A highway abnormal event type identification and position estimation method

Abstract

The application belongs to the field of intelligent transportation, and discloses a highway abnormal event type identification and position estimation method, which adopts a method combining BERT and LSTM models, can accurately identify different types of traffic abnormal events, and improves the accuracy of event type identification. Secondly, by using a CNN-GRU model and a K-Means clustering algorithm, the vehicle travel time and delay time can be accurately predicted, so that the initial position of traffic congestion can be accurately estimated. By comprehensively considering the terrain factors and traffic flow characteristics, the influence of different event types on the traffic flow is comprehensively analyzed, and the applicability and reliability of the model are improved. In addition, through real-time data analysis and dynamic adjustment, the model is adapted to complex traffic situations such as sudden large traffic flow. In summary, the application can improve the safety and efficiency of highway tunnel traffic.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of intelligent transportation, and specifically relates to a highway abnormal event type identification and position estimation method. BACKGROUND

[0002] With the development of economy, people's demand for travel quality is constantly improving, and highways have become the first choice for more and more people. However, the high frequency use of highways brings huge traffic pressure, and the existing traffic infrastructure has been difficult to meet the growing travel demand, leading to increasingly prominent traffic supply and demand contradictions and increasingly serious congestion problems. Common abnormal events on highways, such as traffic accidents, sudden heavy traffic, construction, etc., further exacerbate congestion and lead to frequent queuing. Therefore, how to quickly and accurately identify the type of abnormal events on the highway and estimate their occurrence position has become a key problem in the field of traffic management that needs to be solved urgently.

[0003] On highways, abnormal events such as traffic accidents, adverse weather, road construction, etc. frequently occur, which not only further exacerbate traffic congestion, but also easily cause secondary accidents, causing more serious traffic problems. In existing technologies, some patents propose methods for event detection and position estimation using a single data source, such as patent CN123456A relying on video monitoring data and patent CN234567B using ground sensor data. However, these methods are usually only applicable to certain specific types of events, and the accuracy of identification and position estimation for different types of events is low, making it difficult to cope with the complexity and diversity of abnormal events on highways.

[0004] In view of the dynamics and diversity of abnormal events on highways, an abnormal event occurrence position estimation model needs to be constructed. This model needs to combine multiple data sources, including ETC data, radar and vision fusion trajectory data, etc., and fully consider the influence of different event occurrence types. In the model construction process, after obtaining the event type, the single vehicle travel time threshold is obtained using linear features and historical travel time to more accurately estimate the abnormal event occurrence position and provide accurate input data for subsequent impact range analysis. The research and application of existing technologies in this aspect are not mature enough, and there is an urgent need for an efficient, multi-source data fusion abnormal event identification and position estimation method to improve the intelligent level of traffic management and improve the efficiency and emergency response capability of traffic management. SUMMARY

[0005] Therefore, the purpose of the present application is to provide a highway abnormal event type identification and position estimation method. The present application aims to solve the problem of poor applicability of existing methods, low accuracy of identification and position estimation for different types of events, and difficulty in coping with complex and diverse abnormal events on highways.

[0006] The application provides a highway abnormal event type identification and position estimation method, comprising the following steps:

[0007] S1. Collecting ETC data, trajectory data and camera data with event detection function of the highway, and pre-processing the data to obtain travel time of the vehicle, traffic event description and road segment characteristics for subsequent model training;

[0008] S2. Cleaning and word segmentation of the camera event description text by using the BERT model, extracting deep semantic features of the text, and processing traffic flow parameters by using the LSTM model, then inputting the text features extracted by the BERT model and the traffic flow features extracted by the LSTM model into the fully connected neural network FCNN for feature fusion and event type prediction;

[0009] S3. Extracting the slope and curvature features of the road segment by using the CNN, converting the slope and curvature features into time series to input the GRU model, then predicting the final single vehicle travel time by the fully connected layer based on the output of the GRU;

[0010] S4. Determining the travel time threshold of each road segment and vehicle type combination by using the K-Means clustering algorithm, and identifying traffic delay based on the comparison between the predicted travel time and the threshold;

[0011] S5. Analyzing the delay time by combining the difference between the predicted travel time and the actual travel time, determining the initial point of traffic congestion and estimating the incident location.

[0012] Further, the step S1 comprises the following sub-steps:

[0013] S1.1 ETC data preprocessing;

[0014] I. Obtaining the passing record of each vehicle from the ETC system of the highway, including vehicle code, vehicle type, gantry number, entry and exit toll station time and passing time;

[0015] II. Denoising the ETC data to remove invalid data and redundant data;

[0016] III. Spatiotemporal matching of the cleaned ETC data, matching the passing records of the vehicle in different road segments according to the vehicle code and passing time, calculating the travel time T i,j and the driving speed V i,j of each vehicle in each road segment;

[0017] T i,j =t exit -t entry

[0018] In the formula, Ti,j is the travel time of vehicle i on link j; t exit is the time of vehicle at the exit; t entry is the time of vehicle at the entrance;

[0019] S1.2 Trajectory data preprocessing;

[0020] I. Collect trajectory data of vehicles on highways through installed radars and cameras, including the position, speed and acceleration of vehicles;

[0021] II. Clean and filter the trajectory data, remove abnormal data points;

