065] like Figure 1 As shown, a deep learning method for identifying dangerous lane-changing behavior of motor vehicles includes the following steps:

[0066] S1. Construction based on the historical database of dangerous lane change behavior: By accessing the event record information of the road traffic accident handling information system, a historical data set containing dangerous lane change events is established as the basis for subsequent model training;

[0067] S2. Extraction of Trajectory Characteristics and Environmental Features of Dangerous Lane Changes by Motor Vehicles

[0068] S2-1. Extraction of Dangerous Lane Change Events Based on Trajectory Spatiotemporal Matching Analysis: Collect vehicle trajectory data within a specific time period (e.g., 5 minutes before and after) and a 300-meter spatial range around the accident site to construct a neighborhood trajectory database for the accident location; calculate the lane change index Sft based on the vertical distance between the trajectory points and the road alignment; if the lane change index Sft is greater than a preset threshold of 3 meters, it is determined to be a dangerous lane change accident; transform the vehicle trajectories around the accident site into a discretized spatiotemporal matrix (M(t)) to simulate a continuous dynamic image of vehicle distribution before the accident, and achieve a continuous representation of vehicle motion state by completing the neighborhood vehicle trajectories second by second, i.e., generating spatiotemporally continuous image data; the specific steps are as follows:

[0069] S2-1-1. For the set of vehicle trajectories A_Cars(A,C,G) that can match accident records, where A is the list of corresponding accident records, C is the list of vehicles that match the accident, and G is the set of matching vehicle trajectory data, iterate through each accident record a, and extract the corresponding matching vehicle trajectory data g(a,c) from the matching vehicle c of a within a range of 5 minutes before and after the accident record time.

[0070] S2-1-2. Iterate through each trajectory point of g(a,c) sequentially, and obtain the vehicle speed closest to the time of the accident record as the zero time t_accident_real, which is the precise time of the accident. Five minutes before this time is the corresponding adjacent time period t_seg, and the latitude and longitude coordinates c_accident_real corresponding to this time are the precise location of the accident.

[0071] S2-1-3, Further extract the list of trajectory points within 300 meters of c_accident_real: Tr_pre(a,c)={p1,p2,...,p n}, pn That is, point c_accident_real. The lane change index Sft(a,c) is calculated based on the status of each trajectory point in this list and the spatial relationship with the adjacent road network; for example... Figure 2 and Figure 3 As shown, the calculation method for the lane change index Sft is as follows:

[0072] (1) Traverse each trajectory point p1~p n Calculate the forward direction p i →p i+1 The angle between the search area and adjacent road segments (search range is 80 meters, i.e.) Figure 2 In diagrams AB and CD, the number of adjacent road segments corresponds to the number of included angles. Road segments with included angles less than 90 degrees are considered to be in the same direction. The road segment with the smallest included angle among adjacent road segments in the same direction is selected as the matching driving segment (i.e., segment AB in the diagram), and this included angle is β. i ;

[0073] (2) For the current point p i Further, the system searches the overall trajectory database to select a set of historical vehicle trajectory points within an 80-meter radius that travel in the same direction (meeting the angle between the travel direction and AB is less than 10 degrees and the speed is greater than 10 km / h) HP = {hp1, hp2, ... hp}. m}, and calculate p i The distance perpendicular to AB from each point along the AB direction is calculated to obtain the maximum counterclockwise vertical distance dmax1. i The maximum vertical distance dmax2 clockwise i ;

[0074] (3) Calculate the lane change index Sft based on the perpendicular distance between the trajectory point and the road alignment. The formula for the lane change index Sft is as follows:

[0075]

[0076] That is, from point p1 to p n Cumulative lateral displacement;

[0077] If Sft(a,c)>3 meters, it is considered that the vehicle has made a lateral lane change during the stage of the accident, and the accident is thus identified as a dangerous lane change accident.

[0078] By traversing A from S2-1-1 to S2-1-3, we can obtain the precise time set A_T of dangerous lane change accidents, the corresponding time period set A_T_Seg, and the location set A_C.

[0079] S2-1-4. For the complete vehicle trajectory set T_Cars(C_total, G_total), where C_total is the list of all vehicles and G_total is the list of all vehicle trajectory data, iterate through each vehicle c in C_total. For its trajectory set g(c), obtain the subset of trajectory points g_S(c_A_m) that is consistent with the time range of A_T_seg, where A_m is the subset of the accident record list that is consistent with the time range of A_T_seg for vehicle c. Suppose that there is an accident a_m in A_T_seg that occurred between 14:30:14 and 14:35:14 on 2019-08-18, and vehicle c has a trajectory record set g'(c) during this time period, then add this set to g_S(c, A_m) and add the accident a_m record to A_m.

