Urban road multi-lane real-time traffic state reconstruction method and device thereof

Abstract

This application discloses a method and apparatus for real-time traffic state reconstruction of multi-lane urban roads, belonging to the field of intelligent transportation technology. It matches vehicle license plate data with road network topology information to extract path information and turning history features; generates training soft labels through a soft label fusion mechanism; predicts the turning probability of individual vehicles using an LSTM (Long Short-Term Memory) network to obtain macroscopic turning flow; uses the matched vehicle as a reference point and a DNN (Deep Neural Network) to estimate the cumulative arrival flow of the lane from the reference point to the current time, and the corresponding cumulative arrival curve; reconstructs lane traffic speed using a Newell simplified motion wave model, introducing traffic speed as a constraint into the IDM (Intelligent Driving Model) to reconstruct the micro-level car-following trajectory of individual vehicles, thereby identifying lane-level real-time queuing status and traffic status; aiming to reconstruct continuous and detailed traffic status of the entire road segment from limited point detection data.

Description

Technical Field

[0001] This application belongs to the field of intelligent transportation technology, specifically relating to a method and apparatus for real-time traffic state reconstruction of multi-lane urban roads. Background Technology

[0002] Real-time, accurate, and complete traffic data is the foundation for refined management and control in intelligent transportation systems (ITS). It provides solid data support for traffic congestion assessment, signal timing optimization, and the formulation of multi-source control strategies, playing a crucial role in ITS. However, limited by the density and accuracy of urban road sensing devices, existing traffic data often suffers from uneven spatiotemporal distribution and limited information dimensions, failing to directly and comprehensively reflect traffic conditions. Therefore, reconstructing traffic conditions in missing spatiotemporal areas using traffic models or data-driven methods has become a hot topic in urban traffic research. However, how to achieve refined real-time reconstruction of multi-lane traffic conditions on urban road sections based on limited traffic sensing data from real-world road networks remains a pressing issue.

[0003] Existing research on traffic state reconstruction mainly includes two implementation paths. The first is macroscopic group traffic state reconstruction, which typically utilizes macroscopic traffic flow aggregation parameters such as flow rate, speed, and density of road sections or segments. It then uses probabilistic statistical methods, machine learning, and macroscopic traffic flow theory to simulate the evolution and propagation of traffic flow, thereby reconstructing the traffic state. The second is microscopic individual traffic state reconstruction, which focuses on using partially known vehicle trajectory data and employing machine learning methods, car-following models, lane-changing models, and other techniques to infer the complete trajectory of individual vehicles in missing road segments or time periods. Although there have been explorations combining macroscopic and microscopic methods in existing technologies, these studies currently mainly focus on highway scenarios with relatively simple structures and relatively uniform traffic flow characteristics. For urban road networks with complex structures, numerous interfering factors, and highly heterogeneous traffic flow dynamics, existing research still lacks effective and universally applicable deep fusion reconstruction schemes, making it difficult to balance overall accuracy and local precision in the reconstruction results.

[0004] The existing technology has the following technical defects: (1) Macro-group traffic reconstruction methods focus on characterizing the overall traffic characteristics of roads and capturing the trend of macro-traffic flow changes. They are difficult to capture traffic fluctuations caused by the differentiated characteristics of vehicles, which can easily lead to distortion of local traffic state estimation.

[0005] (2) The micro-level individual traffic state reconstruction method focuses on modeling the driving process and trajectory of individual vehicles and expressing the traffic state through vehicle trajectories. However, the constraint relationship between adjacent vehicles makes it easy for the trajectory estimation error of individual vehicles to accumulate. That is, the estimation error of the trajectory of the preceding vehicle may accumulate on the estimated trajectory of the subsequent vehicle. In the case of long-term trajectory reconstruction, the accumulated estimation error is likely to cause the overall state to be inaccurate.

[0006] (3) Existing traffic state reconstruction methods that integrate macro and micro features are mainly focused on highway scenarios, ignoring the more complex vehicle turning needs and signal phase control constraints at intersections in urban roads. The real-time reconstruction accuracy needs to be further improved. Summary of the Invention

[0007] To address the technical problem of incomplete traffic condition perception, this application provides a method and apparatus for real-time traffic condition reconstruction of multi-lane urban roads, aiming to reconstruct the continuous and detailed traffic condition of the entire road segment, especially vehicle trajectories, from limited point detection data.

[0008] The technical solution is as follows: On the one hand, a method for real-time traffic status reconstruction of multi-lane urban roads is provided, including the following steps: Step S1: Obtain vehicle license plate data and road network topology information, match the vehicle license plate data with the road network topology information, and extract the path information of individual vehicles in the road network and the turning history features at each intersection. Step S2: Based on the path information and turning history characteristics of individual vehicles, training soft labels are generated through a soft label fusion mechanism. The turning probability of individual vehicles at downstream intersections is predicted through a Long Short-Term Memory (LSTM) network. The turning probabilities are accumulated to obtain the macroscopic turning flow. Step S3: Using the matched vehicle as the reference point, estimate the lane cumulative arrival flow from the reference point to the current time using a deep neural network (DNN), and simultaneously estimate the corresponding lane cumulative arrival curve; Step S4: Based on the downstream cumulative departure curve and the estimated lane cumulative arrival curve, the Newell simplified motion wave model is used to reconstruct the lane's full-time and spatiotemporal macroscopic traffic speed. The macroscopic traffic speed is introduced as a constraint into the intelligent driving model (IDM) to reconstruct the microscopic car-following trajectory of individual vehicles. Based on the reconstructed microscopic car-following trajectory, the lane-level real-time queuing status is identified and the lane-level real-time traffic status is reconstructed by fusing macroscopic and microscopic data.

