A road traffic accident detection method and system based on roadside radar perception

Abstract

The present invention relates to the field of radar sensing technology, and in particular to a road traffic accident detection method and system based on roadside radar sensing, comprising the following steps: S1, data collection: continuously collecting motion data of traffic participants; S2, multi-source data fusion and preprocessing: real-time collection of environmental perception data, and fusion of the collected motion data and environmental perception data into multi-source data; S3, real-time dynamic analysis: identifying abnormal behaviors of traffic participants; S4, accident identification and risk assessment: risk assessment of the results of real-time dynamic analysis; S5, virtual accident scene reconstruction: reconstruction of virtual accident scene; S6, alarm and information push: sending alarm information to relevant traffic management departments and emergency rescue units. The present invention can reproduce the entire process of the accident, and provides reliable data support for accident analysis, prediction and responsibility determination.

Description

Technical Field

[0001] The present invention relates to the field of radar sensing technology, and in particular to a road traffic accident detection method and system based on roadside radar sensing. Background Art

[0002] In modern traffic management systems, with the increase in the number of vehicles and the continuous rise in traffic flow, the frequent occurrence of road traffic accidents has become a serious social problem. In order to reduce the frequency of accidents and improve the response efficiency after accidents, more and more intelligent traffic detection and management systems are being used in practice. These systems usually monitor road traffic conditions in real time through a variety of sensor technologies, collect and analyze relevant data, so as to realize early warning, detection and analysis of traffic accidents.

[0003] However, existing technologies still have many shortcomings in traffic accident detection and management. Traditional traffic detection systems mostly rely on a single sensor or data source, which makes it impossible to comprehensively and accurately capture and analyze the entire process of an accident in a complex traffic environment. In addition, these systems usually lack detailed process reconstruction after the accident and accident analysis under multiple conditions, and cannot provide sufficient data support for accident responsibility determination and subsequent processing. At the same time, the response speed of existing systems and the accuracy of accident warnings also have certain limitations under changing traffic conditions, making it difficult to meet the needs of practical applications.

[0004] In order to solve the above problems, the present invention proposes a road traffic accident detection method based on roadside radar perception, which aims to significantly improve the real-time and accuracy of traffic accident detection through multi-source data fusion, real-time dynamic analysis, virtual accident scene reconstruction and accident simulation under multiple conditions. Summary of the invention

[0005] Based on the above objectives, the present invention provides a road traffic accident detection method and system based on roadside radar perception.

[0006] A road traffic accident detection method based on roadside radar perception comprises the following steps:

[0007] S1, data collection: through radar sensors installed on both sides of the road, the movement data of traffic participants are continuously collected, including speed, acceleration, location information and vehicle spacing;

[0008] S2, multi-source data fusion and preprocessing: collect environmental perception data in real time through a variety of collection devices, including visual data, traffic flow data and communication data, fuse the collected motion data and environmental perception data into multi-source data, and preprocess the fused multi-source data;

[0009] S3, real-time dynamic analysis: real-time dynamic analysis of pre-processed multi-source data to identify abnormal behaviors of traffic participants and predict potential traffic accidents;

[0010] S4, accident identification and risk assessment: Conduct risk assessment on the results of real-time dynamic analysis, dynamically adjust the weight of risk assessment according to vehicle speed, vehicle distance, pedestrian density and weather conditions, and determine the probability and severity of the accident;

[0011] S5, virtual accident scene reconstruction: After a traffic accident is detected, the virtual accident generation model is used to analyze the collected multi-source data and reconstruct the virtual accident scene, including:

[0012] S51, Data parsing and feature extraction: Extract relevant features from multi-source data, including ambient lighting conditions, road conditions, traffic signal status, pedestrian behavior characteristics, vehicle types and characteristics;

[0013] S52, Model construction: Constructing a virtual accident generation model using historical multi-source data;

[0014] S53, virtual accident scene generation: inputting the extracted relevant features into the virtual accident generation model to generate a virtual accident scene, including the entire scene before, during and after the accident;

[0015] S54, scenario simulation and multi-condition analysis: Based on the virtual accident scenario, simulate the accident scenarios under different conditions (such as different vehicle speeds, weather, etc.), analyze the impact of various factors on the accident, and generate multiple accident development paths;

[0016] S6, Alarm and information push: After completing the reconstruction of the virtual accident scene, alarm information will be automatically sent to relevant traffic management departments and emergency rescue units according to the severity of the traffic accident, and accident warning information will be pushed to users through road display screens and vehicle terminals.

[0017] Optionally, the data collection in S1 includes:

[0018] S11, radar sensor installation: multiple radar sensors are installed at predetermined intervals on fixed brackets on both sides of the road;

[0019] S12, high-frequency pulse signal transmission: each radar sensor continuously scans the traffic participants in its coverage area by transmitting a high-frequency pulse signal to capture the movement of the target object;

[0020] S13, radar signal reception and processing: The radar sensor receives radar signals reflected by traffic participants and calculates the motion data of each target in real time, including speed, acceleration, position information, and vehicle spacing.

[0021] Optionally, the multi-source data fusion and preprocessing in S2 includes:

[0022] S21, environmental perception data collection: collect environmental perception data in real time through a variety of collection devices, including:

[0023] Visual data: Collect visual data through on-board cameras, including images on the road, pedestrian behavior, and lane marking status;

[0024] Traffic flow data: Traffic flow data is collected through traffic flow meters, including traffic density, speed distribution, and vehicle types;

[0025] Communication data: Communication data is collected through V2X communication equipment, including communication data between vehicles, traffic signal status, and emergency alarms;

[0026] S22, data fusion: Adopt adaptive weighted average algorithm to fuse the collected motion data and environmental perception data into multi-source data, expressed as:

[0027] D fused =w1·D motion +w2·D visual +w3·D traffic +w4·D communication ;

[0028] Among them, D fused represents the fused multi-source data, D motion .D visual ,D traffic, D communication They represent motion data, visual data, traffic flow data, and communication data, respectively, and w1, w2, w3, and w4 are the corresponding weights;

[0029] S23, data preprocessing: using Kalman filter to filter the fused multi-source data.