[0080] S2-1-5. If g_S(c,A_m) is an empty set, then jump to the next vehicle; otherwise, traverse each accident record a_m in A_m and search within a 500-meter radius of the precise location c_accident_real_m of a_m. If the corresponding vehicle trajectory record set g'(c) has trajectory data in the same direction of travel and with an angle of less than 90 degrees within this range, then it can be considered that the trajectory data set g(c,a_m) within the range of the vehicle has a spatiotemporal matching relationship with the accident record a_m.

[0081] After traversing T_Cars(C_total,G_total) through S2-1-4 to S2-1-5, we can obtain the spatiotemporal matching vehicle set M_C, the matching trajectory set M_G, and the neighborhood spatiotemporal continuous image set STG for all accidents.

[0082] like Figure 4 As shown, the trajectory data is processed into continuous spatiotemporal imagery in the following manner:

[0083] (1) Draw a circle with the accident point c_accident_real_m as the center and a radius of 500 meters. Divide the circle vertically into 100-meter ranges above and below the vehicle's direction of travel 100 meters before the accident. Finally, divide the divided area into discrete matrix spatial regions Z with 1-meter intervals. For any time t, each square in the region corresponds to a vector (x,y,m), where x and y are coordinates and m is the vehicle model number corresponding to the trajectory point in the square at the current time. If there is no trajectory point in the square, then m is 0.

[0084] (2) Neighboring vehicle trajectory completion; for the trajectory set g(c,a_m) of a vehicle c that spatially matches the accident point during the time period of accident a_m (i.e., the 5-minute segment before the accident), trajectory completion is performed second by second; Figure 4Taking the trajectory in the image as an example, assuming the first trajectory point p1 corresponds to time t1 and the second trajectory point p2 corresponds to time t2, then the line connecting p1 to p2 is evenly divided according to the interval of (t2-t1-1) seconds. For the last point p4 of the trajectory, the forward distance is calculated second by second according to the travel speed of p4 and the path points are completed according to the extension line from p3 to p4. The continuous trajectory g'(c,a_m) starting from p1 is obtained second by second, and the intersection of the continuous trajectory and Z is obtained to obtain the grid number corresponding to each continuous trajectory point.

[0085] (3) Traverse all neighboring vehicles c (including the accident vehicle) at the accident point to obtain the corresponding complete continuous trajectory. At the same time, select the first trajectory point t_first of the vehicle that appears last during the time period as the start time of generating the spatiotemporal image of the neighboring vehicles.

[0086] (4) Generation of spatiotemporally continuous image data of neighboring vehicles; starting from t_first to the time of the accident, the regional Z vehicle distribution matrix M(t) is generated second by second, as follows:

[0087]

[0088] Where m(P,Q) corresponds to the vehicle model number of the current trajectory point in the corresponding square, and is 0 if there is no trajectory point.

[0089] S2-2. Risk of Speeding Events Identification and Risk of Speeding Index Calculation: Identify whether the vehicle exhibited significant speeding behavior before the accident occurred, and calculate the corresponding "Risk of Speeding Index," which comprehensively considers both the severity of the accident and the relative degree of speeding. Specific steps are as follows:

[0090] S2-2-1. Traverse the estimated vehicle speed set A_V before the accident and record each vehicle speed v_accident. If the spatiotemporal matching vehicle set M_C(a) corresponding to the accident a is not empty, calculate the neighborhood average vehicle speed avg_v(a) and root variance var_v(a) of M_V(a). If v_accident>(avg_v(a)+var_v(a)), then define the accident a as a high-risk speeding accident.

[0091] S2-2-2. For risky overspeed accident a, calculate the risky overspeed index. The formula for calculating the risky overspeed index is as follows:

[0092] indi_v(a)=s(a)*avg_v(a) / (v_accident-avg_v(a)),

[0093] Wherein, s(a) is the severity parameter of accident a, with a value of 1 for "simple" accidents, a value of 5 for "general" accidents, and a value of 20 for accidents above "general" level; in addition, (v_accident-avg_v(a)) is placed in the numerator to indicate that when an overspeeding behavior occurs, the smaller the relative degree of overspeeding, the greater the severity of the accident, and the higher the corresponding risk overspeeding index.