[0022] III. Integrate the trajectory data according to time sequence, form the trajectory record of each vehicle in different time period, and extract the corresponding link density; the extraction formula of link density is:

[0023]

[0024] Wherein, the density of missing link is completed by LWR traffic flow model;

[0025] S1.3 Camera data preprocessing;

[0026] I. Obtain traffic event data from cameras with event detection function, including event description, event type, event occurrence time and camera number;

[0027] II. Use natural language processing technology to clean and segment the event description text, remove irrelevant characters and noise words, and extract the key information of the event;

[0028] III. Associate the processed text data with ETC data and trajectory data to form complete traffic event records;

[0029] S1.4 Feature extraction and fusion;

[0030] I. Extract traffic flow parameters from cleaned ETC data, including the travel time and speed of each vehicle;

[0031] II. Extract trajectory features from trajectory data, including the driving trajectory, position change and speed change of vehicles;

[0032] III. Extract text features of event description from camera data, and combine traffic flow parameters and trajectory features to form comprehensive feature vectors;

[0033] S1.5 Data labeling;

[0034] According to the description of the traffic event and the actual occurrence, the data is labeled to determine the event type and the occurrence position of the record, and the labeled data is used as a label of the model for model training of event type identification and position estimation.

[0035] Further, in the step S1.2, the LWR traffic flow model adopts a hyperbolic partial differential equation to describe the space-time variation law of the traffic flow, which is written as:

[0036]

[0037] In the formula, ρ is the road density, which refers to the average number of vehicles per unit length of the road, and the unit is veh / km; is the time domain; is the space domain; ρ0 is the initial density; ρ up ,ρ dn are the upstream and downstream densities on the road, respectively, which are the boundary conditions of the LWR model; Q is the flow, which refers to the number of vehicles passing through a specific position of the road per unit time, and the unit is veh / h.

[0038] Further, the step S2 includes the following sub-steps:

[0039] S2.1 adopts the BERT model to clean and tokenize the camera event description text, and extracts the deep semantic features of the text;

[0040] That is, using the deep bidirectional characteristics of BERT, an embedding vector rich in context is generated for each word, which is expressed as:

[0041] Embeddings=BERT(TokenizedText)

[0042] In the formula, TokenizedText is the text after preprocessing and tokenization; Embeddings is the word embedding vector generated by the BERT model;

[0043] S2.2 uses LSTM to process the traffic flow parameters;

[0044] I. Preprocessing the traffic flow parameter data to adapt to the input requirements of the LSTM network;

[0045] Standardization: standardizing the traffic volume, speed, and travel time;

[0046]

[0047] In the formula, X is the original data; μ and σ are the mean and standard deviation of the data, respectively; X std is the standardized data;

[0048] Serializing: converting time series data into a format that the LSTM network can process;

[0049] II. Building the LSTM network model and training the LSTM network model with the pre-processed traffic flow parameter data;

[0050] The LSTM network model is:

[0051]

[0052] f t = σ(W f X t + U f h t-1 + b f )

[0053] i t = σ(W i X t + U i h t-1 + b i )

[0054] o t = σ(W o X t + U o h t-1 + b o )

[0055]

[0056] where i t is the input gate; f t is the forget gate; o t is the output gate; is the new candidate cell state; c t is the cell state; h t is the hidden state at the current time; W f , W i , W o , and W c are weight matrices for hidden layer calculation; U f , U i , U o , and U c are weight matrices; b f , b i , b c , and b o are bias vectors; σ is an activation function that controls data flow between [0, 1]; tanh is a nonlinear activation function that controls data within the range [-1, 1];

[0057] S2.3 Construct a fully connected neural network (FCNN) to fuse the text features extracted by BERT and the traffic flow features extracted by LSTM, and realize accurate recognition of different event types;

[0058] I. Fuse traffic flow features from LSTM and quantified highway linear indicators, combining traffic flow features F traffic with quantified slope I grade and curvature I curvature to form a comprehensive feature vector C;

[0059] C = [F traffic ; I grade ; I curvature ]

[0060] II. Design a FCNN containing an input layer, several hidden layers, and a Softmax output layer to process the fused feature vector and predict the event type;

[0061] The construction formula of FCNN is as follows:

[0062] H i = Re LU(W i H i-1 + b i )

[0063] P event = Soft max(W i H n-1 + b i )

[0064] where H0 = C; W i and b i are the weights and biases of each layer; H i is the output of the i-th hidden layer;

[0065] III. Use cross-entropy loss function to optimize FCNN;

[0066]

[0067] where y is the one-hot encoding vector based on the label extracted by BERT; P event.i is the probability distribution vector predicted by the model; N is the total number of event types;

[0068] IV. Use Adam optimization algorithm to adjust model parameters;

[0069]

[0070] where θ is the model parameter; α is the learning rate; m t and v tis a small constant to prevent division by zero;

[0071] V. Train the model, iteratively optimize the weight and bias terms until the model converges;

[0072]

[0073] where θ is the model parameter; η is the learning rate; is the gradient of the loss function with respect to the parameter.