[0009] Furthermore, in step S1, the path information includes: the intersection approach lane number and the corresponding sequence of passage times; The vehicle license plate data includes: license plate detection equipment number, intersection number, vehicle passage time, license plate data, license plate type, exit from turning lane, and direction of approach lane.

[0010] Furthermore, in step S1, the turning history features include: Historical turning probability of individual vehicles at intersections; The distribution of vehicles turning at the intersection in the previous time period; For vehicles without historical turning records, the vehicle turning distribution at the intersection in the previous time period is used.

[0011] Furthermore, in step S2, training soft labels are generated through a soft label fusion mechanism, and the calculation process is as follows: Soft tags are generated by weighting the route information and steering history characteristics of individual vehicles: ; in, A soft tag for individual vehicles. A hard label for individual vehicles. For the steering history characteristics of individual vehicles, This refers to hard label weights.

[0012] Furthermore, in step S2, the turning probability of individual vehicles is predicted using a Long Short-Term Memory (LSTM) network, and the turning probabilities are accumulated to obtain the macroscopic turning flow rate. The calculation process is as follows: The discrete indices of the upstream inlet channel are converted into dense vector representations through a word embedding layer; The output of the word embedding layer is concatenated with the temporal feature information according to the feature dimension to generate a path time vector of dimension T×d. This vector is then input into the LSTM layer to extract the hidden temporal associations of the path information, where T is the time step length (in this embodiment, it is the path sequence length, i.e., the number of upstream intersection entrance lanes passed by the vehicle). d For feature dimensions: Computing the forget gate in Long Short-Term Memory (LSTM) networks : ; in, for , The number of hidden layer units. Here is the weight matrix for the forget gate. For the offset of the forget gate, This is the hidden state from the previous moment. Let the input vector be the input vector at the current time. for function; Computing the input gate in a Long Short-Term Memory (LSTM) network : ; in, This represents the weight matrix of the input gate. This represents the bias term of the input gate; Generate candidate cell state information: ; in, For activation function, w c The weight matrix represents the candidate cells. b c The bias term representing candidate cells; Update cell status: ; Computing the output gate of a Long Short-Term Memory (LSTM) network : ; in, w o This represents the weight matrix of the output gate; This represents the bias term of the output gate; Output the hidden state at the current moment. : ( ); The hidden state and historical features of the last time step output of the Long Short-Term Memory (LSTM) network are input into the fully connected layer. The vehicle historical turning data is encoded as [1,0], and the vehicle turning data of the previous time step at the intersection is encoded as [0,1]. The output of the fully connected layer is transformed into normalized turning probabilities through a Softmax layer. ; in, For category; This is the output of the fully connected layer; Total number of categories; Calculate the loss during training using KL divergence: ; in, The total number of samples; This represents the true probability distribution; To predict the probability distribution; and Represent the true distribution and the predicted distribution in the category, respectively. The probability of it; The predicted turning probabilities of all individual vehicles within the study period are summed to obtain the macroscopic turning flow.

[0013] Furthermore, in step S3, using the matched vehicle as a reference point, the cumulative lane arrival flow from the reference point to the current time is estimated using a deep neural network (DNN), including: Matching vehicles by matching license plate data at upstream and downstream intersections Using the baseline, calculate the cumulative arrival flow of the lane at the current time: ; ; in, for Time Lane The cumulative arrival flow, Number the downstream lane. Number the upstream inlet channel; To match vehicles The corresponding cumulative number of arriving vehicles; This refers to the number of upstream import channels; This is a mapping function between feature information and lane arrival flow. for and Between, leaving the upstream entrance road Traffic; For upstream import channels and lane Nonlinear mapping function between; for and Within a given time period, the turning flow obtained from the aggregation of turning probabilities; For upstream import channels In the departing traffic, lanes are allocated Traffic on the internet For lane The corresponding downstream shift, For upstream import channels The departing traffic, For upstream import channels In the departing flow, Steering ratio in direction, For upstream import channels The resulting turning flow In the middle, assigned to lane Traffic share on; Using deep neural networks (DNNs) to fit a nonlinear mapping relationship between features and lane arrival flow. The arrival flow rate within the lane is obtained, where a deep neural network (DNN) employs the ReLU activation function and sets dropout parameters, using the root mean square error (RMSE) as the loss function. ; in, The total number of samples, For the first The estimated arrival flow for each sample This corresponds to the actual arrival traffic.

