Vehicle pre-collision early warning method and system based on traffic accident element matching

Abstract

The invention discloses a vehicle pre-collision early warning method and system based on traffic accident element matching. The method comprises the following steps: extracting accident parameters from a traffic accident case; performing simulation restoration on the traffic accident process based on the extracted accident parameters, and extracting a multi-dimensional time sequence of traffic accident participants from the traffic accident process; importing the multi-dimensional time sequence into a DS-HDP-HMM model to obtain a corresponding hidden state sequence; marking each hidden state as a dangerous state, a secondary dangerous state and a normal state; storing the dangerous state sub-sequence and the secondary dangerous state sub-sequence into an accident state sub-library matched with the traffic accident; acquiring the type of a current road for a vehicle with an early warning demand, and acquiring real-time motion parameters of the vehicle and surrounding traffic participants; and calculating the similarity between the multi-dimensional time sequence of the vehicle and the sequence in the matching sub-library, and carrying out pre-collision early warning on the vehicle with the early warning demand based on the similarity.

Description

Technical Field

[0001] The present invention belongs to the technical field of traffic warning control, and in particular relates to a vehicle pre-collision warning method and system based on traffic accident element matching. Background Art

[0002] Existing vehicle warning systems widely use on-board sensors and vehicle networking technologies, including radar, lidar, camera, ultrasonic, infrared sensors, etc., to issue warnings by sensing the surrounding environment and potential dangers of the vehicle. At the same time, with the help of vehicle-to-vehicle (V2V) and vehicle-to-infrastructure (V2I) communication technologies, information sharing and coordinated warnings between vehicles and between vehicles and road facilities are realized. Although these technologies have played an important role in improving driving safety, they still face problems such as difficulty in multi-sensor fusion, poor performance of sensors under complex weather conditions, and limited computing power under real-time requirements. At present, the era of intelligence is approaching rapidly. How to ensure the driving safety of vehicles in the complex road environment where various traffic participants exist has become a key issue. Although today's on-board equipment and safety technology can avoid the occurrence of some accidents, there is still a lot of room for the development of automobile safety. Many vehicle collision accidents occur due to late braking triggering time, sensors failing to detect surrounding vehicles or abnormal judgment of drivers. Based on such reasons, this paper extracts the motion state parameters of both vehicles at the time of the accident from the accident case as the trigger conditions for preventing safety accidents. Through early warning, braking and moderate intervention of the driver's operation, the safety of the vehicle and the driver is protected when it is judged to be in line with previous accident cases. Summary of the invention

[0003] The present invention provides a vehicle pre-collision warning method and system based on traffic accident element matching, which solves the problem that traffic accidents may occur due to the current simple reliance on sensors to avoid accidents and the inability to collect movement information of all surrounding traffic participants.

[0004] In order to achieve the above technical objectives, the present invention adopts the following technical solutions:

[0005] A vehicle pre-collision warning method based on traffic accident element matching, comprising:

[0006] Step 1, extracting accident parameter data for restoring the traffic accident from the traffic accident case;

[0007] Step 2: Based on the extracted accident parameter data, simulate and restore the process before and after the traffic accident, and extract multiple motion parameters of the two parties involved in the traffic accident before and after the traffic accident from the restored traffic accident case to obtain a multidimensional time series describing the process before and after the traffic accident;

[0008] Step 3, import the multidimensional time series describing the process before and after the traffic accident into the DS-HDP-HMM model to obtain the corresponding hidden state sequence; mark each hidden state as a dangerous state, a sub-dangerous state and a normal state, and obtain a dangerous state subsequence, a sub-dangerous state subsequence and a normal state subsequence; then store the dangerous state subsequence and the sub-dangerous state subsequence, as well as the multidimensional time series of the corresponding time period, in the sub-library of the accident state library that matches the traffic accident; the accident state library is respectively set up with multiple sub-libraries according to the type of road where the traffic accident occurred and the types of the two parties involved in the traffic accident;