[0030] Optionally, the real-time dynamic analysis in S3 includes:

[0031] S31, data input: input the preprocessed multi-source data into the convolutional neural network (CNN) model to extract the spatiotemporal characteristics of traffic participants’ behaviors and capture the behavior patterns of traffic participants;

[0032] S32, abnormal behavior recognition: Based on the spatiotemporal features extracted by the convolutional neural network (CNN) model, abnormal behavior is recognized by calculating the difference between the current behavior features and the normal behavior features using the Euclidean distance.

[0033] Optionally, the accident identification and risk assessment in S4 includes:

[0034] S41, feature weight adjustment: Based on the results of real-time dynamic analysis, combined with the current vehicle speed, vehicle distance, pedestrian density and weather conditions, the gradient boosting decision tree (GBDT) is used to adjust the weight, which is expressed as:

[0035] S411, define the loss function:

[0036] Among them, y i is the true label, F m (x i ) is the predicted value of the mth tree, l is the squared error loss function, α is the learning rate, and h m (x i ) is the predicted output of the mth tree;

[0037] S412, Calculation of Feature Importance: I j =∑ t∈T ΔL t I(t);

[0038] Among them, I j is the importance score of feature j, ΔL t is the loss reduction brought by each split in the decision tree, and I(t) is the indicator function of whether feature j participates in the split;

[0039] S413, Dynamic Weight Adjustment:

[0040] Among them, w j is the dynamically adjusted weight of feature j, and M is the number of all features;

[0041] S42, accident probability calculation: Based on the adjusted feature weights, the probability of an accident is calculated using a logistic regression model, expressed as:

[0042] ;

[0043] Among them, P (Accident|Features) is the probability of an accident, β0 is the bias term of the model, and β i For each feature x i The regression coefficient, w i is the dynamically adjusted weight;

[0044] S43, severity assessment: Based on the determination of the probability of an accident, the severity of the accident is assessed using a weighted summation algorithm by comprehensively analyzing the speed, vehicle distance, pedestrian density and weather conditions of traffic participants, expressed as:

[0045]

[0046] Where S is the final severity score, w i is the weight of the ith input variable, x i is the standardized value of the ith input variable, and n is the total number of input variables.

[0047] Optionally, the data analysis and feature extraction in S51 include:

[0048] S511, extracting ambient lighting conditions: acquiring a road image through a vehicle-mounted camera, and extracting ambient lighting conditions using a grayscale value average algorithm;

[0049] S512, extraction of road surface conditions: using image features in visual data, through image Canny edge detection algorithm and Gabor filter, to extract road surface texture features, and determine whether the road surface is slippery, potholes or water accumulation;

[0050] S513, extraction of traffic signal status: obtaining real-time traffic signal status through V2X communication equipment, and using color recognition algorithm to identify the color of the traffic light;

[0051] S514, extraction of pedestrian behavior features: by combining the on-board camera with the convolutional neural network (CNN) model, the pedestrian behavior features are extracted, including walking direction, speed, and whether they suddenly cross the road;

[0052] S515, vehicle type and feature extraction: Extract the appearance features of the vehicle, including vehicle type, color, and size, through visual data and convolutional neural network (CNN) model.

[0053] Optionally, the virtual accident generation model in S52 adopts a generative adversarial network (GANs) model, and the generative adversarial network (GANs) model includes:

[0054] S521, multimodal data input and conditional fusion: Multimodal data (such as ambient lighting, road conditions, traffic signal status, etc.) is introduced as input and combined with the noise vector, expressed as:

[0055] z~p z (z);

[0056] Where z is a noise vector sampled from a standard normal distribution;

[0057] y=concat(y1,y2,...,y n );

[0058] Among them, y is the conditional vector, y1,y2,...,y nIt is the standardized vector of each modal data (such as ambient lighting, road conditions), and concat means concatenating multiple modal data into a conditional vector;

[0059] z cond =concat(z,y);

[0060] Among them, z cond is the fusion input;

[0061] S522, condition generator: the generator receives the fused input z cond , the virtual accident scenario that generates the conditional vector y is expressed as:

[0062] x fake =G(z cond ) = G(concat(z,y));

[0063] Among them, x fake is the generated virtual accident scenario, G is the generator;

[0064] S523, multi-scale discriminator: A multi-scale discriminator is introduced to discriminate the generated virtual accident scene at different scales (such as global scale and local scale), which is expressed as:

[0065] Multi-scale loss of the discriminator:

[0066] Among them, D s represents the discriminator at scale s, and E represents the expected value;

[0067] Multi-scale total loss:

[0068] Among them, λ s is the weight of each scale loss, S is the total number of scales;

[0069] S524, multi-scale generation loss combined with reconstruction loss: In the loss function of the generator, the multi-scale generation loss and reconstruction loss are combined, expressed as:

[0070] Multi-scale generation loss:

[0071] Reconstruction loss:

[0072] Among them, x real It is a real accident scene;

[0073] Generator total loss:

[0074] Among them, λ reconis the weight coefficient of reconstruction loss;

[0075] S525, adversarial training and updating: Use adversarial training method to update the generator and discriminator alternately, expressed as:

[0076] Generator Updates:

[0077] Among them, θ G is the parameter of the generator, η is the learning rate;

[0078] Discriminator update:

[0079] Among them, θ D are the parameters of the discriminator and η is the learning rate.

[0080] Optionally, the generation of the virtual accident scene in S53 includes:

[0081] S531, feature fusion input: after standardization, the relevant features extracted from multi-source data are fused into a comprehensive feature vector;

[0082] S532, feature input generation model: combining the fused comprehensive feature vector with the random noise vector and inputting the vector into the virtual accident generation model as an input vector;

[0083] S534, generate accident scene: the virtual accident generation model generates a virtual accident scene according to the input comprehensive feature vector, including the whole scene before, during and after the accident.