[0014] Further, in step S3, the corresponding lane cumulative arrival curve is estimated, including: The original cumulative arrival curve is smoothed using the longest increasing subsequence (LIS) method, by defining... and Given two arrays, iterate through the sequence and update the state for each element in the sequence. Using binary search in the array Find the largest , making If such a thing exists, it is true. Then update Assign the value to the current index. At the same time, the predecessor index of the current element is... Assigned value via array Obtain the last index of the longest increasing subsequence, and then use the array... By tracing the entire subsequence in reverse, the smoothed monotonically increasing cumulative arrival curve can be obtained; in, Used to record a length of The index of the smallest last element of the increasing subsequence in the original sequence. Record the predecessor index of each element for subsequent reconstruction of the increasing subsequence. The length of the subsequence. To use the current element index as the length The last element of an increasing subsequence.

[0015] Furthermore, in step S4, based on the downstream cumulative departure curve and the estimated lane cumulative arrival curve, the Newell simplified motion wave model is used to reconstruct the lane's all-time and all-space macroscopic traffic speed, including: Based on the principles of Newell's simplified motion wave model and the basic diagram of traffic triangles, the cumulative number of vehicles passing through the middle of the road is derived through boundary condition constraints: ; ; ; in, For upstream cumulative arrival curve function; This is the downstream cumulative departure curve function; This is the cumulative number of vehicles passing through, calculated based on upstream and downstream boundary conditions. The wave speed in free flow; The wave velocity is L in the congested flow state; L is the road length. For time; Indicates the location of the cross-section. Traffic density; like If the value is the minimum, then this point is governed by the upstream boundary conditions and is in a free-flow state. If the value is the minimum, then this point is governed by the downstream boundary conditions and is in a congested flow state. According to the principle of the basic traffic triangle diagram, the macroscopic velocity of the lane is obtained as follows: ; ; ; in, Indicates the location and time Vehicle density, This indicates the corresponding traffic flow. To represent a time interval, take , To represent spatial interval, take .

[0016] Furthermore, in step S4, macroscopic traffic speed is introduced as a constraint into the Intelligent Driving Model (IDM) to reconstruct the micro-car-following trajectory of individual vehicles. Based on the reconstructed micro-car-following trajectory, lane-level real-time queuing status and lane-level real-time traffic status are identified, including: Perform macro-micro fusion vehicle status updates to reconstruct the micro-level car-following trajectory of individual vehicles: ; in, To obtain the final speed through a weighted fusion of macro and micro factors; To optimize weighting, the weights are dynamically adjusted based on the distance between the vehicle and the vehicle in front. This represents a macroscopic speed constraint. Indicates macroscopic speed, This indicates that the vehicle status is updated based on IDM. Based on the reconstructed micro-following trajectories of individual vehicles, lane-level real-time queuing status recognition and traffic status reconstruction are performed: ; in, This indicates the position of the vehicle at the end of the convoy. This indicates the stop line position.

[0017] On the other hand, a device for real-time traffic state reconstruction of multi-lane urban roads is provided, comprising: The data acquisition and matching module is used to acquire vehicle license plate data and road network topology information, match the vehicle license plate data with the road network topology information, and extract the path information of individual vehicles in the road network and the turning history characteristics at each intersection. The turning flow prediction module is used to generate training soft labels based on the path information and turning history characteristics of individual vehicles through a soft label fusion mechanism, predict the turning probability of individual vehicles through a long short-term memory network (LSTM), and accumulate the turning probabilities to obtain the macroscopic turning flow. The lane cumulative arrival curve estimation module uses the matched vehicle as a reference point and a deep neural network (DNN) to estimate the lane cumulative arrival flow from the reference point to the current time, while also estimating the corresponding lane cumulative arrival curve. The macro-micro fusion traffic state reconstruction module is used to reconstruct the lane's all-time macro traffic speed based on the downstream cumulative departure curve and the estimated lane cumulative arrival curve, using the Newell simplified motion wave model. The macro traffic speed is introduced as a constraint into the intelligent driving model (IDM) to reconstruct the micro car-following trajectory of individual vehicles. Based on the reconstructed micro car-following trajectory, the real-time lane-level queuing status is identified and the macro-micro fusion lane-level real-time traffic state is reconstructed.

[0018] The technical solution includes at least the following technical effects: 1. In-depth utilization of vehicle license plate data: By matching vehicle license plates, the driving path and turning history characteristics of individual vehicles can be obtained, which forms the basis for high-precision turning prediction and trajectory reconstruction; 2. A macro-micro integrated reconstruction framework is proposed: Micro-initiation: First, use LSTM to predict the steering probability of individual vehicles; Macro transition: Aggregate micro-turn probabilities into macro-turn flow, and then use DNN to estimate lane cumulative arrival curve; Macro-micro closed loop: The full-time-space velocity is reconstructed using the Newell macro model, and this macro velocity is then used as a constraint to feed back into the micro IDM car-following model to correct and generate more reasonable individual vehicle trajectories.