[0094] S2-2-3. Obtain the complete set of risk overspeed events O_A through S2-2-1 and S2-2-2.

[0095] S2-3, Construction of Multi-Dimensional Feature Vectors for Risky Speeding Events: By traversing the set of risky speeding events O_A, for the current event a, the event features are analyzed and multi-dimensional vectors are constructed from the following five perspectives, as shown in Table 1.

[0096] (1) Static traffic characteristics

[0097] The fixed attributes of the accident section are extracted, including the attribute parameters of neighboring roads and the geometric parameters of the neighboring roads at the precise location c_accident_real of event a. The attribute parameters of neighboring roads include parameter level, number of lanes, design speed, and road type, which includes bridges, tunnels, highways, and urban roads. The attribute parameters of neighboring roads are obtained by searching the corresponding GIS map data attributes in the vicinity. The geometric parameters of the neighboring roads include slope, radii, and turning radius, which are calculated based on the trajectory location and elevation information.

[0098] (2) Dynamic traffic characteristics

[0099] Traffic flow data for the time period of the accident is collected, including various statistical features of the matching vehicle set M_C(a) and trajectory set M_G(a) corresponding to event a. These statistical features include traffic flow levels and a list of vehicle types and corresponding average speeds (e.g., {C1: private car 54.6 km / h, C2: large truck 47.1 km / h, C3: commercial passenger vehicle 50.3 km / h}). The traffic flow levels are represented as follows:

[0100] vol(a) = num(M_C(a)) * θ,

[0101] Where num(M_C(a)) is the number of vehicles in M_C(a); θ is the average sampling rate, which is calculated by the ratio of traffic flow at actual observation points in the region (such as traffic flow observation data from various checkpoints, highway entrances and exits, traffic volume observation stations, etc.) to the number of vehicles in the corresponding matching trajectory; the number of vehicles in the trajectory is the number of vehicles corresponding to the spatiotemporal matching data of the observation point during the observation period and the entire vehicle trajectory set T_Cars(C_total,G_total);

[0102] (3) Environmental factors affecting the event

[0103] The environmental factors affecting the event include environmental information such as weather, visibility, temperature, humidity, and road construction conditions recorded at the time of the accident.

[0104] (4) Overspeed trajectory attributes and behavioral characteristics;

[0105] The speeding trajectory attributes and behavioral characteristics include the type of vehicle involved in the accident, the continuous driving time before the accident (obtained by statistically analyzing the continuous driving trajectory time of the vehicle before the accident, with the interruption interval determined as 30 minutes), the driver's attributes (age and gender), and the list of observed vehicle speeds corresponding to each trajectory point within 100 to 300 meters before the accident, Tr-V_pre(a,c), such as {65.1km / h, 67.2km / h...};

[0106] (5) Construction of multi-dimensional feature vectors for risk overspeed;

[0107] The aforementioned features are integrated into a comprehensive multi-dimensional vector, which is then used as the input to the deep learning model. See the table below:

[0108]

[0109]

[0110] S3. Construction of a deep learning model for dangerous lane-changing behavior based on multi-dimensional features: Using the multi-dimensional feature vectors constructed above, a dual-task deep learning prediction model for vehicle speeding risk is trained. By jointly processing structured traffic features and spatiotemporal matrix data, the synchronous output of speeding behavior classification and risk index regression is achieved.

[0111] like Figure 5As shown, specifically, the deep learning model consists of four modules: a data input layer, a structured feature processing module, a spatiotemporal matrix processing module, a multimodal fusion module, and a dual-task output module. The structured feature processing module integrates multi-source heterogeneous road and traffic information, and includes a static traffic feature encoding unit, a dynamic feature temporal processing unit, an environmental factor encoding unit, and a feature fusion unit. The road and traffic information includes inherent static attributes of the road and dynamic traffic data. The inherent static attributes of the road have a total of 12 dimensions, including road grade, number of lanes, design maximum speed, road type, and geometric parameters describing slope and curve curvature. The dynamic traffic data includes 8... The system comprises several dimensions, primarily covering traffic flow indicators and statistical measures such as the average speed and speed standard deviation of three main vehicle types. The spatiotemporal matrix processing module is used to process high-dimensional, gridded spatiotemporal traffic state data, and includes a 3D convolutional unit, a spatiotemporal attention layer, and a feature compression unit. The multimodal fusion module is used to deeply fuse the structured attributes output by the structured feature processing module and the deep features of the spatiotemporal grid output by the spatiotemporal matrix processing module. The multimodal fusion module includes a cross-network unit and a weighted fusion unit. The dual-task output module simultaneously executes two prediction tasks based on the fused 384-dimensional feature vector, including a classification head and a regression head.