[0074] Further, the step S3 comprises the following sub-steps:

[0075] S3.1 input the terrain image to the CNN model, extract the slope and curvature features of each road segment;

[0076] F CNN = CNN(I)

[0077] where I represents the input terrain image; F CNN represents the features extracted by CNN;

[0078] S3.2 After extracting the features from CNN, convert the extracted feature map F into a sequence S so as to input into the GRU model;

[0079] The conversion process involves a Transform function, Transform represents the process of converting the output feature map F of CNN into a sequence S suitable for input into GRU, the expression is as follows:

[0080] S = Transform(F)

[0081] S3.3 Assuming that at time t, the feature vector S t is input into the GRU model, the GRU model updates the hidden state h t of the feature vector S t to reflect the sequence information observed so far;

[0082] The feature vector S t is converted from the feature map processed by CNN, including the slope and curvature information of the road segment;

[0083] S3.4 Use a fully connected layer to get the final bicycle travel time prediction F(t) from the output of GRU.

[0084] Further, the step S4 comprises the following sub-steps:

[0085] S4.1 Extract the travel time data of each road segment and vehicle type combination;

[0086] S4.2 Apply K-Means clustering algorithm to divide the travel time data into different clusters;

[0087]

[0088] where C k is the kth cluster; μ k is the center of the kth cluster; K is the number of clusters; x is the data point; ||x-μ k || 2 is the distance between the data point and the cluster center.

[0089] S4.3 Calculate the travel time threshold for each cluster;

[0090] Set the travel time threshold for each road segment and vehicle type combination based on the clustering result; for each cluster C i , calculate the average travel time of all data points in the corresponding cluster The expression of the travel time threshold is:

[0091]

[0092] where v is the vehicle type; σ is the allowed time deviation.

[0093] Based on the theoretical basis of the multiple regression equation, a single vehicle travel time prediction model based on the influence of composite road segments is designed, and the formula of the weight parameter is updated as:

[0094]

[0095] where is the travel time of each road segment; is the influence degree of the ith slope quantified by VISSIM simulation results; is the influence degree of the ith curvature quantified by VISSIM simulation results; a, β are undetermined coefficients; θ and k are slope and curvature; V is the average speed of the road segment; v k is the one-hot encoding of the vehicle type; β1, β 2,i , β 3,j , β 4,k , β5 are regression coefficients; ∈ is the error term.

[0096] S4.4 Identify whether there is traffic delay based on the comparison between the predicted travel time and the threshold;

[0097] Estimate the travel time y t , the average travel time, and the average travel speed v 拥堵流 in the congestion state of each vehicle in order to calculate the difference between the real travel time y(t).

[0098] Further, the step S5 comprises the following sub-steps:

[0099] S5.1 Analyzing the trend of the vehicle delay time according to the traffic delay data identified in step S4;

[0100]

[0101] where Δt i represents the delay time of the ith vehicle; represents the predicted travel time; T ij k represents the travel time threshold of the ith road segment and the jth vehicle type combination;

[0102] S5.2 Calculating the difference between the predicted travel time and the set threshold;

[0103]

[0104] where ΔT represents the difference between the overall predicted travel time and the set threshold, represents the overall predicted travel time, T threshold represents the set travel time threshold;

[0105] S5.3 Determining the initial point of traffic congestion and estimating the incident location in combination with the trend of the delay time and the difference;

[0106]

[0107] where x 停车 is the distance from the target vehicle to the road segment entrance position to the road segment congestion position.

[0108] Advantages:

[0109] The application provides a high-efficiency and accurate expressway abnormal event type identification and position estimation scheme by combining deep learning technology and traffic engineering theory. At present, most of the research work on the scene of the incident is mainly concentrated in the congestion area estimation of urban roads, which mainly involves the area with high daily traffic flow. Therefore, more is for the prediction of the scene of the incident, and there is a lack of research on the position estimation after the occurrence of the event, especially without fully considering the dynamic change of the incident position under different event types. This limitation leads to the fact that the existing methods cannot accurately estimate the specific position after the occurrence of the event when dealing with abnormal events on the expressway, which affects the efficiency of traffic management and emergency response. First, the method combining BERT and LSTM model can accurately identify different types of traffic abnormal events, and improve the accuracy of event type identification. Secondly, the CNN-GRU model and K-Means clustering algorithm can be used to accurately predict the vehicle travel time and delay time, so as to accurately estimate the initial position of traffic congestion. By comprehensively considering the terrain factors and traffic flow characteristics, the influence of different event types on traffic flow is analyzed, and the applicability and reliability of the model are improved. In addition, through real-time data analysis and dynamic adjustment, the adaptability of the model to complex traffic situations such as sudden large traffic flow is enhanced.

[0110] Other advantages, objects, and features of the present application will be apparent to those skilled in the art from the following specification, and it is intended to be covered by the following claims. The objects and other advantages of the present application can be realized and attained by the below description. BRIEF DESCRIPTION OF DRAWINGS

[0111] Figure 1 The flowchart of the expressway abnormal event type identification and position estimation method of the present application is shown in the figure.

[0112] Figure 2 The BERT-LSTM-FCNN model structure diagram is shown in the figure.

[0113] Figure 3 The flowchart of the BERT-LSTM-FCNN algorithm is shown in the figure.

[0114] Figure 4 The flowchart of the incident position estimation algorithm based on ETC data is shown in the figure. DETAILED DESCRIPTION

[0115] In order to make the technical solutions, advantages and objectives of the present application clearer, the technical solutions of the embodiments of the present application will be described clearly and completely below in conjunction with the drawings of the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the described embodiments of the present application, all other embodiments obtained by a person of ordinary skill in the art without any inventive effort fall within the protection scope of the present application.