[0084] Optionally, the scenario simulation and multi-condition analysis in S54 include:

[0085] S541, condition variable setting: based on the generated virtual accident scenario, multiple different condition variables are set, including different vehicle speeds, weather conditions, traffic densities, and vehicle types, to create different simulation scenario conditions;

[0086] S542, multi-condition scenario generation: input the set condition variables into the virtual accident generation model respectively to generate accident scenarios under different conditions, which are expressed as:

[0087]

[0088] in, is the conditional variable y i The fused input with the noise z, is the conditional variable y i The accident scenario generated below;

[0089] S543, Factor Impact Analysis: By comparing the accident scenarios generated under different conditions, analyzing the impact of various conditional variables (such as vehicle speed changes, weather deterioration) on the severity and probability of accidents, identifying key factors and their influence on the development of accidents, expressed as:

[0090]

[0091] in, is the conditional variable y i The severity of the accident assessed under base is the accident severity under baseline conditions, ΔS i is the conditional variable y i incremental impact on accident severity;

[0092] S544, Multi-path Accident Development Generation: Simulate and generate multiple accident development paths under different conditions, and analyze the results of the accident under different circumstances, expressed as:

[0093]

[0094] in, is the accident in the condition variable y i The probability path of development is is the conditional variable y i Next, based on the generated scene The path probability.

[0095] A road traffic accident detection system based on roadside radar perception is used to implement the above-mentioned road traffic accident detection method based on roadside radar perception, and includes the following modules:

[0096] Radar sensor module: Through radar sensors installed on both sides of the road, the motion data of traffic participants are continuously collected, including speed, acceleration, location information and vehicle spacing;

[0097] Multi-source data fusion and preprocessing module: collects environmental perception data in real time through a variety of collection devices, including visual data, traffic flow data and communication data, and fuses and preprocesses environmental perception data with motion data;

[0098] Real-time dynamic analysis module: performs real-time analysis on pre-processed multi-source data, identifies abnormal behaviors of traffic participants, and predicts potential traffic accidents;

[0099] Accident identification and risk assessment module: Conduct risk assessment based on real-time dynamic analysis results, dynamically adjust the risk assessment weight, and determine the probability and severity of the accident;

[0100] Virtual accident scene reconstruction module: After a traffic accident is detected, the virtual accident generation model is used to analyze the collected multi-source data and reconstruct the virtual accident scene;

[0101] Alarm and information push module: According to the severity of the virtual accident scene, send alarm information to the traffic management department and emergency rescue units, and push accident warning information to users through road display screens and vehicle terminals.

[0102] Beneficial effects of the present invention:

[0103] The present invention collects and integrates the motion data of traffic participants and environmental perception data in real time through radar sensors and various acquisition devices installed on both sides of the road. Through the application of convolutional neural networks and gradient boosting decision trees, the system can accurately identify the abnormal behaviors of traffic participants and dynamically evaluate the probability and severity of accidents, thereby significantly improving the real-time and accuracy of accident detection.

[0104] The present invention generates a virtual accident scene by generating an adversarial network model and combining multimodal data, a multi-scale discriminator and a reconstruction loss optimization generation process. The virtual accident scene not only conforms to the real traffic environment in terms of global characteristics, but also is highly consistent with the actual accident scene in terms of local details. This virtual scene can reproduce the entire process of the accident, covering the entire scene before, during and after the accident, providing reliable data support for accident analysis, prediction and responsibility determination.

[0105] The present invention, through multi-condition simulation and analysis of virtual accident scenes, can flexibly set different variables such as vehicle speed, weather, traffic density, etc., and deeply analyze the impact of various factors on the severity and development path of the accident. The generated multiple accident development paths provide a scientific basis for the formulation of accident prevention, emergency response and management measures, effectively reducing the occurrence of secondary accidents and improving the overall efficiency and safety of the traffic management system. BRIEF DESCRIPTION OF THE DRAWINGS

[0106] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings in the following description are only for the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0107] Figure 1 A schematic diagram of a detection method flow chart of an embodiment of the present invention;

[0108] Figure 2 Schematic diagram of system function modules according to an embodiment of the present invention. DETAILED DESCRIPTION

[0109] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. At the same time, it is explained here that in order to make the embodiments more detailed, the following embodiments are the best and preferred embodiments, and those skilled in the art may also adopt other alternatives to implement some known technologies; and the accompanying drawings are only for more specific description of the embodiments, and are not intended to specifically limit the present invention.

[0110] It should be noted that the references to "one embodiment", "an embodiment", "an exemplary embodiment", "some embodiments" and the like in the specification indicate that the embodiments described may include specific features, structures or characteristics, but not every embodiment may include the specific features, structures or characteristics. In addition, when a specific feature, structure or characteristic is described in conjunction with an embodiment, it should be within the knowledge of a person skilled in the art to implement such feature, structure or characteristic in conjunction with other embodiments (whether or not explicitly described).

[0111] In general, a term can be understood, at least in part, from its use in context. For example, depending, at least in part, on the context, the term "one or more" as used herein can be used to describe any feature, structure, or characteristic in the singular sense, or can be used to describe a combination of features, structures, or characteristics in the plural sense. Additionally, the term "based on" can be understood as not necessarily intended to convey an exclusive set of factors, but can instead, depending, at least in part, on the context, allow for the presence of other factors that are not necessarily explicitly described.