[0112] like Figure 5 As shown, the working process of the deep learning model for dangerous lane-changing behavior based on multi-dimensional features is as follows:

[0113] (1) The inherent static attributes of the road first enter the static traffic feature encoding unit through the data input layer. This unit uses a 64-dimensional embedding layer to map the original low-dimensional or categorical features to a high-dimensional dense vector space to capture their potential semantic information. At the same time, dynamic traffic data is sent to the dynamic feature temporal processing unit through the data input layer. The core of the dynamic feature temporal processing unit is a Long Short-Term Memory (LSTM) network with 32 memory units, which is specifically used to capture the complex sequence patterns and dependencies of traffic flow state evolution over time. In addition, the 4-dimensional environment data input through the data input layer... Parameters (such as weather condition codes, visibility levels, ambient temperature, and ambient humidity) are received by the environmental factor encoding unit and nonlinearly transformed through a fully connected layer with 16 neurons. Subsequently, the feature fusion unit concatenates the three deep-processed 64-dimensional static embedded features, the 32-dimensional temporal features extracted by LSTM, and the 16-dimensional environmental coding features along the feature dimensions to form a 112-dimensional joint feature vector. Finally, a fully connected layer equipped with a ReLU activation function further compresses the 112-dimensional joint feature vector and enhances it to a 128-dimensional comprehensive traffic feature representation.

[0114] (2) The data input layer receives a spatiotemporal tensor with dimensions T×100×100×3 (i.e., corresponding to T consecutive time frames, a 100m×100m spatial grid, and 3 types of vehicle codes); the spatiotemporal tensor is first input into a three-dimensional convolutional unit, which applies 32 three-dimensional convolutional kernels of size 3×5×5 (i.e., 3 frames in the depth direction and 5x5 grid in the spatial direction) to simultaneously perform sliding convolution operations in the time and spatial dimensions, extracting pattern features in the local spatiotemporal neighborhood, and generating a tensor containing 32 feature channels; the spatiotemporal attention layer has 4 attention heads, which learn and calculate the inter-frame correlation weights of the feature tensor in the time axis direction and the inter-grid correlation weights in the spatial plane in parallel, and use these weights to perform weighted aggregation of features; finally, the feature compression unit adopts a global spatial pooling operation to aggregate the spatiotemporal grid structure feature tensor after three-dimensional convolution and attention enhancement into a 256-dimensional global spatiotemporal feature vector.

[0115] (3) The cross-network unit receives a 128-dimensional feature vector from the structured feature processing module; the cross-network unit explicitly calculates the cross product between features through three layers of dedicated cross operations, thereby generating a high-order combination of features and outputting a 128-dimensional vector containing rich feature interaction information; subsequently, the weighted fusion unit performs weighted concatenation of the 128-dimensional structured features after cross processing and the 256-dimensional spatiotemporal features output by the spatiotemporal matrix processing module; finally, a unified feature vector of 384 dimensions that integrates all modal information is generated; the key setting here is that the two features are assigned different importance weights, with the 256-dimensional spatiotemporal feature weight being 2 and the 128-dimensional structured feature weight being 1.

[0116] (4) The classification head consists of a fully connected layer containing only a single neuron, and the Sigmoid activation function is applied to its output to map the output value of the neuron to the range of 0-1. The output value of the neuron is directly interpreted as the probability prediction value of the vehicle's "speeding behavior" in the current spatiotemporal environment. In parallel, the regression head consists of a single neuron using a linear activation function, and its output value is directly used as the continuous value of the predicted "risk index" to quantify and assess the degree of potential traffic safety risks in the current scenario. The outputs of the classification head and the regression head together constitute the model's comprehensive assessment result of traffic scenario safety.

[0117] S4. Identification and alarm of dangerous lane change behavior based on dynamic trajectory: The real-time acquired vehicle trajectory data and related environmental information are input into the pre-trained deep learning model. The deep learning model analyzes and judges the real-time data. When dangerous lane change behavior is identified, an alarm is triggered immediately to notify relevant regulatory departments or take intervention measures.