[0116] As shown in Figures 1-4 , the present application provides a highway abnormal event type recognition and position estimation method, comprising the following steps:

[0117] S1. Collecting ETC data, trajectory data and camera data with event detection function of the highway, and pre-processing the data to obtain travel time, traffic event description and road segment characteristics of the vehicle for subsequent model training;

[0118] S1.1 ETC data preprocessing;

[0119] I. Obtain the pass record of each vehicle from the ETC system of the highway, including vehicle code, vehicle type, gantry number, entry and exit toll station time and passing time;

[0120] II. Denoising the ETC data to remove invalid data and redundant data, such as invalid vehicle type data and redundant data of continuous multiple passes through the same gantry in a very short time interval;

[0121] III. Temporal and spatial matching of the cleaned ETC data, matching the pass records of the vehicle in different road segments according to the vehicle code and passing time, and calculating the travel time T i,j and the driving speed V i,j of each vehicle in each road segment;

[0122] T i,j = t exit -t entry

[0123] In the formula, T i,j is the travel time of vehicle i in road segment j; t exit is the time of the vehicle at the exit; t entry is the time of the vehicle at the entrance;

[0124] S1.2 Trajectory data preprocessing;

[0125] I. Collecting the trajectory data of the vehicle through the radar and camera installed on the highway, including the position, speed and acceleration of the vehicle;

[0126] II. Clean and filter trajectory data, remove abnormal data points caused by detection device failure or environmental interference;

[0127] III. Integrate trajectory data in time series, form trajectory records of each vehicle in different time periods, and extract corresponding road segment density; the extraction formula of road segment density is:

[0128]

[0129] Where the density of missing road segments is completed by LWR (Lighthill Whitham Richards, LWR) traffic flow model;

[0130] LWR traffic flow model uses hyperbolic partial differential equation to describe the space-time variation of traffic flow, which is written as:

[0131]

[0132] In the formula, ρ is the road density, which refers to the average number of vehicles per unit length of road, with the unit of veh / km; is the time domain; is the space domain; ρ0 is the initial density; ρ up ,ρ dn are the upstream and downstream densities on the road, respectively, which are the boundary conditions of the LWR model; Q is the flow, which refers to the number of vehicles passing through a specific position on the road per unit time, with the unit of veh / h.

[0133] S1.3 Camera data preprocessing;

[0134] I. Obtain traffic event data from cameras with event detection function, including event description, event type, event time and camera number;

[0135] II. Use natural language processing technology to clean and tokenize event description text, remove irrelevant characters and noise words, and extract key information of the event;

[0136] III. Associate the processed text data with ETC data and trajectory data to form complete traffic event records;

[0137] (1) Text cleaning: remove irrelevant characters in the text, such as special symbols, non-text content, etc.;

[0138] (2) Tokenization: Tokenization divides text into lexical units (Tokens), and BERT uses WordPiece algorithm for tokenization, which means it can effectively handle unknown words;

[0139] (3) Add special tokens: Add `[CLS]` and `[SEP]` tokens to the text, which represent the beginning and end of the text sequence, respectively.

[0140] (4) Generate attention masks: This is to let the model distinguish the focus of attention and the padding part.

[0141] S1.4 Feature extraction and fusion;

[0142] I. Extract traffic flow parameters from cleaned ETC data, including travel time and speed of each vehicle;

[0143] II. Extract trajectory features from trajectory data, including vehicle travel trajectory, position change and speed change;

[0144] III. Extract text features from camera event description, and combine traffic flow parameters and trajectory features to form a comprehensive feature vector;

[0145] S1.5 Data labeling;

[0146] According to the description and actual occurrence of the traffic event, the data is labeled to determine the type and location of the recorded event. The labeled data is used as a label for the model, which is used for event type identification and location estimation model training.

[0147] S2. Use BERT model to clean and tokenize camera event description text, extract deep semantic features of text, and use LSTM model to process traffic flow parameters, then input text features extracted by BERT model and traffic flow features extracted by LSTM model into fully connected neural network FCNN for feature fusion and event type prediction;

[0148] S2.1 Use BERT model to clean and tokenize camera event description text, extract deep semantic features of text;

[0149] I. Use BERT model to clean text data, remove irrelevant characters and special symbols such as HTML tags, special symbols, non-text content, etc., to ensure data purity;

[0150] II. Tokenize the text data, break the text into token units, use the WordPiece algorithm to handle unknown words, and add special tokens such as [CLS] (representing the beginning of the text) and [SEP] (representing the end of the text) to each text sequence. The WordPiece algorithm formula is as follows:

[0151] argmax (a,b) Count(ab)

[0152] where Count(a, b) represents the frequency of the subword pair (a, b) in the vocabulary;

[0153] III. Generate attention masks to mark which tokens need to be focused on by the model and which are padding parts, to ensure that the model can distinguish between actual text and padding when processing. Let the input sequence be tokens = [t1, t2,..., tn], and the padding part be padding = [p1, p2,..., pm], then the attention mask attention_mask can be represented as: n m

[0154] attention_mask = [1, 1,..., 1, 0, 0,..., 0]

[0155] where the first n elements are 1 and the last m elements are 0.