[0009] Step 4: The vehicle with warning needs obtains the current road type and the multiple motion parameters of itself and surrounding traffic participants in the recent long time periods including the current moment; wherein the motion parameters of surrounding traffic participants are obtained through the vehicle-road-cloud integrated device; the vehicle is combined with each surrounding traffic participant to obtain multiple groups of traffic accident participants to be predicted, and each group of traffic accident participants to be predicted corresponds to a group of multidimensional time series T 实 ;

[0010] Step 5: For each group of parties involved in the traffic accident to be predicted, locate the corresponding sub-library in the accident status database according to the road type and the type of the parties involved, and convert the corresponding multidimensional time series T 实 Calculate similarity with the multi-dimensional time series in the sub-database;

[0011] Step 6: Based on the similarity calculation value, a pre-collision warning is issued to the vehicle that needs a warning.

[0012] Furthermore, multiple motion parameters are extracted, including: speed, position, heading angle and relative distance of both traffic participants.

[0013] Furthermore, the road types include intersections, T-junctions, and highways.

[0014] Furthermore, the types of participants in traffic accidents include passenger vehicles, two-wheeled vehicles, and pedestrians.

[0015] Furthermore, the marking classification standard for each hidden state is: the state within a preset time before the collision is marked as a dangerous state, the state within a certain time before the dangerous state is marked as a secondary dangerous state, and the state within a certain time before the secondary dangerous state is marked as a normal state.

[0016] Furthermore, similarity calculation is performed by normalized cross-correlation, including:

[0017] First, the two sets of multidimensional time series involved in the calculation are normalized;

[0018] Then, the normalized cross-correlation is calculated on each dimension to calculate the shape similarity between the two time series on that dimension:

[0019]

[0020] in, Represent the two normalized time series of the j-th dimension, Representing time series The shape similarity between them; τ is the time offset, and different values ​​of τ represent the translation of one sequence relative to another sequence; i represents the time series The sample point number, n represents the time series Length; It means that under all possible time offsets τ, a τ is selected so that the two time series The normalized cross-correlation reaches its maximum value;

[0021] Then combine the shape similarities between the two time series in all dimensions to obtain the similarity between the two sets of multidimensional time series:

[0022]

[0023] Among them, T1 and T2 represent the two sets of multidimensional time series involved in the calculation, S(T1,T2) represents the similarity between the two sets of multidimensional time series T1 and T2, and w j represents the weight of the j-th dimension, and m represents the dimension of the multidimensional time series.

[0024] Furthermore, the method for performing pre-collision warning according to the similarity calculation value is:

[0025] First, the current multidimensional time series T of the vehicle 实 Compared with the similarity between each dangerous state subsequence in the positioning sub-library, if T 实 If the similarity between the sequence and any of the dangerous state subsequences exceeds the threshold, a warning is issued to remind the current driver that the vehicle is in a dangerous driving state; otherwise, T 实 Compared with the similarity between each sub-dangerous state subsequence in the positioning sub-library, if T 实 If the similarity between the current driving state and any of the sub-dangerous state subsequences exceeds a threshold, a warning is issued to remind the current driving vehicle that there is a risk of an accident.

[0026] A vehicle pre-collision warning system based on traffic accident element matching, comprising:

[0027] The accident parameter extraction module is used to: extract accident parameter data for restoring traffic accidents from traffic accident cases;

[0028] The accident restoration and data export module is used to: simulate and restore the process before and after the traffic accident based on the extracted accident parameter data, and extract multiple motion parameters of both parties involved in the traffic accident before and after the traffic accident from the restored traffic accident case to obtain a multi-dimensional time series describing the process before and after the traffic accident;

[0029] The accident state library construction module is used to: import the multidimensional time series describing the process before and after the traffic accident into the DS-HDP-HMM model to obtain the corresponding hidden state sequence; mark each hidden state as a dangerous state, a sub-dangerous state and a normal state, and obtain a dangerous state subsequence, a sub-dangerous state subsequence and a normal state subsequence; then store the dangerous state subsequence and the sub-dangerous state subsequence, as well as the multidimensional time series of the corresponding time period, in the sub-library of the accident state library that matches the traffic accident; the accident state library is provided with multiple sub-libraries according to the type of road where the traffic accident occurred and the types of the parties involved in the traffic accident;