[0118] This invention achieves dynamic identification of dangerous lane-changing behaviors of motor vehicles through a deep learning-based discrimination method based on trajectory data mining. This method breaks through the limitations of traditional passive post-event management, and can proactively and promptly detect dangerous lane-changing behaviors, significantly improving the monitoring coverage of dangerous lane-changing behaviors on roads, thereby effectively reducing the risk of related accidents.

[0119] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions within the technical scope disclosed in this application should be included within the scope of protection of this application. Based on the disclosure and teachings of the above specification, those skilled in the art can also make changes and modifications to the above embodiments. Therefore, this invention is not limited to the specific embodiments disclosed and described above, and some modifications and changes to the invention should also fall within the scope of protection of the claims of this invention. Furthermore, although some specific terms are used in this specification, these terms are only for convenience of explanation and do not constitute any limitation on this invention.

Claims

1. A deep learning discrimination method for dangerous lane-changing behavior of a motor vehicle, characterized in that, The method comprises the following steps: S1, constructing a dangerous lane-changing behavior history database based on the database: by accessing the event record information of the road traffic accident handling information system, a historical data set containing dangerous lane-changing events is established as the basis for subsequent model training; S2, extracting the trajectory features and environmental features of the dangerous lane-changing behavior of the motor vehicle: collecting the vehicle trajectory data of the accident, calculating the lane-changing index according to the vertical distance between the trajectory point and the road alignment, and performing time-space continuous image processing on the vehicle trajectory data; identifying whether the vehicle has a speeding behavior, calculating the corresponding risk speeding index, and constructing a risk speeding event multi-dimensional feature vector from multiple angles; S3, constructing a dangerous lane-changing behavior deep learning model based on multi-dimensional features: using the multi-dimensional feature vector constructed above, a vehicle speeding risk double-task deep learning prediction model is trained, which realizes the synchronous output of speeding behavior classification and risk index regression by jointly processing structured traffic features and time-space matrix data; S4, dangerous lane-changing behavior identification and alarm based on dynamic trajectory: input the real-time acquired vehicle trajectory data and related environmental information into the trained deep learning model, which analyzes and discriminates the real-time data, and triggers an alarm when a dangerous lane-changing behavior is identified, notifying the relevant regulatory departments or taking intervention measures. 2.The method of claim 1, wherein the method further comprises: The S2 specifically comprises the following steps: S2-1, dangerous lane-changing event extraction based on trajectory time-space matching analysis: collecting vehicle trajectory data within a specific time period before and after the accident and within a 300-meter space range of the accident point, constructing an accident site neighborhood trajectory database; calculating the lane-changing index Sft according to the vertical distance between the trajectory point and the road alignment; if the lane-changing index Sft is greater than the preset threshold value of 3 meters, it is determined as a dangerous lane-changing accident; the vehicle trajectory around the accident point is converted into a discretized time-space matrix M(t), simulating a continuous vehicle distribution dynamic image before the accident, and the vehicle motion state is represented continuously by completing the neighborhood vehicle trajectory second by second, i.e. generating time-space continuous image data; S2-2, risk speeding event identification and risk speeding index calculation: identifying whether the vehicle has a significant speeding behavior before the accident, and calculating the corresponding "risk speeding index", which considers the accident severity and relative speeding degree; S2-3, risk speeding event multi-dimensional feature vector construction: traversing the risk speeding event set, analyzing the event features and constructing the multi-dimensional vector from five angles, i.e. static traffic features, dynamic traffic features, event environmental impact factors, speeding trajectory attributes and behavior features, and risk speeding multi-dimensional feature vector construction. 3.The method of claim 2, wherein: The specific steps of S2-1 are as follows: S2-1-1, for the vehicle trajectory set A_Cars(A,C,G) that can match the accident record, where A is the corresponding accident record list, C is the accident matching vehicle list, and G is the matching vehicle trajectory data set, traverse each accident record a, and extract the corresponding matching vehicle trajectory data g(a,c) from the accident record time before and after 5 minutes. S2-1-2, sequentially traverse each trajectory point of g(a, c) to obtain the time t_accident_real when the vehicle speed is zero closest to the time when the accident record occurs as the accurate time when the accident occurs, the time 5 minutes before the time is the corresponding accident occurrence adjacent time period t_seg, and the coordinates c_accident_real corresponding to the time are taken as the accurate accident occurrence position; S2-1-3, further extract the trajectory to reach the c_accident_real point to 300 meters within the trajectory point list Tr_pre(a,c) = {p1, p2,..., p n}, p n That is the c_accident_real point, according to the list of each trajectory point state and adjacent road network spatial relationship to calculate the lane change index Sft(a,c); After traversing A through S2-1-1 to S2-1-3, the accurate accident occurrence time set A_T of the dangerous lane-changing accident, the corresponding occurrence time period set A_T_Seg, and the position set A_C can be obtained; S2-1-4, traverse each vehicle c in the vehicle list C_total and the trajectory data list G_total of all vehicles T_Cars(C_total, G_total) to obtain the trajectory point subset g_S(c,A_m) consistent with the time range of A_T_seg, wherein A_m is the subset of the corresponding accident record list of vehicle c consistent with the time range of A_T_seg. Assuming that the occurrence time period of a_m in A_T_seg is 2019-08-1814:30:14-14:35:14, and vehicle c has a trajectory record set g'(c) in the time period, the set is added to g_S(c,A_m), and the accident a_m record is added to A_m; S2-1-5, if g_S(c,A_m) is an empty set, jump to the next vehicle, otherwise traverse each accident record a_m in A_m, search the accurate occurrence position c_accident_real_m of a_m within the adjacent 500-meter range, if the corresponding vehicle trajectory record set g'(c) exists within the range and the heading direction is the same and the forward direction angle is less than 90 degrees, it is considered that the trajectory data set g(c,a_m) of the vehicle within the range has a spatiotemporal matching relationship with the accident record a_m; After traversing T_Cars(C_total, G_total) through S2-1-4 to S2-1-5, the corresponding spatiotemporal matching vehicle set M_C, the matching trajectory set M_G, and the neighborhood spatiotemporal continuous image set STG of all accidents can be obtained; The calculation method of the lane-changing index Sft is as follows: (1) Traverse each trajectory point p1~p n , calculate the angle between the forward direction p i →p i+1 and the adjacent road section of the road network, and record the road section with an angle less than 90 degrees as the same direction; The road section with the minimum included angle with the adjacent road section is selected as the matching driving road section, and the included angle is β i The search range of the road section is 80 meters, and there are several included angles with several adjacent road sections. (2) to the current point p i , further search from the total trajectory database to select the adjacent 80-meter range historical passing vehicle same direction trajectory point set HP = {hp1, hp2,... hp m}, and calculate the vertical direction distance of p i to each point, so as to obtain the counterclockwise maximum vertical distance dmax1 i , and the clockwise maximum vertical distance dmax2 i ; (3) According to the vertical distance between the trajectory point and the road alignment, the lane-changing index Sft is calculated, and the formula of the lane-changing index Sft is as follows: i.e. from point p1 to p n cumulative lateral displacement level; If Sft(a, c) > 3 meters, it is considered that the vehicle has a lateral lane-changing behavior during the accident occurrence stage, and the accident is identified as a dangerous lane-changing accident.