[0019] 3. This application proposes a method and apparatus for real-time traffic state reconstruction of multi-lane urban roads. By mining the path information of individual vehicles and feeding back the reconstruction results of the macroscopic motion wave model as constraints to the microscopic car-following model, it achieves integrated and high-precision reconstruction of macroscopic and microscopic traffic states. This technical solution differs significantly from existing traffic light control schemes and traffic flow prediction schemes in terms of core tasks, data foundation, technical path, and model fusion mechanism.

[0020] It should be understood that the above general description and the following detailed description are merely exemplary and do not limit this application. Attached Figure Description

[0021] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0022] Figure 1 is a flowchart of a method for real-time traffic state reconstruction of multi-lane urban roads according to a preferred embodiment of this application; Figure 2 A schematic diagram of a multi-lane real-time traffic status reconstruction device for urban roads provided in a preferred embodiment of this application; Figure 3 This is a schematic diagram of vehicle steering prediction provided in a preferred embodiment of this application; Figure 4 This is a schematic diagram of arrival and departure on a multi-vehicle road segment provided in a preferred embodiment of this application; Figure 5 A schematic diagram of cumulative arrival flow based on matched vehicles provided in a preferred embodiment of this application; Figure 6 A schematic diagram illustrating the cumulative number of vehicles passing through according to a preferred embodiment of this application; Figure 7a A predicted left-turn traffic flow diagram of the west entrance of test intersection 3 provided in a preferred embodiment of this application; Figure 7b A downstream traffic flow prediction map for the west entrance lane of intersection 3 to test the straight-through traffic flow. Figure 8 The test intersection 3 west entrance straight lane 3 provided in a preferred embodiment of this application has the morning and evening peak hour cumulative arrival curves; Figure 9 The test intersection 3 west entrance straight lane 3 provided in a preferred embodiment of this application provides a smoothed estimated cumulative arrival curve for morning and evening peak hours; Figure 10 Spatiotemporal map of morning and off-peak speed distribution of lane 3 at the west entrance of test intersection 3 provided in a preferred embodiment of this application; Figure 11 This application provides a preferred embodiment of the reconstructed trajectory diagram of the straight lane 3 at the test intersection 3. Figure 12 This is a reconstruction result diagram of the evening rush hour queue length of 3 straight lanes at a test intersection provided in a preferred embodiment of this application; Detailed Implementation Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0023] Terminology Explanation: Traffic dynamic wave model: Based on the basic triangle diagram assumption, it considers different traffic waves under free flow and congested conditions, and uses the boundary conditions at both ends of the road segment to estimate the traffic state at any time and space within the road segment.

[0024] Intelligent Driving Model (IDM): A typical micro-vehicle car-following model that quantifies the driver's decision-making process as a continuous function. It aims to describe how the driver adjusts their acceleration based on the dynamic information of the vehicle in front and road conditions during driving, thereby achieving smooth and safe car-following behavior.

[0025] Cumulative arrival / departure curves: A time-varying flow rate graph describing the arrival and dissipation of all vehicles at the intersection.

[0026] As shown in Figure 1, in a preferred embodiment, a method for real-time traffic state reconstruction of multi-lane urban roads includes the following steps: Step S1: Obtain vehicle license plate data and road digital geographic information, match and map the vehicle license plate data with the actual road network, and extract the path information of individual vehicles in the road network and the turning history characteristics of vehicles at each entrance lane (intersection). The specific process is as follows: Step S11: Obtain vehicle license plate data through license plate recognition equipment, including license plate detection equipment number, intersection number, vehicle passage time, license plate data, license plate type, exit from turning lane, and approach lane direction information.

[0027] Step S12: Based on the latitude and longitude coordinates of the license plate recognition device, obtain the digital road network intersection information to which the device belongs through map matching, and complete the relationship matching between the license plate recognition device number and the digital road network intersection entrance lane by detecting the entrance lane direction information in the device configuration file.

[0028] Step S13: Generate a sequence of vehicle license plate recognition device detections in the road network based on time sequence, convert the license plate recognition device numbers into digital road network entrance lane numbers, and obtain the driving path of individual vehicles in the road network. The path information includes the intersection entrance lane numbers and the corresponding passage time sequence. To facilitate model learning of input features, the vehicle driving path sequence adopts a uniform length, using the passage information of the 5 closest intersection entrance lanes in time as the vehicle's path sequence. If the original path sequence length is less than 5, it is padded with 0 values.

[0029] Step S14: Use the historical turning probability of individual vehicles at the intersection and the vehicle turning distribution at the intersection in the previous time period as historical features. The turning probability distribution of all vehicles passing through the downstream intersection in the historical data is as follows: in, For vehicle license plates; This refers to the number of left turns a vehicle makes at intersections downstream of the road segment. The number of straight trips. This represents the number of right turns.

[0030] For vehicles without historical turning records, the vehicle turning distribution at the intersection in the previous time period is used as the historical feature, with a time period length of 15 minutes.