[0030] The real-time data acquisition module is used to: for vehicles with early warning requirements, obtain the current road type, obtain the multiple motion parameters of the vehicle and surrounding traffic participants in the recent long time periods including the current moment; wherein the motion parameters of the surrounding traffic participants are obtained through the vehicle-road-cloud integrated device; combine the vehicle with each surrounding traffic participant to obtain multiple groups of traffic accident participants to be predicted, and each group of traffic accident participants to be predicted corresponds to a group of multidimensional time series T 实 ;

[0031] The similarity calculation module is used to locate the corresponding sub-library in the accident status library for each group of traffic accident participants to be predicted according to the road type and the type of both parties, and convert the corresponding multidimensional time series T 实 Calculate similarity with the multi-dimensional time series in the sub-database;

[0032] The warning module is used to: provide pre-collision warning to vehicles that currently have warning needs based on the similarity calculation value.

[0033] Beneficial Effects

[0034] This solution is different from the transmission's vehicle safety warning technology. It does not simply rely on the vehicle's on-board sensor acquisition equipment, but adds vehicle networking technology. In the future era of vehicle-road-cloud integrated traffic information sharing, it can obtain all-round movement information of surrounding traffic participants, help vehicles avoid the types of accidents that have occurred, and respond in time to emergency situations that cannot be detected by sensors, using "eyes" outside the vehicle to help improve vehicle safety while driving. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 is a schematic diagram of the vehicle pre-collision warning system of the present invention;

[0036] Figure 2 and Figure 3 It is a flow chart of the vehicle pre-collision warning method of the present invention. DETAILED DESCRIPTION

[0037] The following is a detailed description of an embodiment of the present invention. This embodiment is based on the technical solution of the present invention, and provides a detailed implementation method and a specific operation process to further explain the technical solution of the present invention.

[0038] Example 1

[0039] This embodiment provides a vehicle pre-collision warning method based on traffic accident factor matching, referring to Figure 1-3 As shown, including:

[0040] Step 1: extract accident parameter data for restoring traffic accidents from traffic accident cases.

[0041] Obtain relevant data and related information from traffic accident cases for subsequent step 2 to restore traffic accidents, such as accident cases recorded in CIDAS (China Traffic Accident In-depth Investigation Data) and GIDAS (German Traffic Accident In-depth Investigation Data). The extracted accident parameters include accident scene photos and videos (panoramic photos, close-up photos of vehicle collision points and damaged parts, brake marks and other physical evidence), physical measurement data (brake mark length and position, precise location of collision point, vehicle stop position), vehicle data recorder (EDR) data (vehicle speed, acceleration, brake and accelerator pedal position, yaw angular velocity, yaw rate, etc.), and other data (vehicle position and trajectory, road adhesion coefficient, accident scene base map, vehicle brand and model, vehicle inertia parameters, tire type and size, etc.).

[0042] Step 2: Based on the extracted accident parameter data, simulate and restore the process before and after the traffic accident, and extract multiple motion parameters of the two parties involved in the traffic accident before and after the traffic accident from the restored traffic accident case to obtain a multi-dimensional time series that describes the process before and after the traffic accident.

[0043] The simulation restoration process includes:

[0044] (1) Construct the accident base map using AutoCAD or software with the same function, drag it into the PC-Crash work interface, and set the various parameters in the scene according to the actual accident case, including: 1. Vehicle data: vehicle size, mass, center of gravity position, suspension system, tire characteristics and braking efficiency. In addition, vehicle performance characteristics data, such as steering angle and braking force distribution, are also required; 2. Accident scene information: data on the accident site are required, including road layout, slope, road friction coefficient (especially wet or icy roads) and brake marks; 3. Collision conditions: the initial speed, direction and angle of the vehicle at the time of collision need to be estimated. PC-Crash uses momentum-based reduction to calculate the interaction between vehicles during and after the collision; 4. Pedestrian or motorcycle data: If the accident involves pedestrians or motorcycles, pedestrian mass and movement, motorcycle dynamic data, etc. need to be entered. PC-Crash has special models for these situations, such as multi-body systems for pedestrian dynamics; 5. Post-collision movement: PC-Crash can simulate the movement of vehicles after the collision, including trajectory, rollover possibility and final stop position, and can also calculate secondary collisions and their consequences.