[0156] IV. Input the preprocessed text data into the BERT model to generate high-quality feature vectors that capture the deep semantic information and context dependency of the text;

[0157] That is, using the deep bidirectional characteristics of BERT, generate a context-rich embedding vector for each word, expressed as:

[0158] Embeddings = BERT(TokenizedText)

[0159] In the formula, TokenizedText is the preprocessed and tokenized text; Embeddings is the word embedding vector generated by the BERT model;

[0160] S2.2 Use LSTM to process traffic flow parameters;

[0161] I. Preprocess the traffic flow parameter data to adapt to the input requirements of the LSTM network;

[0162] Standardization: Standardize the traffic volume, speed, and travel time; The standardization process includes subtracting the mean and dividing by the standard deviation to ensure that data of different dimensions are comparable and to speed up model convergence;

[0163]

[0164] where X is the original data; μ and σ are the mean and standard deviation of the data, respectively; X std is the standardized data;

[0165] ​​Serialization: converting time series data into a format that the LSTM network can process; typically, this means organizing the data into sequences of time steps, each containing the traffic flow parameters at that time instant;

[0166] II. Building the LSTM network model and training the LSTM network model with the pre-processed traffic flow parameter data; the LSTM network avoids the long-term dependency problem through its special gating mechanism, making the model able to effectively learn the features in the time series data; the LSTM network model is:

[0167]

[0168] f t =σ(W f X t +U f h t-1 +b f )

[0169] i t =σ(W i X t +U i h t-1 +b i )

[0170] o t =σ(W o X t +U o h t-1 +b o )

[0171]

[0172] where i t is the input gate; f t is the forget gate; o t is the output gate; is the new candidate cell state; c t is the cell state; h t is the hidden state at the current time instant; W f , W i , W o , and W c are weight matrices for hidden layer calculation; U f , U i , U o , and U c are weight matrices; b f , b i , b c , and b ois the bias vector; σ is an activation function that controls the data stream between [0, 1]; tanh is a nonlinear activation function that controls the data in the range [-1, 1];

[0173] S2.3 Construct a full connection neural network FCNN to fuse the text features extracted by BERT and the traffic flow features extracted by LSTM, and realize accurate identification of different event types;

[0174] In order to fuse the text features extracted by BERT and the traffic flow parameters processed by LSTM together for distinguishing different types of traffic events, the present application constructs a full connection neural network (Full Connect Neural Network, FCNN). FCNN is a basic deep learning model composed of multiple full connection layers.

[0175] In this network, each neuron in a layer is connected to all neurons in the previous layer, and this dense connection enables the network to capture and learn complex features and patterns of input data. The core of the full connection layer is its weight matrix and bias vector b. Given the input vector x, the output y of the layer can be calculated by the following formula:

[0176] O=f(W·C+b)

[0177] Where f is the activation function, W is the weight matrix, and b is the bias term.

[0178] I. Fusion of traffic flow features from LSTM and quantified highway linear indicators, combined with traffic flow features F traffic and quantified slope I grade and curvature I curvature , to form a comprehensive feature vector C;

[0179] C=[F traffic ;I grade ;I curvature ]

[0180] II. Design FCNN containing input layer, several hidden layers and Softmax output layer to process the fused feature vector and predict event type;

[0181] The construction formula of FCNN is as follows:

[0182] H i =ReLU(W i H i-1 +b i )

[0183] P event =Softmax(W i H n-1+b i )

[0184] where H0=C; W i and b i are the weights and biases of each layer; H i is the output of the i-th hidden layer;

[0185] III. Optimizing FCNN with cross-entropy loss function;

[0186]

[0187] where y is the one-hot encoding vector of the label extracted based on BERT; P event.i is the probability distribution vector predicted by the model; N is the total number of event types;

[0188] IV. Adjusting model parameters using Adam optimization algorithm;

[0189]

[0190] where θ is the model parameter; α is the learning rate; m t and v t are the momentum and acceleration terms, respectively; ∈ is a small constant to prevent division by zero;

[0191] V. Training the model, iteratively optimizing the weight and bias terms until the model converges;

[0192]

[0193] where θ is the model parameter; η is the learning rate; is the gradient of the loss function with respect to the parameter.

[0194] S3. Extracting the slope and curvature features of the road segment through CNN, converting the slope and curvature features into time series input GRU model, and then predicting the final bicycle travel time through a fully connected layer based on the output of the GRU;

[0195] S3.1 Inputting the terrain image into the CNN model to extract the slope and curvature features of each road segment;

[0196] F CNN = CNN(I)

[0197] where I represents the input terrain image; F CNN represents the features extracted by CNN;

[0198] S3.2 After extracting the features from CNN, the extracted feature map F is converted into a sequence S so as to be input into the GRU model;

[0199] The transformation process involves a Transform function, which represents the process of transforming the output feature map F of the CNN into a sequence S suitable for input into the GRU. The expression is as follows:

[0200] S = Transform(F)

[0201] S3.3 Assume that at time t, the eigenvector S t (Eigenvector S) t The feature map (derived from the CNN-processed feature map, including information on the slope and curvature of the road segment) is input into the GRU model, which updates the feature vector S. t The hidden state h t This is to reflect the sequence information observed so far; the specific steps are as follows:

[0202] Update Gate Calculation:

[0203] z t =σ(W z ·[h t-1 ,S t ]+b z )

[0204] It determines the extent to which the model should retain the previous hidden states.