[0031] Step S2: Based on the path information and turning history characteristics of individual vehicles, the actual turning behavior of vehicles and prior historical information are fused to generate training soft labels. The Long Short Memory Neural Network (LSTM) is used to predict the turning probability of individual vehicles arriving from upstream at downstream intersections. The turning probabilities are accumulated to obtain the turning flow, and then the micro turning probabilities are aggregated into macro turning flow. The specific process is as follows, such as Figure 3 As shown: Step S21: Classify vehicle turns at downstream intersections into four categories: left turn, straight, right turn, and others (long-term roadside parking and turning out from a side road in the middle of the road segment). Since other types of vehicles cannot be detected by downstream license plate recognition equipment, and right-turn lanes at intersections are usually not equipped with license plate recognition equipment (i.e., right-turning vehicles are not detected), right-turning vehicles and other vehicles are grouped together under the right-turn label. Therefore, left turns, right turns, and straight-ahead are coded as [1,0,0], [0,1,0], and [0,0,1], respectively. Integrating individual behavioral specificities (vehicle paths) and group statistical patterns (historical characteristics), soft labels are generated as follows: ; in, For individual vehicles, hard labels (steering truth values, such as [1,0,0] for left turn). For the historical characteristics of individual vehicles, Hard label weights are used to control the balance between individual and group information.

[0032] Step S22: The route sequence information of individual vehicles is as follows ,in For the downstream intersection entrance lane, This indicates the route of the target vehicle, starting from the intersection entrance lane. Counting from upstream Vehicle passage information at each intersection's approach lanes, including the path sequence information containing the index sequence of the upstream approach lanes. and through time series .

[0033] The discrete indices of the upstream entrance lanes are converted into dense vector representations through an embedding layer. The embedding matrix has a dimension of V×d (V is the total number of index sequences, and d is the embedding dimension). Each intersection entrance lane index corresponds to a row vector in the matrix, and its value is dynamically optimized during model training.

[0034] The output of the word embedding layer is concatenated with the temporal feature information according to the feature dimension to generate a path temporal vector with dimension T×d (T is the time step length, which in this embodiment is the path sequence length, i.e. the number of upstream intersection entrance lanes passed by the vehicle, and d is the feature dimension). This vector is then input into the LSTM layer to extract the hidden temporal associations of the path information.

[0035] Computing the forget gate in LSTM : ; in, for , The number of hidden layer units. Here is the weight matrix for the forget gate. For the offset of the forget gate, This is the hidden state from the previous moment. Let the input vector be the input vector at the current time. for function.

[0036] Computing the input gate in a Long Short-Term Memory (LSTM) network : ; in, This represents the weight matrix of the input gate. This represents the bias term of the input gate; Generate candidate cell state information: ; in, For activation function, w c The weight matrix represents the candidate cells. b c The bias term representing candidate cells; Update cell status: ; Computing the output gate of a Long Short-Term Memory (LSTM) network : ; in, w o This represents the weight matrix of the output gate; This represents the bias term of the output gate; Output the hidden state at the current moment. : ( ); The hidden state and historical features of the last time step output of the Long Short-Term Memory (LSTM) network are input into the fully connected layer. The vehicle historical turning data is encoded as [1,0], and the vehicle turning data of the previous time step at the intersection is encoded as [0,1]. The output of the fully connected layer is transformed into normalized turning probabilities through a Softmax layer. ; in, For category; This is the output of the fully connected layer; This represents the total number of categories.

[0037] Calculate the loss during training using KL divergence: ; in, The total number of samples; This represents the true probability distribution; To predict the probability distribution; and Represent the true distribution and the predicted distribution in the category, respectively. The probability of it.

[0038] Step S23: Accumulate the predicted turning probabilities of all individual vehicles within the study period to obtain the macroscopic turning flow.

[0039] Step S3: Using the matched vehicle as the reference point, and taking the departure flow of the upstream entrance lane and the predicted turning flow as input, a deep neural network (DNN) is used to estimate the cumulative arrival flow of the lane from the reference point to the current time. At the same time, the cumulative arrival curve of the lane is estimated, and the longest increasing subsequence algorithm is used to correct the estimated cumulative arrival curve. The specific process is as follows, such as Figure 4 and Figure 5 As shown: Step S31: Designate the downstream exit lane as a dedicated vehicle lane. The arrival flow of vehicles exiting from the corresponding lane downstream is taken as the lane's arrival flow. Considering the lack of license plate recognition equipment and potential missed or incorrect detections at the upstream intersection's right-turn direction, matching vehicles whose license plate data can be matched at both upstream and downstream intersections is considered. Using this as a reference point, calculate the cumulative arrival flow of the lane at the current time. : ; ; in, for Time Lane The cumulative arrival flow, Number the downstream lane. Number the upstream inlet channel; To match vehicles The corresponding cumulative number of arriving vehicles; This refers to the number of upstream import channels; This is a mapping function between feature information and lane arrival flow. for and Between, leaving the upstream entrance road Traffic; For upstream import channels and lane Nonlinear mapping function between; for and Within a given time period, the turning flow obtained from the aggregation of turning probabilities; For upstream import channels In the departing traffic, lanes are allocated Traffic on the internet For lane The corresponding downstream shift, For upstream import channels The departing traffic, For upstream import channels In the departing flow, Steering ratio in direction, For upstream import channels The resulting turning flow In the middle, assigned to lane The percentage of traffic on the platform.