[0045] (2) Model the accident participants based on the information given in the case, and set the optimal collision contact point, collision angle, accident participant position and other parameters. The optimal collision contact point and angle mainly refer to those parameter settings that can restore the accident most realistically and in accordance with physical laws under specific accident scenarios and meet the research objectives. These parameters can be optimized through software simulation and repeated adjustments to achieve a restoration effect that is closest to the actual accident.

[0046] (3) Accident reconstruction simulation: Run the accident simulation using the parameters set in steps (2) and (3). Compare the running results (check whether the restored vehicle motion trajectory is consistent with physical evidence such as brake marks, tire marks, and the vehicle's final resting position at the accident scene. This is the most direct and effective verification method for accident reconstruction) with the actual accident information to determine whether the restoration is consistent. If not, return to step (2) to adjust the relevant parameters until the accident is restored.

[0047] Multiple motion parameters extracted from the restored traffic accident cases, including: speed, acceleration and heading angle, are specifically exported from the simulated restored Pc-crash software: select "Calculation Output" in the "Calculation" menu in the Pc-crash software, and in the "Calculation Output" window, check the "Speed", "Acceleration" and "Heading Angle" options, and select the traffic participants to be exported. Export data: Select "Export Data" in the "File" menu to export and save the file containing speed, acceleration and heading angle data.

[0048] Multiple motion parameters of traffic accident participants before and after the traffic accident are obtained to form a multi-dimensional time series that can describe the process before and after the traffic accident.

[0049] Step 3, import the multidimensional time series describing the process before and after the traffic accident into the DS-HDP-HMM model to obtain the corresponding hidden state sequence; mark each hidden state as a dangerous state, a sub-dangerous state and a normal state, and obtain a dangerous state subsequence, a sub-dangerous state subsequence and a normal state subsequence; then store the dangerous state subsequence and the sub-dangerous state subsequence, as well as the multidimensional time series of the corresponding time period, in the sub-sequence of the accident state library that matches the traffic accident; the accident state library is composed of multiple sub-libraries according to the type of road on which the traffic accident occurred and the type of the two parties involved in the traffic accident.

[0050] Step 3.1, import the multidimensional time series describing the process before and after the traffic accident into the DS-HDP-HMM model to obtain the corresponding hidden state sequence.

[0051] Among them, the DS-HDP-HMM model is mainly divided into HDP (Hierarchical Dirichlet Model) and HMM model (Hidden Markov Model). The HDP model is a statistical method for non-parametric Bayesian clustering and topic modeling. This method is introduced in the model to expand the traditional HMM model, allowing the model to automatically determine the number of states actually required. The hierarchical Dirichlet process models the state transfer matrix as a random process, so that we no longer need to define the number of implicit states of traffic participants during the accident process. For the HMM model, HDP is used to give the state transfer matrix in the HMM model, so that it can capture more complex state transfer structures. At the same time, the global shared random variables in the HDP model (the basis distribution of the Dirichlet process) provide a common probability space for different states, thereby realizing the sharing and constraint of state transition probabilities.

[0052] The basic components of the Hidden Markov Model (HMM) will not be described in detail. Here we briefly explain the working principle of the DS-HDPHMM model. First, the state transfer matrix in the HMM is constructed using HDP. HDP is mainly divided into two types of Dirichlet processes. One is the overall global Dirichlet process: β~GEM(γ), GEM (Griffiths-Engen-McCloskey) distribution is used to generate hidden states; the other is the local Dirichlet process:

[0053]

[0054] k i represents the state self-sustaining probability controlled by the two hyperparameters ρ1 and ρ2; α is the centralized parameter controlling the distribution of transition probability, π' irepresents the state transition probability generated by the Dirichlet distribution with the basic metric β; δ in the formula i represents the Dirac measure on state i.