[0112] like Figure 1 As shown, a road traffic accident detection method based on roadside radar perception includes the following steps:

[0113] S1, Data Collection: Radar sensors installed on both sides of the road continuously collect the movement data of traffic participants, including speed, acceleration, location information and vehicle spacing. Radar sensors can work normally in all weather conditions and accurately detect different types of targets;

[0114] S2, multi-source data fusion and preprocessing: collect environmental perception data in real time through a variety of collection devices, including visual data, traffic flow data and communication data, fuse the collected motion data and environmental perception data into multi-source data, and preprocess the fused multi-source data to improve the overall reliability and accuracy of the data;

[0115] S3, real-time dynamic analysis: real-time dynamic analysis of pre-processed multi-source data to identify abnormal behaviors of traffic participants and predict potential traffic accidents;

[0116] S4, accident identification and risk assessment: Conduct risk assessment on the results of real-time dynamic analysis, dynamically adjust the weight of risk assessment according to vehicle speed, vehicle distance, pedestrian density and weather conditions, and determine the probability and severity of the accident;

[0117] S5, virtual accident scene reconstruction: After a traffic accident is detected, the virtual accident generation model is used to analyze the collected multi-source data and reconstruct the virtual accident scene. The virtual scene can reproduce the entire process of the accident and simulate accident scenes under different conditions, including:

[0118] S51, Data parsing and feature extraction: Extract relevant features from multi-source data, including ambient lighting conditions, road conditions, traffic signal status, pedestrian behavior characteristics, vehicle types and characteristics;

[0119] S52, Model construction: Constructing a virtual accident generation model using historical multi-source data;

[0120] S53, virtual accident scene generation: input the extracted relevant features into the virtual accident generation model to generate a virtual accident scene, including the entire scene before, during and after the accident, to ensure the continuity of the scene and the accuracy of details;

[0121] S54, scenario simulation and multi-condition analysis: Based on the virtual accident scenario, simulate the accident scenarios under different conditions (such as different vehicle speeds, weather, etc.), analyze the impact of various factors on the accident, and generate multiple accident development paths to assist in accident cause analysis and responsibility determination;

[0122] S6, Alarm and information push: After completing the reconstruction of the virtual accident scene, it will automatically send alarm information to the relevant traffic management department and emergency rescue unit according to the severity of the traffic accident, and push accident warning information to other road users through road display screens and vehicle terminals to reduce the occurrence of secondary accidents;

[0123] Through the above content, not only the real-time and accuracy of accident detection are improved, but also detailed accident process reproduction and multi-condition analysis can be provided through virtual scene reconstruction after the accident, which greatly assists in accident cause analysis and responsibility determination. At the same time, through timely alarms and information push, the risk of secondary accidents is effectively reduced, and the efficiency and safety of overall traffic management are improved.

[0124] Data collection in S1 includes:

[0125] S11, radar sensor installation: multiple radar sensors are installed on fixed brackets on both sides of the road at predetermined intervals to ensure that each radar sensor can cover the traffic monitoring range of the area where it is located;

[0126] S12, high-frequency pulse signal transmission: each radar sensor continuously scans the traffic participants in its coverage area by transmitting a high-frequency pulse signal to capture the movement of the target object;

[0127] S13, radar signal reception and processing: The radar sensor receives the radar signal reflected by the traffic participants and calculates the motion data of each target in real time, including speed, acceleration, position information and vehicle spacing;

[0128] The speed is calculated by the Doppler frequency shift formula, which is expressed as:

[0129]

[0130] Where v is the target velocity, Δf is the Doppler frequency shift, c is the speed of light, and f0 is the frequency of the radar transmission signal;

[0131] The vehicle spacing is calculated by measuring the signal round trip time to calculate the target distance, which is expressed as:

[0132]

[0133] Where d is the target distance, Δt is the signal round trip time, and c is the speed of light;

[0134] Acceleration is calculated by continuously measuring the rate of change of velocity, expressed as:

[0135]

[0136] Where a is the target acceleration, Δv is the velocity change, and Δt is the time interval;

[0137] Through the above steps, not only the coverage and accuracy of traffic monitoring are improved, but also a solid data foundation is provided for subsequent traffic accident detection and analysis, significantly improving the overall reliability and response speed of the system.

[0138] Multi-source data fusion and preprocessing in S2 include:

[0139] S21, environmental perception data collection: collect environmental perception data in real time through a variety of collection devices, including:

[0140] Visual data: Collect visual data through on-board cameras, including images on the road, pedestrian behavior, and lane marking status;

[0141] Traffic flow data: Traffic flow data is collected through traffic flow meters, including traffic density, speed distribution, and vehicle types;

[0142] Communication data: Communication data is collected through V2X communication equipment, including communication data between vehicles, traffic signal status, and emergency alarms;

[0143] S22, data fusion: Adopt adaptive weighted average algorithm to fuse the collected motion data and environmental perception data into multi-source data, expressed as:

[0144] D fused =w1·D motion +w2·D visual +w3·D traffic +w4·D communication ;

[0145] Among them, D fused represents the fused multi-source data, D motion .D visual ,D traffic, D communication They represent motion data, visual data, traffic flow data, and communication data, respectively, and w1, w2, w3, and w4 are the corresponding weights;

[0146] S23, data preprocessing: Kalman filter is used to filter the fused multi-source data to eliminate noise and outliers, which is expressed as:

[0147]

[0148] in, is the estimated state at the current moment, is the predicted state at the previous moment, K k is the Kalman gain, z k is the measurement value at the current moment, and H is the measurement matrix;

[0149] Through the above steps, multi-source data can be effectively integrated, the overall reliability and accuracy of the data can be improved, and a high-quality data foundation can be provided for subsequent traffic accident detection and analysis.