[0040] Step S32: Use a deep neural network (DNN) to fit a nonlinear mapping relationship between features and lane arrival flow. The arrival flow rate within the lane is obtained, where a deep neural network (DNN) employs the ReLU activation function and sets dropout parameters, using the root mean square error (RMSE) as the loss function. ; in, The total number of samples, For the first The estimated arrival flow for each sample This corresponds to the actual arrival traffic.

[0041] Step S33: Obtain lane cumulative arrival flow and generate lane cumulative arrival curve. Smooth the original cumulative arrival curve using the longest increasing subsequence (LIS) method, by defining... and Given two arrays, iterate through the sequence and update the state for each element in the sequence. Using binary search in the array Find the largest , making If such a thing exists, it is true. Then update Assign the value to the current index. At the same time, the predecessor index of the current element is... Assigned value via array Obtain the last index of the longest increasing subsequence, and then use the array... By tracing the entire subsequence in reverse, the smoothed monotonically increasing cumulative arrival curve can be obtained; in, Used to record a length of The index of the smallest last element of the increasing subsequence in the original sequence. Record the predecessor index of each element for subsequent reconstruction of the increasing subsequence. The length of the subsequence. To use the current element index as the length The last element of an increasing subsequence.

[0042] Step S4: Based on the downstream cumulative departure curve and the estimated real-time cumulative arrival curve of the lane, the macroscopic traffic speed in all time and space on the lane is reconstructed using the Newell simplified motion wave model. An Intelligent Driving Model (IDM) incorporating Newell speed constraints is established to reconstruct the vehicle trajectory on the lane and further reconstruct the lane traffic state, such as... Figure 6 The diagram shows the cumulative number of vehicles passing through.

[0043] The specific process is as follows: Step S41: Based on the principles of the simplified motion wave model (Newell) and the basic diagram of the traffic triangle, the cumulative number of vehicles passing through the middle of the road is derived through boundary condition constraints: ; ; ; in, It is the upstream cumulative arrival curve function (i.e., the function of the cumulative arrival flow at the upstream boundary as a function of time). It is the downstream cumulative departure curve function (i.e., the function of the cumulative departure flow at the downstream boundary changing with time). , To calculate the cumulative number of vehicles passing through based on upstream and downstream boundary conditions, The wave speed in free flow. The wave velocity under congested flow conditions. L For road length, For time, Indicates the location of the cross-section. Traffic density.

[0044] Step S42: According to step S41, if If the value is the minimum, then this point is governed by the upstream boundary conditions and is in a free-flow state. If the value is the minimum, then this point is governed by the downstream boundary conditions and is in a congested flow state. According to the principle of the basic diagram, the macroscopic lane speed can be obtained as follows: ; ; ; in, Indicates the location and time Vehicle density, This indicates the corresponding traffic flow. To represent a time interval, take , To represent spatial interval, take .

[0045] Step S43: Input vehicle set and lane macro speed Based on the fundamental principles of Intelligent Driving Model (IDM), vehicle state updates are performed through macro- and micro-level fusion: Input vehicle set Macroscopic speed reconstructed based on the Newell model initial weight values And related parameters of the Intelligent Driving Model (IDM): maximum acceleration Comfort deceleration Vehicle desired speed Acceleration index Minimum safe distance Safe distance and time step .

[0046] For the Vehicle, get current location ,speed Read macroscopic speed constraints ,in, Indicates macroscopic speed; when (For the lead car) When the IDM acceleration is calculated based on the free flow component, it is: ; when When following another vehicle, calculate the net distance between your vehicle and the vehicle in front: ; in, This refers to the length of the vehicle.

[0047] Calculate the safe distance that the vehicle is expected to maintain in the current state: ; The acceleration is then: ; Update vehicle status based on IDM: ; Calculate and adjust fusion weights : when When there is ample space ahead, take ; when When (approaching the vehicle in front). ; The final speed is obtained through a weighted fusion of macro and micro factors: ; Write the updated speed back to the vehicle object until all vehicle objects have been traversed.

[0048] Step S44: Based on the reconstructed micro-level car-following trajectories of individual vehicles, perform lane-level real-time queuing status identification and fusion of macro and micro traffic state reconstruction. Vehicles whose speed remains below a set threshold (5 km / h) for 10 seconds are identified as being in a queuing state, and the lane queue length is calculated. ; in, This indicates the position of the vehicle at the end of the convoy. This indicates the stop line position.

[0049] Let's analyze a specific case as an example: The vehicle license plate data records in step S1 are shown in Table 1.

[0050] Table 1 The historical characteristics of the license plate of the west entrance lane of a certain intersection in step S1 are shown in Table 2.

[0051] Table 2 The performance of the vehicle steering prediction model in step S2 on different intersection approach lanes is shown in Table 3.

[0052] Table 3 The prediction effect of turning traffic flow at the intersection approach in step S2 is as follows: Figure 7a and Figure 7b As shown, the model-predicted steering flow rate has a good fit to the actual steering flow rate. The predicted curve and the actual curve are basically consistent in their changing trends, and the difference between the two is small.