[0055] By flexibly controlling the state transition through ρ1, ρ2, and α, the model can more flexibly handle the state division of complex long sequences. At the same time, the parameter θ of the observation function is required to determine the distribution of the observation data generated in each hidden state. Assuming that the observation data is generated by a Gaussian distribution, that is, For each time point t, the observed variable x t Generated by the corresponding hidden state: where F is the observed distribution, is the hidden state s t Corresponding parameters. In the Bayesian framework, the parameters θ of the observation function are usually sampled from a prior distribution H. The choice of the prior distribution depends on the expectation of the observed data. Formally, it can be expressed as:

[0056] θ~H

[0057] Among them, H is a prior distribution, which can be a known standard distribution (such as normal distribution, Gamma distribution, etc.) or a customized distribution based on domain knowledge.

[0058] After the model framework is built, it is necessary to determine whether to infer the hidden state through Gibbs sampling. The following are the brief steps:

[0059] (1) Initialization: Initialize all variables, including the state sequence z, transition probability π, global transition probability β, and observation model parameters θ.

[0060] (2) Model parameter update:

[0061] 1. Sample state sequence. For each time step t, update the hidden state z according to the current parameters and observation data:

[0062]

[0063] Among them, z -t represents the state sequence of other time steps, p(y t |z t ,θ) represents the likelihood function of a given state of the observation model, and They represent the probability of transition to the previous and next states respectively.

[0064] 2. Sampling state transition probability:

[0065] π i ~Dir(αβ+κδ i +n i )

[0066] Among them, n k is the number of transitions from state i to other states; α and k are the hyperparameters of the model, δ k An indicator representing state i.

[0067] 3. Sampling observation model parameters:

[0068] θ i ~p(θ i |y i ,H)

[0069] Among them, y i and H represent the observed data and prior distribution under a given state i, respectively.

[0070] 4. Probability of sampling self-sustainability:

[0071]

[0072] in, Indicates that from state z t-1 The total number of transitions to other states. Resample the auxiliary variable s according to the probability calculated above t The value of , thereby determining whether the state of the current time step remains unchanged.

[0073] (3) Iterative convergence: Repeat the above steps and make the Gibbs sampling chain converge through multiple iterations to obtain the final hidden state sequence.

[0074] Step 3.2, mark each hidden state as a dangerous state, a sub-dangerous state, and a normal state, and obtain a dangerous state subsequence, a sub-dangerous state subsequence, and a normal state subsequence.

[0075] The marking classification standard for each hidden state is as follows: the state within a preset time length (set to 2 seconds in this embodiment) before the collision is marked as a dangerous state, the state before the dangerous state (set to 2 seconds in this embodiment) is marked as a secondary dangerous state, and the state within a certain time length before the secondary dangerous state is marked as a normal state.

[0076] Step 3.3, the dangerous state subsequence and the sub-dangerous state subsequence, as well as the multi-dimensional time series of the corresponding time period, are stored in the sub-library of the accident state library that matches the traffic accident.

[0077] The accident status database described in this embodiment is combined with multiple sub-databases according to the road type where the traffic accident occurred and the types of other traffic accident participants, so as to quickly locate the sub-database during real-time warning to reduce the amount of calculation required for warning. The road types include crossroads, T-junctions, highways, etc. (not limited to these, all road types are included because accidents occur on various road types), and the types of other traffic accident participants include passenger vehicles, two-wheeled vehicles, pedestrians, etc.

[0078] Step 4: The vehicle that needs early warning obtains the current road type and the multiple motion parameters of itself and surrounding traffic participants in the most recent 3-second period including the current moment.