[0150] Real-time dynamic analysis in S3 includes:

[0151] S31, data input: The preprocessed multi-source data is input into the convolutional neural network (CNN) model to extract the spatiotemporal characteristics of traffic participants’ behaviors and capture the behavior patterns of traffic participants, which can be expressed as:

[0152]

[0153] in, is the activation value of position (i, j) in layer l, is the convolution kernel weight, is the input of the previous layer, b (l) is the bias term, σ is the activation function;

[0154] S32, abnormal behavior recognition: Based on the spatiotemporal features extracted by the convolutional neural network (CNN) model, the abnormal behavior is recognized by calculating the difference between the current behavior features and the normal behavior features using the Euclidean distance, which is expressed as:

[0155]

[0156] Among them, x k and k are the vectors of current behavior features and normal behavior features respectively. When the Euclidean distance exceeds the threshold and is lower than the threshold, it is marked as abnormal behavior;

[0157] Through the above content, the system's real-time analysis capabilities and sensitivity to abnormal situations have been effectively improved, providing more reliable basic data for subsequent accident identification and risk assessment.

[0158] Accident identification and risk assessment in S4 include:

[0159] S41, feature weight adjustment: Based on the results of real-time dynamic analysis, combined with the current vehicle speed, vehicle distance, pedestrian density and weather conditions, the gradient boosting decision tree (GBDT) is used to adjust the weight, which is expressed as:

[0160] S411, define the loss function:

[0161] Among them, y i is the true label, F m (x i ) is the predicted value of the mth tree, l is the squared error loss function, α is the learning rate, and h m (x i ) is the predicted output of the mth tree;

[0162] S412, Calculation of Feature Importance: I j =∑ t∈T ΔL t I(t);

[0163] Among them, I j is the importance score of feature j, ΔL t is the loss reduction brought by each split in the decision tree, I(t) is the indicator function of whether feature j participates in the split. If feature j participates in the node split of the decision tree, I(t) = 1, otherwise it is 0;

[0164] S413, Dynamic Weight Adjustment:

[0165] Among them, w j is the dynamically adjusted weight of feature j, and M is the number of all features;

[0166] S42, accident probability calculation: Based on the adjusted feature weights, the probability of an accident is calculated using a logistic regression model, expressed as:

[0167] ;

[0168] Among them, P (Accident|Features) is the probability of an accident, β0 is the bias term of the model, and β i For each feature x i The regression coefficient, w i is the dynamically adjusted weight;

[0169] S43, severity assessment: Based on the determination of the probability of an accident, the severity of the accident is assessed using a weighted summation algorithm by comprehensively analyzing the speed, vehicle distance, pedestrian density and weather conditions of traffic participants, expressed as:

[0170]

[0171] Where S is the final severity score, w i is the weight of the i-th input variable, indicating the influence of the variable on the severity of the accident. The sum of the weights is usually 1 (i.e. ), x i is the standardized value of the ith input variable, representing the relevant characteristics of the accident severity, such as speed, vehicle distance, pedestrian density, and weather conditions, and n is the total number of input variables;

[0172] Through the above steps, the feature weights are precisely adjusted through the gradient boosting decision tree algorithm, which improves the accuracy and reliability of accident probability and severity assessment, providing strong support for the early warning and handling of traffic accidents.

[0173] Data parsing and feature extraction in S51 include:

[0174] S511, extraction of ambient lighting conditions: Obtain a road image through a vehicle-mounted camera, and use a gray value average algorithm to extract the ambient lighting conditions, expressed as:

[0175]

[0176] Among them, L is the average brightness of the ambient light, R i ,G i ,B i are the red, green, and blue components of the i-th pixel, respectively, and N is the total number of pixels in the image;

[0177] S512, extraction of road conditions: using the image features in the visual data, through the image Canny edge detection algorithm and Gabor filter, the texture features of the road surface are extracted, and it is determined whether the road surface is slippery, potholes or waterlogged, which is expressed as:

[0178]

[0179] Where T is the characteristic intensity of road texture, I(i,j) is the pixel value of the image at position (i,j), M and N are the width and height of the image respectively;

[0180] S513, extraction of traffic signal status: obtaining the real-time traffic signal status through the V2X communication device, and using the color recognition algorithm to identify the color of the traffic light, expressed as:

[0181]

[0182] Among them, H is the hue component of the pixel, which is used to determine the color state of the traffic light;

[0183] S514, extraction of pedestrian behavior features: By combining the on-board camera with the convolutional neural network (CNN) model, the pedestrian behavior features are extracted, including walking direction, speed, and whether they suddenly cross the road, which are expressed as:

[0184] P f =σ(W·X+b);

[0185] Among them, P f is the feature vector of pedestrian behavior, W is the weight matrix, X is the input data (such as pedestrian position and speed), b is the bias, and σ is the ReLU activation function;

[0186] S515, vehicle type and feature extraction: The appearance features of the vehicle are extracted through visual data and a convolutional neural network (CNN) model, including vehicle type, color, and size, expressed as:

[0187] V f =CNN(I);

[0188] Among them, V f is the vehicle feature vector, I is the input image, and CNN extracts the type and appearance features of the vehicle through multi-layer convolution operations;

[0189] Through the above steps, key environmental and behavioral features can be accurately extracted from multi-source data, providing high-quality data input for subsequent traffic accident detection and virtual scene reconstruction.

[0190] The virtual accident generation model in S52 adopts the Generative Adversarial Network (GANs) model, which includes:

[0191] S521, multimodal data input and conditional fusion: Multimodal data (such as ambient lighting, road conditions, traffic signal status, etc.) is introduced as input and combined with the noise vector to ensure that the generated accident scene can match the actual traffic environment, expressed as:

[0192] z~p z (z);

[0193] Where z is a noise vector sampled from a standard normal distribution;

[0194] y=concat(y1,y2,...,y n );

[0195] Among them, y is the conditional vector, y1,y2,...,y n It is the standardized vector of each modal data (such as ambient lighting, road conditions), and concat means concatenating multiple modal data into a conditional vector;

[0196] z cond =concat(z,y);

[0197] Among them, z cond It is the fusion input used by the generator to generate accident scenarios;

[0198] S522, condition generator: the generator receives the fused input z cond , the virtual accident scenario that generates the conditional vector y is expressed as:

[0199] x fake =G(z cond ) = G(concat(z,y));

[0200] Among them, x fake is the generated virtual accident scenario, G is the generator;

[0201] S523, multi-scale discriminator: A multi-scale discriminator is introduced to discriminate the generated virtual accident scene at different scales (such as global scale and local scale), so as to ensure that the generated scene can meet the characteristics of the real accident in both global and detailed aspects, which is expressed as:

[0202] Multi-scale loss of the discriminator:

[0203] Among them, D s represents the discriminator at scale s, and E represents the expected value;

[0204] Multi-scale total loss:

[0205] Among them, λ s is the weight of each scale loss, S is the total number of scales;

[0206] S524, Combination of multi-scale generation loss and reconstruction loss: In the loss function of the generator, the multi-scale generation loss and reconstruction loss are combined to ensure that the generated scene is realistic at multiple scales and consistent with the characteristics of the real accident scene, expressed as:

[0207] Multi-scale generation loss:

[0208] Reconstruction loss:

[0209] Among them, x real It is a real accident scene, and the reconstruction loss ensures that the generated scene is consistent with the characteristics of the real scene;

[0210] Generator total loss:

[0211] Among them, λ recon is the weight coefficient of the reconstruction loss, and the total loss of the generator is calculated by weighted summation;

[0212] S525, adversarial training and updating: Use the adversarial training method to alternately update the generator and the discriminator, so that the generated accident scenes become more and more realistic, and the discriminator becomes increasingly difficult to distinguish between real and generated scenes, expressed as:

[0213] Generator Updates:

[0214] Among them, θ G is the parameter of the generator, η is the learning rate;

[0215] Discriminator update:

[0216] Among them, θ D is the parameter of the discriminator, η is the learning rate;

[0217] Through the above content, the generated accident scene can not only reflect the global traffic environment, but also be highly consistent with the real accident scene in local details, thereby providing more reliable and accurate data support for the analysis, prediction and simulation of traffic accidents. This multimodal and multi-scale generation method is particularly suitable for accident simulation in complex traffic scenarios, which helps to improve the accuracy and robustness of the system in practical applications.

[0218] The virtual accident scenario generation in S53 includes:

[0219] S531, feature fusion input: after standardization, the relevant features extracted from multi-source data are fused into a comprehensive feature vector;

[0220] S532, feature input generation model: the fused comprehensive feature vector is combined with the random noise vector and input into the virtual accident generation model as an input vector, so that an accident scene consistent with the actual traffic environment can be generated according to the input features;

[0221] S534, generating an accident scene: the virtual accident generation model generates a virtual accident scene according to the input comprehensive feature vector, including the whole scene before, during and after the accident, ensuring that the generated scene has a complete time series and details that conform to the actual accident characteristics;

[0222] Through the above steps, a virtual accident scene that is highly consistent with the real traffic environment can be generated, covering the entire process of the accident, and providing comprehensive data support for accident analysis and responsibility determination.

[0223] Scenario simulation and multi-condition analysis in S54 include:

[0224] S541, condition variable setting: based on the generated virtual accident scenario, multiple different condition variables are set, including different vehicle speeds, weather conditions, traffic densities, and vehicle types, to create different simulation scenario conditions;

[0225] S542, multi-condition scenario generation: input the set condition variables into the virtual accident generation model respectively, generate accident scenarios under different conditions, and each simulation scenario corresponds to a condition combination, so as to observe the influence of different conditions on the development of the accident, which is expressed as:

[0226]

[0227] in, is the conditional variable y i The fusion input with the noise z is used to generate accident scenarios under different conditions. is the conditional variable y i The accident scenario generated below;

[0228] S543, Factor Impact Analysis: By comparing the accident scenarios generated under different conditions, analyzing the impact of various conditional variables (such as vehicle speed changes, weather deterioration) on the severity and probability of accidents, identifying key factors and their influence on the development of accidents, expressed as:

[0229]

[0230] in, is the conditional variable y iThe severity of the accident assessed under base is the accident severity under baseline conditions, ΔS i is the conditional variable y i incremental impact on accident severity;

[0231] S544, Multi-path Accident Development Generation: Simulate and generate multiple accident development paths under different conditions, analyze the results of accidents under different circumstances, and provide reference for accident prediction and formulation of preventive measures, expressed as:

[0232]

[0233] in, is the accident in the condition variable y i The probability path of development is is the conditional variable y i Next, based on the generated scene The path probability of

[0234] Through the above content, it is possible to flexibly generate and analyze accident scenarios under different conditions, identify and evaluate various influencing factors, and generate multiple possible accident development paths, thereby providing data support and decision-making basis for accident prediction, management and prevention.

[0235] like Figure 2 As shown, a road traffic accident detection system based on roadside radar perception is used to implement the above-mentioned road traffic accident detection method based on roadside radar perception, including the following modules:

[0236] Radar sensor module: Through radar sensors installed on both sides of the road, the motion data of traffic participants are continuously collected, including speed, acceleration, location information and vehicle spacing;

[0237] Multi-source data fusion and preprocessing module: collects environmental perception data in real time through a variety of collection devices, including visual data, traffic flow data and communication data, and fuses and preprocesses environmental perception data with motion data;

[0238] Real-time dynamic analysis module: performs real-time analysis on pre-processed multi-source data, identifies abnormal behaviors of traffic participants, and predicts potential traffic accidents;

[0239] Accident identification and risk assessment module: Conduct risk assessment based on real-time dynamic analysis results, dynamically adjust the risk assessment weight, and determine the probability and severity of the accident;

[0240] Virtual accident scene reconstruction module: After a traffic accident is detected, the virtual accident generation model is used to analyze the collected multi-source data and reconstruct the virtual accident scene;

[0241] Alarm and information push module: According to the severity of the virtual accident scene, send alarm information to the traffic management department and emergency rescue units, and push accident warning information to users through road display screens and vehicle terminals.