4. The method of claim 3, wherein the method further comprises: In S2-1, the trajectory data is processed by continuous spatiotemporal imaging in the following manner: (1) Taking the accident occurrence point c_accident_real_m as the center and 500 meters as the radius, a circle is drawn, and the circle is divided vertically up and down 100 meters in the range of 100 meters before the accident in the vehicle driving direction, and finally the divided area is divided at intervals of 1 meter to obtain a discrete matrix space area Z. For any time t, each grid in the area corresponds to a vector (x, y, m), x and y are coordinates, and m is the vehicle type number corresponding to the trajectory point in the grid at the current time. If there is no trajectory point in the grid, m is 0; (2) Neighbor vehicle trajectory completion; for the trajectory set g(c, a_m) of a certain vehicle c that matches the accident point space at the accident a_m occurrence period, the trajectory is completed second by second; assuming that the first trajectory point p1 corresponds to time t1, and the second trajectory point p2 corresponds to time t2, then the line connecting p1 and p2 is divided at intervals of (t2-t1-1) seconds, and for the last trajectory point p4, the forward distance is calculated second by second according to the driving speed of p4, and the path point is completed according to the extension line from p3 to p4. The continuous trajectory g'(c, a_m) starting from p1 is obtained, and the continuous trajectory and Z are intersected to obtain the grid number corresponding to each continuous trajectory point; the accident a_m occurrence period refers to a 5-minute segment before the accident occurrence time; (3) Traversing each neighbor vehicle c of the accident point, including the accident vehicle, to obtain the corresponding completed continuous trajectory, and selecting the first trajectory point time t_first of the latest vehicle appearing in the occurrence period as the starting generation time of the neighbor vehicle space-time image; (4) Neighbor vehicle space-time continuous image data generation; starting from t_first to the accident occurrence time, a vehicle distribution matrix M(t) of the area Z is generated second by second, which is represented as follows: Wherein, m(P, Q) is the vehicle type number to which the current trajectory point in the grid belongs, and 0 if there is no trajectory point.