[0079] Among them, the vehicle-road-cloud integrated device obtains real-time information of traffic participants around the vehicle, including the driving direction, heading angle, speed and position changes of traffic participants. The vehicle is combined with each surrounding traffic participant to obtain multiple groups of traffic accident participants to be predicted, and each group of traffic accident participants to be predicted corresponds to a set of multidimensional time series T 实 .

[0080] Step 5: For each group of parties involved in the traffic accident to be predicted, locate the corresponding sub-library in the accident status database according to the road type and the type of the parties involved, and convert the corresponding multidimensional time series T 实 Calculate the similarity with each multidimensional time series in the sub-database.

[0081] This embodiment calculates similarity by normalizing the cross-correlation, including:

[0082] First, the two sets of multidimensional time series involved in the calculation are normalized;

[0083] Then, the normalized cross-correlation is calculated on each dimension to calculate the shape similarity between the two time series on that dimension:

[0084]

[0085] in, Represent the two normalized time series of the j-th dimension, Representing time series The shape similarity between them; τ is the time offset, and different values ​​of τ represent the translation of one sequence relative to another sequence; i represents the time series The sample point number, n represents the time series Length; It means that under all possible time offsets τ, a τ is selected so that the two time series The normalized cross-correlation reaches its maximum value;

[0086] Then combine the shape similarities between the two time series in all dimensions to obtain the similarity between the two sets of multidimensional time series:

[0087]

[0088] Among them, T1 and T2 represent the two sets of multidimensional time series involved in the calculation, S(T1,T2) represents the similarity between the two sets of multidimensional time series T1 and T2, and w j represents the weight of the j-th dimension, and m represents the dimension of the multidimensional time series.

[0089] Step 6: Based on the similarity calculation value, a pre-collision warning is issued to the vehicle that needs a warning.

[0090] First, the multidimensional time series T of the two parties involved in each group of traffic accidents to be predicted 实 Compared with the similarity between each dangerous state subsequence in the positioning sub-library, if T 实 If the similarity between the sequence and any of the dangerous state subsequences exceeds the threshold, a warning is issued to remind the current driver that the vehicle is in a dangerous driving state; otherwise, T 实 Compared with the similarity between each sub-dangerous state subsequence in the positioning sub-library, if T 实 If the similarity between the current driver and any of the sub-dangerous state subsequences exceeds a threshold, a warning is issued to remind the current driver of the risk of an accident. Otherwise, there is no need to issue a warning signal and the vehicle can maintain normal driving.

[0091] In a better implementation scheme, while reminding the current driver that the vehicle is in a dangerous driving state, it continues to determine whether the TTC between the current vehicle and the dangerous conflicting vehicle has reached the threshold. If the TTC threshold is not reached, the AEB emergency braking system is intervened; if the TTC threshold is exceeded, it is considered that braking cannot avoid the occurrence of the accident, and then a certain degree of intervention is made in the control of the vehicle to reduce the possibility of an accident.

[0092] Example 2

[0093] This embodiment provides a vehicle pre-collision warning system based on traffic accident factor matching, which is characterized by comprising:

[0094] The accident parameter extraction module is used to: extract accident parameter data for restoring traffic accidents from traffic accident cases;

[0095] The accident restoration and data export module is used to: simulate and restore the process before and after the traffic accident based on the extracted accident parameter data, and extract multiple motion parameters of the two parties involved in the traffic accident before and after the traffic accident from the restored traffic accident case to obtain a multi-dimensional time series describing the process before and after the traffic accident;

[0096] The accident state library construction module is used to: import the multidimensional time series describing the process before and after the traffic accident into the DS-HDP-HMM model to obtain the corresponding hidden state sequence; mark each hidden state as a dangerous state, a sub-dangerous state and a normal state, and obtain a dangerous state subsequence, a sub-dangerous state subsequence and a normal state subsequence; then store the dangerous state subsequence and the sub-dangerous state subsequence, as well as the multidimensional time series of the corresponding time period, in the sub-library of the accident state library that matches the traffic accident; the accident state library is provided with multiple sub-libraries according to the type of road where the traffic accident occurred and the types of the two parties involved in the traffic accident;