[0053] The lane cumulative arrival curve estimation results in step S3 are as follows: Figure 8As shown, under low matching rates, the model-predicted arrival flow curves for lanes at different times and in different directions can fit the actual cumulative arrival curves of the lanes well. The overall trend of the model-predicted curves and the actual curves is consistent, accurately reflecting the dynamic changes in lane-level traffic flow. Especially under high arrival flow conditions, the overlap between the predicted and actual curves is high, indicating that the model can maintain good prediction accuracy even with large traffic flows. This demonstrates that the model effectively learns the mapping relationship between input features and arrival flow, and can adapt to arrival flow estimation tasks under different traffic levels and lane types. However, in certain time periods, there are still some deviations between the predicted and actual curves, mainly manifested as differences in local fluctuation amplitudes. This may be because the model predicts over a large time span under low matching rates, while lane arrival flow is affected by special traffic events or micro-behaviors, exhibiting a certain degree of randomness. In long-term predictions, the impact of anomalies accumulates, and since the model mainly learns regular characteristics based on historical statistical data, there may be some errors.

[0054] The smoothing and historical correction effects of the cumulative arrival curve in step S3 are as follows: Figure 9 As shown in the figure, the dashed line represents the upstream arrival time of the latest matched vehicle. The left side of the dashed line corresponds to the historical cumulative arrival curve estimation part (for the known arrival flow between matched vehicles), and the right side of the dashed line corresponds to the real-time cumulative arrival curve estimation part (for the unknown arrival flow after the latest matched vehicle). As can be seen from the figure, in different lanes and at different time periods, the smoothed estimated cumulative arrival curve, compared with the original estimated cumulative arrival curve, not only eliminates local small fluctuations and maintains a monotonically increasing trend, but also better matches the actual cumulative arrival curve, demonstrating the effectiveness of the smoothing algorithm based on the longest increasing subsequence.

[0055] The lane macro velocity reconstruction result in step S4 is as follows: Figure 10 As shown, the reconstructed speed distribution spatiotemporal map presents a clear downstream congestion-release process, reflecting the vehicle queuing and queue dissipation phenomena at the downstream signalized intersection, and effectively depicting the dynamic changes in traffic conditions on the lanes.

[0056] The lane trajectory reconstruction effect in step S4 is as follows: Figure 11As shown, the reconstructed vehicle trajectories generally conform to the spatiotemporal evolution trends of the actual trajectories. Vehicles move forward gradually over time, exhibiting clear continuity and smoothness in dynamic processes such as acceleration, deceleration, stopping, and starting. In some areas, especially during periods of high-density congestion (e.g., 17:27-17:28), vehicle trajectories tend to be denser and smoother, showing a clear process of queue formation and dissipation. In free-flow conditions, vehicle trajectories are more dispersed and at higher speeds. Although there are slight deviations or temporal lags in local trajectories, the overall trend is basically consistent with the actual data. The trajectory reconstruction results demonstrate that the IDM car-following model combined with Newell speed constraints has high accuracy and stability in reconstructing the driving state of road vehicles. The model can not only finely characterize the car-following behavior of individual vehicles and capture local traffic fluctuations, but also take into account the evolution trend of macroscopic traffic flow, ensuring the continuity and rationality of the overall lane state, thereby achieving refined lane traffic state reconstruction.

[0057] The real-time reconstruction result of lane traffic status in step S4 is as follows: Figure 12 As shown in the figure, the changes in the number of vehicles queuing per second and the queue length are illustrated. The orange dashed line represents the upstream arrival time of the latest matched vehicle, used to distinguish between historical and real-time reconstructed parts. The reconstructed queue state closely matches the actual queue state, indicating that the analysis method based on reconstructed trajectory data has high accuracy and can accurately characterize the dynamic changes in lane traffic state within a signal cycle. This method demonstrates excellent sensitivity and robustness in characterizing queue formation, peak fluctuations, and dissipation processes, providing reliable data support for traffic condition monitoring. The reconstructed queue state precisely depicts the dynamic evolution of vehicle queues within a signal cycle. The number of vehicles queuing and the queue length exhibit periodic fluctuations with the signal cycle. During the red light period, the number of vehicles queuing increases rapidly, and the queue length rises accordingly. During the green light phase, the queue gradually dissipates, and the queue length drops to 0 after the last vehicle in the queue starts moving, demonstrating a typical queue formation and release process.

[0058] As attached Figure 2 As shown, in a preferred embodiment, a real-time traffic state reconstruction device for multi-lane urban roads includes the following modules: The data acquisition and matching module is used to acquire vehicle license plate data and road network topology information, match the vehicle license plate data with the road network topology information, and extract the path information of individual vehicles in the road network and the turning history characteristics at each intersection. The turning flow prediction module is used to generate training soft labels based on the path information and turning history characteristics of individual vehicles through a soft label fusion mechanism, predict the turning probability of individual vehicles through a long short-term memory network (LSTM), and accumulate the turning probabilities to obtain the macroscopic turning flow. The lane cumulative arrival curve estimation module uses the matched vehicle as a reference point and a deep neural network (DNN) to estimate the lane cumulative arrival flow from the reference point to the current time, while also estimating the corresponding lane cumulative arrival curve. The macro-micro fusion traffic state reconstruction module is used to reconstruct the lane's full-time and spatiotemporal macro traffic speed based on the downstream cumulative departure curve and the estimated lane cumulative arrival curve, using the Newell simplified motion wave model. The macro traffic speed is introduced as a constraint into the intelligent driving model (IDM) to reconstruct the micro-car-following trajectory of individual vehicles. Based on the reconstructed micro-car-following trajectory, the real-time queuing status at the lane level is identified and the real-time traffic status at the lane level is reconstructed.