5. The method of claim 4, wherein the method further comprises: The specific steps of S2-2 are as follows: S2-2-1, traversing each recorded vehicle speed v_accident in the estimated driving vehicle speed set A_V, if the time-space matching vehicle set M_C(a) of the accident a is not an empty set, then the average speed avg_v(a) of M_V(a) and the root variance var_v(a) are calculated, if v_accident>(avg_v(a)+var_v(a)), then the accident a is defined as a risk speeding accident; S2-2-2, for the risk speeding accident a, the risk speeding index is calculated, and the risk speeding index calculation formula is as follows: indi_v(a)=s(a)*avg_v(a) / (v_accident-avg_v(a)), Wherein, s(a) is the severity parameter of the accident a, the value of "simple" accident is 1, the value of "general" accident is 5, and the value of "general" accident above is 20; in addition, (v_accident-avg_v(a)) is in the numerator, when the speeding behavior occurs, the smaller the relative speeding degree, the greater the severity of the accident, and the higher the risk speeding index. S2-2-3, obtaining the whole risk speeding event set O_A through S2-2-1 and S2-2-2. 6.The method of claim 5, wherein the method further comprises: In the S2-3, through traversing the risk speeding event set O_A, the event feature analysis and multi-dimensional vector construction are performed from the following five aspects for the current event a: (1) static traffic feature; The fixed attributes of the accident road section are extracted, including the adjacent road attribute parameters and the neighborhood road line geometry parameters of the event a precise position c_accident_real; the adjacent road attribute parameters include parameter level, lane number, design speed, and road type, and the road type includes bridge, tunnel, highway, and urban road; the adjacent road attribute parameters are obtained by searching the adjacent corresponding GIS map data attributes; the neighborhood road line geometry parameters include slope, curvature, and turning radius, which are calculated according to the trajectory position and elevation information; (2) dynamic traffic feature; The traffic flow data at the time of the accident is collected, including various statistical features of the event a corresponding matching vehicle set M_C(a) and trajectory set M_G(a), including traffic flow level and corresponding average vehicle speed list of vehicle type; the traffic flow level is represented as follows: vol(a)=num(M_C(a))*θ, wherein num(M_C(a)) is the number of vehicles in M_C(a); θ is the average sampling rate, which is calculated by the ratio of the actual observation point traffic flow and the corresponding matching trajectory vehicle number; the trajectory vehicle number is the corresponding vehicle number of the observation point observation period and the whole vehicle trajectory set T_Cars(C_total, G_total) spatio-temporal matching data; (3) event environment influence factor; the event environment influence factor includes weather, visibility, temperature, humidity, and road construction conditions recorded at the time of the accident event; (4) speeding trajectory attribute and behavior feature; the speeding trajectory attribute and behavior feature includes accident vehicle type, continuous driving time before the accident, driver attribute, and observation speed list Tr-V_pre(a, c) of each trajectory point within 100-300 meters before the accident; the continuous driving time before the accident is obtained by counting the continuous driving trajectory time of the vehicle before the accident, and the interruption driving interval judgment time is 30 minutes; the driver attribute at least includes age and gender; (5) risk speeding multi-dimensional feature vector construction; the various features in (1)-(4) are integrated into a comprehensive multi-dimensional vector, which is used as the input of the deep learning model.