[0097] The real-time data acquisition module is used to: for vehicles with early warning requirements, obtain the current road type, obtain the multiple motion parameters of the vehicle and surrounding traffic participants in the recent long time periods including the current moment; wherein the motion parameters of the surrounding traffic participants are obtained through the vehicle-road-cloud integrated device; combine the vehicle with each surrounding traffic participant to obtain multiple groups of traffic accident participants to be predicted, and each group of traffic accident participants to be predicted corresponds to a group of multidimensional time series T 实 ;

[0098] The similarity calculation module is used to locate the corresponding sub-library in the accident status library for each group of traffic accident participants to be predicted according to the road type and the type of both parties, and convert the corresponding multidimensional time series T 实 Calculate similarity with the multi-dimensional time series in the sub-database;

[0099] The warning module is used to: provide pre-collision warning to vehicles that currently have warning needs based on the similarity calculation value.

[0100] The working mode of each module in this embodiment is the same as that in Embodiment 1 and will not be described again.

[0101] The above embodiments are preferred embodiments of the present application. Ordinary technicians in this field can also make various changes or improvements on this basis. Without departing from the overall concept of the present application, these changes or improvements should fall within the scope of protection required by the present application.

Claims

1. A vehicle pre-collision warning method based on traffic accident element matching, characterized in that: include: Step 1, extracting accident parameter data for restoring the traffic accident from the traffic accident case; Step 2: Based on the extracted accident parameter data, simulate and restore the process before and after the traffic accident, and extract multiple motion parameters of the two parties involved in the traffic accident before and after the traffic accident from the restored traffic accident case to obtain a multidimensional time series describing the process before and after the traffic accident; Step 3, import the multidimensional time series describing the process before and after the traffic accident into the DS-HDP-HMM model to obtain the corresponding hidden state sequence; mark each hidden state as a dangerous state, a sub-dangerous state and a normal state, and obtain a dangerous state subsequence, a sub-dangerous state subsequence and a normal state subsequence; then store the dangerous state subsequence and the sub-dangerous state subsequence, as well as the multidimensional time series of the corresponding time period, in the sub-library of the accident state library that matches the traffic accident; the accident state library is respectively set up with multiple sub-libraries according to the type of road where the traffic accident occurred and the types of the two parties involved in the traffic accident; Step 4: The vehicle with the warning demand obtains the current road type and the multiple motion parameters of itself and surrounding traffic participants in the recent long time periods including the current moment; wherein the motion parameters of surrounding traffic participants are obtained through the vehicle-road-cloud integrated device; the vehicle is combined with each surrounding traffic participant to obtain multiple groups of traffic accident participants to be predicted, and each group of traffic accident participants to be predicted corresponds to a group of multidimensional time series T 实 ; Step 5: For each group of parties involved in the traffic accident to be predicted, locate the corresponding sub-library in the accident status database according to the road type and the type of the parties involved, and convert the corresponding multidimensional time series T 实 Calculate similarity with the multi-dimensional time series in the sub-database; Step 6: Based on the similarity calculation value, a pre-collision warning is issued to the vehicle that needs a warning.

2. The vehicle pre-collision warning method based on traffic accident element matching according to claim 1 is characterized in that: Multiple motion parameters extracted include: speed, position, heading angle and relative distance of both traffic participants.

3. The vehicle pre-collision warning method based on traffic accident element matching according to claim 1 is characterized in that: The road types include intersections, T-junctions, and expressways.

4. The vehicle pre-collision warning method based on traffic accident element matching according to claim 1, characterized in that: Types of traffic accident participants, including passenger vehicles, two-wheelers, and pedestrians.

5. The vehicle pre-collision warning method based on traffic accident element matching according to claim 1 is characterized in that: The marking classification standard for each hidden state is: the state within a preset time before the collision is marked as a dangerous state, the state within a certain time before the dangerous state is marked as a secondary dangerous state, and the state within a certain time before the secondary dangerous state is marked as a normal state.