[0205] Reset Gate Calculation:

[0206] r t =σ(W r ·[h t-1 ,S t ]+b r )

[0207] It determines how much information about previous hidden states the model should ignore when computing new candidate hidden states.

[0208] Candidate Hidden State Calculation:

[0209]

[0210] The candidate hidden state is a temporary update that is used to update the current hidden state based on the result of the update gate.

[0211] Final hidden state update:

[0212]

[0213] Among them, S t The eigenvector at time t; zt is the update gate; r t is the reset gate; is the candidate hidden state; h t is the hidden state at time t; σ denotes the sigmoid function; ⊙ denotes element-wise multiplication; W z , W r , W are the weight matrices for the update gate, the reset gate, and the candidate state, respectively; b z , b r , b are the bias terms;

[0214] S3.4 Obtain the final bike travel time prediction F(t) from the output of the GRU using a fully connected layer;

[0215] F(t) = FC(h)

[0216] S4. Determine the travel time threshold for each road segment and vehicle type combination using the K-Means clustering algorithm, and identify traffic delays based on the comparison between the predicted travel time and the threshold;

[0217] S4.1 Extract the travel time data for each road segment and vehicle type combination;

[0218] S4.2 Apply the K-Means clustering algorithm to divide the travel time data into different clusters;

[0219] Determine the features on which the cluster analysis is based;

[0220] F1 = T 行程 / D 路段 : the average travel speed of vehicles on a specific road segment, used to reflect the travel characteristics of different vehicle types on that road segment.

[0221] F2 = V 类型 : the vehicle type, used as the basis for classification.

[0222] The k-means clustering aims to assign n observation values to k clusters, with each observation belonging to the cluster represented by the nearest mean (i.e., cluster center).

[0223] Cluster centers: C i = (c i1 , c i2 ), where i = 1, 2,..., k, C i1 represents the average travel time of a certain vehicle type under non-congestion conditions, while C i2 represents the average speed of that vehicle type.

[0224] Assignment rule: assign each data point to the nearest cluster center, minimizing the Euclidean distance of each point to the cluster center.

[0225]

[0226] Update cluster centers: Calculate the mean of all points in each cluster and update as new cluster centers.

[0227] where S i represents the set of all vehicles in the i-th cluster, x contains the travel time, speed, and other features of the vehicle that are directly related to congestion.

[0228]

[0229] where C k represents the k-th cluster; μ k represents the center of the k-th cluster; K represents the number of clusters; x represents the data point; ||x-μ k || 2 represents the distance from the data point to the cluster center;

[0230] S4.3 Calculate the travel time threshold for each cluster;

[0231] Based on the clustering results, set the travel time threshold for each road segment and vehicle type combination; for each cluster C i , calculate the average travel time of all data points in the corresponding cluster The expression of the travel time threshold is:

[0232]

[0233] where v is the vehicle type; σ is the allowed time deviation to consider fluctuations in different time periods and traffic conditions;

[0234] The present invention attempts to synthesize clustering data by considering the topographic features of road segments to obtain a more comprehensive understanding of the impact of different road traffic flow states. By collecting and analyzing topographic data, the present invention adjusts the clustering threshold to better reflect the actual impact of topography on driving speed. The traditional dynamic model defines the impact of slope and curve on travel time, considering the length of the road segment, the slope angle, the vehicle acceleration, the curve radius, and the acceleration due to gravity. Further, the adjusted clustering data is synthesized using the weighted average method, aiming to provide a comprehensive perspective for traffic management and planning, taking into account the impact of topographic factors on traffic flow speed, and the formula for the weight parameter is updated as:

[0235]

[0236] where, is the travel time for each road segment; is the i-th slope impact degree quantified by VISSIM simulation results; is the quantified i-th curvature impact degree; a, β are undetermined coefficients; θ and k are slope and curvature; V is the average speed of the road segment; v k is the one-hot encoding of the vehicle type; β1, β 2,i , β 3,j , β 4,k , β5 are regression coefficients; ∈ is the error term;

[0237] S4.4 Based on the comparison between the predicted travel time and the threshold value, identify whether there is traffic delay;

[0238] Estimate the travel time y t , the average travel time, and the average travel speed v 拥堵流 of each vehicle in order to calculate the difference between the real travel time y(t);

[0239] T 差异 = F(t) - y(t)

[0240] S5. Analyze the delay time, determine the initial point of traffic congestion and estimate the incident location by combining the difference between the predicted travel time and the actual travel time;

[0241] S5.1 Analyze the trend of vehicle delay time according to the traffic delay data identified in step S4;

[0242]

[0243] In the formula, Δt i represents the delay time of the i-th vehicle; represents the predicted travel time; T ij k represents the travel time threshold value of the i-th road segment and the j-th vehicle type combination;

[0244] S5.2 Calculate the difference between the predicted travel time and the set threshold value;

[0245]

[0246] In the formula, ΔT represents the difference between the overall predicted travel time and the set threshold value, represents the overall predicted travel time, T threshold represents the set travel time threshold value;

[0247] S5.3 Determine the initial point of traffic congestion and estimate the incident location by combining the trend and difference of delay time;

[0248]

[0249] In the formula, x 停车is a distance from the entrance position of the road segment to the congestion position of the road segment.