[0242] The present invention covers any substitution, modification, equivalent method and scheme made on the essence and scope of the present invention. In order to make the public have a thorough understanding of the present invention, specific details are described in detail in the following preferred embodiments of the present invention, but those skilled in the art can fully understand the present invention without the description of these details. In addition, in order to avoid unnecessary confusion about the essence of the present invention, well-known methods, processes, procedures, components and circuits are not described in detail.

[0243] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A road traffic accident detection method based on roadside radar perception, characterized in that: The following steps are involved: S1, data collection: through radar sensors installed on both sides of the road, the movement data of traffic participants are continuously collected, including speed, acceleration, location information and vehicle spacing; S2, multi-source data fusion and preprocessing: collect environmental perception data in real time through a variety of collection devices, including visual data, traffic flow data and communication data, fuse the collected motion data and environmental perception data into multi-source data, and preprocess the fused multi-source data; S3, real-time dynamic analysis: real-time dynamic analysis of pre-processed multi-source data to identify abnormal behaviors of traffic participants and predict potential traffic accidents; S4, accident identification and risk assessment: Conduct risk assessment on the results of real-time dynamic analysis, dynamically adjust the weight of risk assessment according to vehicle speed, vehicle distance, pedestrian density and weather conditions, and determine the probability and severity of the accident; S5, virtual accident scene reconstruction: After a traffic accident is detected, the virtual accident generation model is used to analyze the collected multi-source data and reconstruct the virtual accident scene, including: S51, Data parsing and feature extraction: Extract relevant features from multi-source data, including ambient lighting conditions, road conditions, traffic signal status, pedestrian behavior characteristics, vehicle types and characteristics; S52, Model construction: Constructing a virtual accident generation model using historical multi-source data; S53, virtual accident scene generation: inputting the extracted relevant features into the virtual accident generation model to generate a virtual accident scene, including the entire scene before, during and after the accident; S54, scenario simulation and multi-condition analysis: Based on the virtual accident scenario, simulate the accident scenarios under different conditions, analyze the impact of various factors on the accident, and generate multiple accident development paths; S6, Alarm and information push: After completing the reconstruction of the virtual accident scene, alarm information will be automatically sent to relevant traffic management departments and emergency rescue units according to the severity of the traffic accident, and accident warning information will be pushed to users through road display screens and vehicle terminals.

2. A road traffic accident detection method based on roadside radar perception according to claim 1, characterized in that: The data collection in S1 includes: S11, radar sensor installation: multiple radar sensors are installed at predetermined intervals on fixed brackets on both sides of the road; S12, high-frequency pulse signal transmission: each radar sensor continuously scans the traffic participants in its coverage area by transmitting a high-frequency pulse signal to capture the movement of the target object; S13, radar signal reception and processing: The radar sensor receives radar signals reflected by traffic participants and calculates the motion data of each target in real time, including speed, acceleration, position information, and vehicle spacing.

3. The method for detecting road traffic accidents based on roadside radar perception according to claim 2, characterized in that: The multi-source data fusion and preprocessing in S2 include: S21, environmental perception data collection: collect environmental perception data in real time through a variety of collection devices, including: Visual data: Collect visual data through on-board cameras, including images on the road, pedestrian behavior, and lane marking status; Traffic flow data: Traffic flow data is collected through traffic flow meters, including traffic density, speed distribution, and vehicle types; Communication data: Communication data is collected through V2X communication equipment, including communication data between vehicles, traffic signal status, and emergency alarms; S22, data fusion: Adopt adaptive weighted average algorithm to fuse the collected motion data and environmental perception data into multi-source data, expressed as: D fused =w1·D motion +w2·D visual +w3·D traffic +w4·D communication ; Among them, D fused represents the fused multi-source data, D motion .D visual ,D traffic, D communication They represent motion data, visual data, traffic flow data, and communication data, respectively, and w1, w2, w3, and w4 are the corresponding weights; S23, data preprocessing: using Kalman filter to filter the fused multi-source data.

4. The method for detecting road traffic accidents based on roadside radar perception according to claim 1, characterized in that: The real-time dynamic analysis in S3 includes: S31, data input: input the preprocessed multi-source data into the convolutional neural network model to extract the spatiotemporal characteristics of traffic participants' behaviors and capture the behavior patterns of traffic participants; S32, abnormal behavior recognition: Based on the spatiotemporal features extracted by the convolutional neural network model, abnormal behavior is recognized by calculating the difference between the current behavior features and the normal behavior features using the Euclidean distance.

5. The method for detecting road traffic accidents based on roadside radar perception according to claim 1, characterized in that: The accident identification and risk assessment in S4 includes: S41, feature weight adjustment: Based on the results of real-time dynamic analysis, combined with the current vehicle speed, vehicle distance, pedestrian density and weather conditions, the gradient boosting decision tree is used to adjust the weight, expressed as: S411, define the loss function: ; Among them, y i is the true label, F m (x i ) is the predicted value of the mth tree, l is the squared error loss function, α is the learning rate, and h m (x i ) is the predicted output of the mth tree; S412, calculation of feature importance: ; Among them, I j is the importance score of feature j, is the loss reduction brought by each split in the decision tree, and I(t) is the indicator function of whether feature j participates in the split; S413, Dynamic Weight Adjustment: ; Among them, w j is the dynamically adjusted weight of feature j, and M is the number of all features; S42, accident probability calculation: Based on the adjusted feature weights, the probability of an accident is calculated using a logistic regression model, expressed as: ； Among them, P (Accident|Features) is the probability of an accident, β0 is the bias term of the model, and β i For each feature x i The regression coefficient, w i is the dynamically adjusted weight; S43, severity assessment: Based on the determination of the probability of an accident, the severity of the accident is assessed using a weighted summation algorithm by comprehensively analyzing the speed, vehicle distance, pedestrian density and weather conditions of traffic participants, expressed as: ； Where S is the final severity score, w i is the weight of the ith input variable, x i is the standardized value of the ith input variable, and n is the total number of input variables.