6. The vehicle pre-collision warning method based on traffic accident element matching according to claim 1 is characterized in that: Similarity calculation is performed through normalized cross-correlation, including: First, the two sets of multidimensional time series involved in the calculation are normalized; Then, the normalized cross-correlation is calculated on each dimension to calculate the shape similarity between the two time series on that dimension: in, Represent the two normalized time series of the j-th dimension, Representing time series The shape similarity between them; τ is the time offset, and different values ​​of τ represent the translation of one sequence relative to another sequence; i represents the time series The sample point number, n represents the time series Length; It means that under all possible time offsets τ, a τ is selected so that the two time series The normalized cross-correlation of reaches its maximum value; Then combine the shape similarities between the two time series in all dimensions to obtain the similarity between the two sets of multidimensional time series: Among them, T1 and T2 represent the two sets of multidimensional time series involved in the calculation, S(T1,T2) represents the similarity between the two sets of multidimensional time series T1 and T2, and w j represents the weight of the j-th dimension, and m represents the dimension of the multidimensional time series.

7. The vehicle pre-collision warning method based on traffic accident element matching according to claim 1, characterized in that: The method for pre-collision warning based on similarity calculation value is: First, the current multidimensional time series T of the vehicle 实 Compared with the similarity between each dangerous state subsequence in the positioning sub-library, if T 实 If the similarity between the vehicle and any of the dangerous state subsequences exceeds a threshold, a warning is issued to remind the current driver that the vehicle is in a dangerous driving state; Otherwise, T 实 Compared with the similarity between each sub-dangerous state subsequence in the positioning sub-library, if T 实 If the similarity between the current driver and any of the sub-dangerous state subsequences exceeds a threshold, a warning is issued to remind the current driver that there is a risk of an accident.

8. A vehicle pre-collision warning system based on traffic accident element matching, characterized in that: include: The accident parameter extraction module is used to: extract accident parameter data for restoring traffic accidents from traffic accident cases; The accident restoration and data export module is used to: simulate and restore the process before and after the traffic accident based on the extracted accident parameter data, and extract multiple motion parameters of both parties involved in the traffic accident before and after the traffic accident from the restored traffic accident case to obtain a multi-dimensional time series describing the process before and after the traffic accident; The accident state library construction module is used to: import the multidimensional time series describing the process before and after the traffic accident into the DS-HDP-HMM model to obtain the corresponding hidden state sequence; mark each hidden state as a dangerous state, a sub-dangerous state and a normal state, and obtain a dangerous state subsequence, a sub-dangerous state subsequence and a normal state subsequence; then store the dangerous state subsequence and the sub-dangerous state subsequence, as well as the multidimensional time series of the corresponding time period, in the sub-library of the accident state library that matches the traffic accident; the accident state library is provided with multiple sub-libraries according to the type of road where the traffic accident occurred and the types of the parties involved in the traffic accident; The real-time data acquisition module is used to: for vehicles with early warning requirements, obtain the current road type, obtain the multiple motion parameters of the vehicle and surrounding traffic participants in the recent long time periods including the current moment; wherein the motion parameters of the surrounding traffic participants are obtained through the vehicle-road-cloud integrated device; combine the vehicle with each surrounding traffic participant to obtain multiple groups of traffic accident participants to be predicted, and each group of traffic accident participants to be predicted corresponds to a group of multidimensional time series T 实 ; The similarity calculation module is used to locate the corresponding sub-library in the accident status library for each group of traffic accident participants to be predicted according to the road type and the type of both parties, and convert the corresponding multidimensional time series T 实 Calculate similarity with the multi-dimensional time series in the sub-database; The warning module is used to: provide pre-collision warning to vehicles that currently have warning needs based on the similarity calculation value.

9. The vehicle pre-collision warning device based on traffic accident element matching according to claim 8, characterized in that: Multiple motion parameters extracted include: speed, position, heading angle and relative distance of both traffic participants.

Images