[0250] In this declaration, the above embodiments are only used to illustrate the technical solutions of the present application but not limit the present application. Although the present application is described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or equivalently replaced without departing from the purpose and scope of the present application, and all should be covered in the scope of the claims of the present application.

Claims

1. A method for identifying and estimating the location of abnormal events on highways, characterized in that, Includes the following steps: S1. Collect ETC data, trajectory data, and camera data with event detection function on the highway, and preprocess the data to obtain vehicle travel time, traffic event description and road segment characteristics for subsequent model training; S2. The BERT model is used to clean and segment the camera event description text, extract the deep semantic features of the text, and use the LSTM model to process traffic flow parameters. Then, the text features extracted by the BERT model and the traffic flow features extracted by the LSTM model are input into the fully connected neural network FCNN for feature fusion and event type prediction. S2.1 uses the BERT model to clean and segment the camera event description text, and extracts the deep semantic features of the text; This leverages BERT's deep bidirectional nature to generate a context-rich embedding vector for each word, expressed by the formula: Embeddings=BERT(Tokenized Text) In the formula, Tokenized Text is the preprocessed and segmented text; Embeddings are word embedding vectors generated by the BERT model. S2.2 Utilizes LSTM to process traffic flow parameters; I. Preprocess the traffic flow parameter data to adapt it to the input requirements of the LSTM network; Standardization: Standardize traffic flow, speed, and travel time; In the formula, X represents the original data; μ and σ are the mean and standard deviation of the data, respectively; X std It is standardized data; Serialization: Converting time series data into a format that can be processed by an LSTM network; II. Construct an LSTM network model and train the LSTM network model using preprocessed traffic flow parameter data; The LSTM network model is as follows: f t =σ(W f X t +U f h t-1 +b f ) i t =σ(W i X t +U i h t-1 +b i ) the t =σ(W o X t +U o h t-1 +b o ) In the formula, i t It is an input gate; f t It is the Gate of Oblivion; t It is an output gate; This is the state of a new candidate cell; c t It refers to the cellular state; h t It is the hidden state at the current moment; W f W i W o and W c The weight matrix is ​​used for calculating the hidden layer; U f U i U o and U c b is the weight matrix; f ,b i ,b c and b o σ is the bias vector; σ is the activation function that controls the data flow within the range [0,1]; tanh is the nonlinear activation function that controls the data flow within the range [-1,1]. S2.3 Construct a fully connected neural network FCNN to fuse the text features extracted by BERT and the traffic flow features extracted by LSTM, thereby achieving accurate identification of different event types; I. Integrate traffic flow features from LSTM and quantified highway linear indices, combined with traffic flow feature F traffic With quantified slope I grade and curvature I curvature , which constitute the comprehensive feature vector C; C=[F traffic ;I grade ;I curvature ] II. Design an FCNN consisting of an input layer, several hidden layers, and a Softmax output layer to process fused feature vectors and predict event types; The formula for constructing FCNN is as follows: A i =ReLU(W i A i-1 +b i ) P event =Softmax(W i H n-1 +b i )) In the formula, H0 = C; W i and b i These are the weights and biases of each layer; H i It is the output of the i-th hidden layer; III. Optimize FCNN using the cross-entropy loss function; In the formula, y i It is the one-hot encoded vector of the label extracted based on BERT; P event.i It is the probability distribution vector predicted by the model; N is the total number of event types; IV. Adjust model parameters using the Adam optimization algorithm; In the formula, θ represents the model parameters; α represents the learning rate; m t and v t These are the momentum and acceleration terms, respectively; ∈ is a small constant to prevent division by zero; V. Train the model, iteratively optimize the weights and biases until the model converges; In the formula, θ represents the model parameters; η represents the learning rate; This represents the gradient of the loss function with respect to the parameters. S3. Extract the slope and curvature features of the road segment through CNN, convert the slope and curvature features into time series inputs into the GRU model, and then predict the final single-vehicle travel time based on the output of GRU through a fully connected layer. S4. Use the K-Means clustering algorithm to determine the travel time threshold for each road segment and vehicle type combination, and identify traffic delays based on the comparison between the predicted travel time and the threshold; S5. Based on the difference between the predicted travel time and the actual travel time, analyze the delay time, determine the initial point of traffic congestion, and estimate the location of the incident.