7. The method of claim 6, wherein the method further comprises: The deep learning model is composed of four modules, including a data input layer, a structured feature processing module, a spatio-temporal matrix processing module, a multi-modal fusion module, and a double-task output module; the structured feature processing module is used for integrating multi-source heterogeneous road and traffic information, and includes a static traffic feature coding unit, a dynamic feature time series processing unit, an environmental factor coding unit, and a feature fusion unit; the road and traffic information includes inherent static attributes of the road and dynamic traffic data; the inherent static attributes of the road include 12 dimensions, including road grade, lane number, designed maximum speed, road type, and geometric parameters describing slope and curve radian; the dynamic traffic data includes 8 dimensions, mainly covering traffic flow indicators and statistical quantities such as average speed and speed standard deviation of three main vehicle types; the inherent static attributes of the road are first input into the static traffic feature coding unit through the data input layer, and the unit uses a 64-dimensional embedding layer to map the original low-dimensional or categorical features to a high-dimensional dense vector space to capture the potential semantic information; at the same time, the dynamic traffic data is sent to the dynamic feature time series processing unit through the data input layer, and the core of the dynamic feature time series processing unit is a long short-term memory network (LSTM) with 32 memory units, which is specially used to capture the complex sequence patterns and dependencies of the evolution of traffic flow state over time; in addition, the 4-dimensional environmental parameters input through the data input layer are received by the environmental factor coding unit and are subjected to nonlinear transformation and feature abstraction through a fully connected layer with 16 neurons; then, the feature fusion unit splices the 64-dimensional static embedding features, the 32-dimensional time series features extracted by the LSTM, and the 16-dimensional environmental coding features along the feature dimension to form a 112-dimensional joint feature vector; finally, a fully connected layer with a ReLU activation function further compresses, fuses and elevates the 112-dimensional joint feature vector to a 128-dimensional, highly structured comprehensive traffic feature representation output; the environmental parameters include weather condition code, visibility level, environmental temperature and environmental humidity. 8.The method of claim 7, wherein the method further comprises: The spatio-temporal matrix processing module is used for processing high-dimensional and gridded spatio-temporal traffic state data, and comprises a three-dimensional convolution unit, a spatio-temporal attention layer and a feature compression unit; the data input layer receives a spatio-temporal tensor with a dimension of T x 100 x 100 x 3, wherein the spatio-temporal tensor with a dimension of T x 100 x 100 x 3 corresponds to a spatio-temporal tensor of T continuous time frames, 100 m x 100 m space grids and 3 types of vehicle encodings; the spatio-temporal tensor is first input to the three-dimensional convolution unit, the three-dimensional convolution unit applies 32 three-dimensional convolution kernels with a size of 3 x 5 x 5 to simultaneously perform sliding convolution operation in the time and space dimensions, extracts pattern features in the local spatio-temporal neighborhood, and generates a tensor containing 32 feature channels; the spatio-temporal attention layer has 4 attention heads, which learn and calculate the inter-frame correlation weight of the feature tensor in the time axis direction and the inter-grid correlation weight in the spatial plane in parallel, and enhance the features by using these weights, i.e., weighted aggregation; finally, the feature compression unit adopts a global spatial pooling operation to aggregate the feature tensor, which still maintains the spatio-temporal grid structure after three-dimensional convolution and attention enhancement, into a fixed-length global spatio-temporal feature vector with a dimension of 256; the three-dimensional convolution kernel with a size of 3 x 5 x 5 is a three-dimensional convolution kernel with a depth direction of 3 frames, a spatial direction of 5 x 5 grids. 9.The method of claim 8, wherein: The multi-modal fusion module is used for deep fusion of the structured attributes output by the structured feature processing module and the deep features of the spatio-temporal grid output by the spatio-temporal matrix processing module; the multi-modal fusion module comprises a cross-network unit and a weighted fusion unit; first, the cross-network unit receives a 128-dimensional feature vector from the structured feature processing module; the cross-network unit explicitly calculates the cross product between the features through 3 layers of special cross operations, thereby generating high-order combinations of the features, and outputs a vector with a same dimension of 128 but containing rich feature interaction information; subsequently, the weighted fusion unit performs weighted splicing on the 128-dimensional structured features after cross processing and the 256-dimensional spatio-temporal features output by the spatio-temporal matrix processing module; finally, a unified feature vector with a dimension of 384, which fuses all modal information, is generated; here, the 128-dimensional structured features and the 256-dimensional spatio-temporal features are assigned different importance weights, the weight of the 256-dimensional spatio-temporal features is 2, and the weight of the 128-dimensional structured features is 1. 10.The method of claim 9, wherein: The dual-task output module simultaneously performs two prediction tasks based on the fused 384-dimensional feature vector, including a classification head and a regression head; the classification head is composed of a fully connected layer containing only a single neuron, and a Sigmoid activation function is applied at the output end to map the neuron output value to between 0 and 1, wherein the neuron output value is a probability prediction value of the vehicle performing "over-speed behavior" in the current space-time environment; in parallel, the regression head is also composed of a single-neuron fully connected layer, but adopts a linear activation function, i.e., no additional transformation, and the output value is directly used as a predicted "risk index" continuous value for quantitatively evaluating the potential traffic safety risk degree in the current scene; the outputs of the classification head and the regression head jointly constitute the comprehensive evaluation result of the model on the safety of the traffic scene.

Images