6. A road traffic accident detection method based on roadside radar perception according to claim 5, characterized in that: The data analysis and feature extraction in S51 include: S511, extracting ambient lighting conditions: acquiring a road image through a vehicle-mounted camera, and extracting ambient lighting conditions using a grayscale value average algorithm; S512, extraction of road surface conditions: using image features in visual data, through image Canny edge detection algorithm and Gabor filter, to extract road surface texture features, and determine whether the road surface is slippery, potholes or water accumulation; S513, extraction of traffic signal status: obtaining real-time traffic signal status through V2X communication equipment, and using color recognition algorithm to identify the color of the traffic light; S514, extraction of pedestrian behavior features: by combining the on-board camera with the convolutional neural network model, the behavior features of pedestrians are extracted, including walking direction, speed, and whether they suddenly cross the road; S515, vehicle type and feature extraction: Extract the appearance features of the vehicle, including vehicle type, color, and size, through visual data and convolutional neural network models.

7. A road traffic accident detection method based on roadside radar perception according to claim 6, characterized in that: The virtual accident generation model in S52 adopts a generative adversarial network model, and the generative adversarial network model includes: S521, multimodal data input and conditional fusion: Multimodal data is introduced as input and combined with the noise vector, expressed as: from~p z (With); Where z is a noise vector sampled from a standard normal distribution; y=concat(y1,y2,...,y n ); Among them, y is the conditional vector, y1,y2,...,y n is the standardized vector of each modal data, and concat means concatenating multiple modal data into a conditional vector; from cond = concat(z,y); Among them, z cond is the fusion input; S522, condition generator: the generator receives the fused input z cond , the virtual accident scenario that generates the conditional vector y is expressed as: x fake =G(z cond )=G(concat(z,y)); Among them, x fake is the generated virtual accident scenario, G is the generator; S523, multi-scale discriminator: a multi-scale discriminator is introduced to discriminate the generated virtual accident scene at different scales, which is expressed as: Multi-scale loss of the discriminator: Among them, D s represents the discriminator at scale s, and E represents the expected value; Multi-scale total loss: Among them, λ s is the weight of each scale loss, S is the total number of scales; S524, multi-scale generation loss combined with reconstruction loss: In the loss function of the generator, the multi-scale generation loss and reconstruction loss are combined, expressed as: Multi-scale generation loss: Reconstruction loss: Among them, x real It is a real accident scene; Generator total loss: Among them, λ recon is the weight coefficient of reconstruction loss; S525, adversarial training and updating: Use adversarial training method to update the generator and discriminator alternately, expressed as: Generator Updates: Among them, θ G is the parameter of the generator, η is the learning rate; Discriminator update: Among them, θ D are the parameters of the discriminator and η is the learning rate.

8. The method for detecting road traffic accidents based on roadside radar perception according to claim 7, characterized in that: The generation of the virtual accident scene in S53 includes: S531, feature fusion input: after standardization, the relevant features extracted from multi-source data are fused into a comprehensive feature vector; S532, feature input generation model: combining the fused comprehensive feature vector with the random noise vector and inputting the vector into the virtual accident generation model as an input vector; S534, generate accident scene: the virtual accident generation model generates a virtual accident scene according to the input comprehensive feature vector, including the whole scene before, during and after the accident.

9. A road traffic accident detection method based on roadside radar perception according to claim 8, characterized in that: The scenario simulation and multi-condition analysis in S54 include: S541, condition variable setting: based on the generated virtual accident scenario, multiple different condition variables are set, including different vehicle speeds, weather conditions, traffic densities, and vehicle types, to create different simulation scenario conditions; S542, multi-condition scenario generation: input the set condition variables into the virtual accident generation model respectively to generate accident scenarios under different conditions, which are expressed as: in, is the conditional variable y i The fused input with the noise z, is the conditional variable y i The accident scenario generated below; S543, Factor Impact Analysis: By comparing the accident scenarios generated under different conditions, the impact of each condition variable on the severity and probability of the accident is analyzed, and the key factors and their influence on the development of the accident are identified, expressed as: in, is the conditional variable y i The severity of the accident assessed under base is the accident severity under baseline conditions, ΔS i is the conditional variable y i incremental impact on accident severity; S544, Multi-path Accident Development Generation: Simulate and generate multiple accident development paths under different conditions, and analyze the results of the accident under different circumstances, expressed as: in, is the accident in the condition variable y i The probability path of development is is the conditional variable y i Next, based on the generated scene The path probability.

10. A road traffic accident detection system based on roadside radar perception, used to implement a road traffic accident detection method based on roadside radar perception as claimed in any one of claims 1 to 9, characterized in that: Includes the following modules: Radar sensor module: Through radar sensors installed on both sides of the road, the motion data of traffic participants are continuously collected, including speed, acceleration, location information and vehicle spacing; Multi-source data fusion and preprocessing module: collects environmental perception data in real time through a variety of collection devices, including visual data, traffic flow data and communication data, and fuses and preprocesses environmental perception data with motion data; Real-time dynamic analysis module: performs real-time analysis on pre-processed multi-source data, identifies abnormal behaviors of traffic participants, and predicts potential traffic accidents; Accident identification and risk assessment module: Conduct risk assessment based on real-time dynamic analysis results, dynamically adjust the risk assessment weight, and determine the probability and severity of the accident; Virtual accident scene reconstruction module: After a traffic accident is detected, the virtual accident generation model is used to analyze the collected multi-source data and reconstruct the virtual accident scene; Alarm and information push module: According to the severity of the virtual accident scene, send alarm information to the traffic management department and emergency rescue units, and push accident warning information to users through road display screens and vehicle terminals.

Images