2. The method for identifying and estimating the location of abnormal events on highways according to claim 1, characterized in that: Step S1 includes the following sub-steps: S1.1 ETC data preprocessing; I. Obtain the passage records of each vehicle from the ETC system of the highway, including vehicle code, vehicle type, gantry number, entry and exit time of toll station and passage time; II. Denoise the ETC data, removing invalid and redundant data; III. Perform spatiotemporal matching on the cleaned ETC data. Based on the vehicle code and passage time, match the vehicle's passage records on different road segments, and calculate the travel time T for each vehicle on each road segment. i,j and driving speed V i,j ; T i,j =t exit -t entry In the formula, T i,j Let i be the travel time of vehicle i on road segment j; t exit t is the time the vehicle spends at the exit. entry The time the vehicle spends at the entrance; S1.2 Trajectory data preprocessing; I. Vehicle trajectory data, including vehicle position, speed, and acceleration, is collected through radar and cameras installed on highways; II. Clean and filter the trajectory data, removing outlier data points; III. Integrate the trajectory data according to the time series to form the trajectory record of each vehicle in different time periods, and extract the corresponding road segment density; The formula for extracting road segment density is: The density of missing road segments is filled in using the LWR traffic flow model; S1.3 Camera Data Preprocessing; I. Obtain traffic event data from cameras with event detection capabilities, including event description, event type, event occurrence time, and camera number; II. Use natural language processing technology to clean and segment the event description text, remove irrelevant characters and noise words, and extract key information about the event; III. The processed text data is linked with ETC data and trajectory data to form a complete traffic incident record; S1.4 Feature Extraction and Fusion; I. Extract traffic flow parameters from the cleaned ETC data, including the travel time and speed of each vehicle; II. Extract trajectory features from trajectory data, including vehicle driving trajectory, position changes, and speed changes; III. Extract textual features describing events from camera data and combine them with traffic flow parameters and trajectory features to form a comprehensive feature vector; S1.5 data annotation; Based on the description and actual occurrence of traffic incidents, the data is labeled to determine the recorded event type and location. The labeled data is used as the model's label for training the model to identify the event type and estimate the location.

3. The method for identifying and estimating the location of abnormal events on highways according to claim 2, characterized in that: In step S1.2, the LWR traffic flow model uses hyperbolic partial differential equations to describe the spatiotemporal variation of traffic flow, written as: In the formula, ρ is the road density, which refers to the average number of vehicles per unit length of road, with the unit: veh / km; For the time domain; For spatial domain; ρ0 is the initial density; ρ up ,ρ dn , representing the density on the road and downstream respectively, are boundary conditions for the LWR model; Q is the flow rate, which refers to the number of vehicles passing through a certain location on the road per unit time, in veh / h.

4. The method for identifying and estimating the location of abnormal events on highways according to claim 3, characterized in that: Step S3 includes the following sub-steps: S3.1 Input the terrain image into the CNN model and extract the slope and curvature features of each road segment; F CNN =CNN(I) In the formula, I represents the input terrain image; F CNN This represents the features extracted using a CNN; S3.2 After extracting features from the CNN, the extracted feature map F is converted into a sequence S so that it can be input into the GRU model; The transformation process involves a Transform function, which represents the process of transforming the output feature map F of the CNN into a sequence S suitable for input into the GRU. The expression is as follows: S = Transform(F) S3.3 Assume that at time t, the eigenvector S t The feature vector S is fed into the GRU model, and the GRU model updates the feature vector S. t The hidden state h t This is to reflect the sequence information observed so far; eigenvector S t It is derived from the feature map processed by CNN, including information on the slope and curvature of the road segment; S3.4 uses a fully connected layer to obtain the final single-vehicle travel time prediction F(t) from the output of GRU.

5. The method for identifying and estimating the location of abnormal events on highways according to claim 4, characterized in that: Step S4 includes the following sub-steps: S4.1 Extract travel time data for each road segment and vehicle type combination; S4.2 applies the K-Means clustering algorithm to divide the travel time data into different clusters; In the formula, C k Represents the k-th cluster; μ k The k-th cluster center is represented by ||x-μ|. K represents the number of clusters; x represents a data point; ||x-μ| k || 2 This represents the distance from a data point to the cluster center; S4.3 Calculate the travel time threshold for each cluster; Based on the clustering results, a travel time threshold is set for each road segment and vehicle type combination; for each cluster C i Calculate the average travel time of all data points in the corresponding cluster. The expression for the travel time threshold is: In the formula, v is the vehicle type; σ is the allowable time deviation; Based on the theoretical foundation of multiple regression equations, a single-vehicle travel time prediction model based on the influence of composite road segments is designed, and the formula for its weight parameters is updated as follows: In the formula, T is the travel time for each road segment; To quantify the influence of the i-th slope using VISSIM simulation results; The degree of influence of the curvature of the j-th curve after quantification; a and β are undetermined coefficients; θ and CUR are the gradient and curvature; V is the average speed of the road segment; ve p This is the one-hot encoding for the p-th vehicle type; β1,β 2,i ,β 3,j ,β 4,p β5 is the regression coefficient; ∈ is the error term; S4.4 Identifies whether traffic delays exist by comparing predicted travel time with a threshold; Estimate the travel time y for each vehicle t Average travel time and average travel speed v under congested conditions. 拥堵流 This is to calculate the difference between the actual travel time y(t) and the actual travel time.

6. The method for identifying and estimating the location of abnormal events on highways according to claim 5, characterized in that: Step S5 includes the following sub-steps: S5.1 Analyze the trend of vehicle delay time based on the traffic delay data identified in step S4; In the formula, Δt i This represents the delay time of the i-th vehicle; Indicates the predicted travel time; This represents the travel time threshold for the i-th road segment and the j-th vehicle type combination; S5.2 Calculate the difference between the predicted travel time and the set threshold; In the formula, ΔT represents the difference between the total predicted travel time and the set threshold. T represents the overall predicted travel time. threshold This indicates the set travel time threshold; S5.3 Combine the trends and differences in delay times to determine the initial point of traffic congestion and estimate the location of the incident; In the formula, x 停车 It is the distance from the target vehicle to the entrance of the road segment to the congested location of the road segment.

Images