[0059] In summary, the present invention provides a real-time traffic state reconstruction method for multi-lane urban roads that fully utilizes vehicle license plate data. Addressing the limitations of existing lane cumulative arrival curve estimation methods, which often rely on the assumption of homogeneous traffic flow and are ill-suited to the large short-term fluctuations in turning flow in urban road sections, this invention proposes a real-time lane cumulative arrival curve estimation method that integrates turning information. This method predicts micro-turning probabilities based on the characteristic information of individual vehicles and aggregates them into macro-turning flow. Furthermore, it constructs an arrival flow estimation model that integrates turning flow and captures the lane selection patterns implicit in the turning flow through a deep neural network. This solves the problem of unknown lane selection for vehicles arriving from upstream of the road segment in real-time estimation, effectively supporting accurate and real-time estimation of lane cumulative arrival curves.

[0060] Secondly, considering the different characteristics of macro-group reconstruction and micro-individual reconstruction, a lane traffic state reconstruction method that integrates macro-features and micro-car-following behavior is proposed: the macro-velocity of the lane in all time and space is reconstructed based on the Newell simplified motion wave model. On this basis, a micro-car-following model with macro-constraints is constructed to improve the robustness of real-time traffic state reconstruction of urban multi-lane lanes.

[0061] Compared with existing technologies, it has the following technological advantages: 1. In the process of lane cumulative arrival curve estimation, the implicit correlation between individual vehicle driving path, historical steering preference and actual steering decision and corresponding lane selection is considered. A real-time estimation method for lane cumulative arrival curve that integrates vehicle steering information is proposed, which solves the problem of unknown steering lane selection of vehicles arriving upstream of the road segment in real-time estimation and provides reliable lane upstream arrival information.

[0062] 2. In the process of traffic state reconstruction, the characteristics of macro-group trajectory reconstruction and micro-individual reconstruction are considered at the same time. Macro speed is introduced as a constraint into the micro-car-following model, which fully integrates macro traffic speed and micro-car-following behavior. This makes the reconstructed traffic state not only conform to the overall traffic flow characteristics and alleviate the accumulation of individual estimation errors, but also accurately depict the local traffic state.

[0063] Those skilled in the art will understand that embodiments of this application can be provided as methods, apparatus, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0064] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in one or more blocks of the flowchart illustrations and / or one or more blocks of the block diagrams.

[0065] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means that implement the functions specified in one or more flowcharts and / or one or more block diagrams.

[0066] These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, such that the instructions, which execute on the computer or other programmable apparatus, provide steps for implementing the functions specified in one or more flowcharts and / or one or more block diagrams.

[0067] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

Claims

1. A method for reconstructing real-time traffic states of multi-lane urban roads, characterized in that, include: Step S1: Obtain vehicle license plate data and road network topology information, match the vehicle license plate data with the road network topology information, and extract the path information of individual vehicles in the road network and the turning history features at each intersection. Step S2: Based on the path information and turning history characteristics of individual vehicles, training soft labels are generated through a soft label fusion mechanism. The turning probability of individual vehicles is predicted through an LSTM long short-term memory network. The turning probabilities are accumulated to obtain the turning flow. Step S3: Using the matched vehicle as the reference point, estimate the lane cumulative arrival flow from the reference point to the current time through a DNN deep neural network, and estimate the corresponding lane cumulative arrival curve. Step S4: Based on the downstream cumulative departure curve and the estimated lane cumulative arrival curve, the macroscopic traffic speed of the lane is reconstructed using the Newell simplified motion wave model. The traffic speed is introduced as a constraint into the IDM intelligent driving model to reconstruct the micro-car-following trajectory of individual vehicles. Based on the reconstructed micro-car-following trajectory, the queuing status and the fusion of macro and micro traffic status are identified. Among them, based on the downstream cumulative departure curve and the estimated lane cumulative arrival curve, the Newell simplified motion wave model is used to reconstruct the lane's all-time and all-space macroscopic traffic speed, including: Based on the principles of Newell's simplified motion wave model and the basic diagram of traffic triangles, the cumulative number of vehicles passing through the middle of the road is derived through boundary condition constraints: ； ； ； wherein, is an upstream cumulative arrival curve function; is a downstream cumulative departure curve function; is a cumulative number of vehicles solved based on upstream and downstream boundary conditions; is a wave speed in free flow state; is a wave speed in jammed flow state; L is a road length; is a time; denotes a cross-section position, is a traffic density; like If the value is the minimum, then this point is governed by the upstream boundary conditions and is in a free-flow state. If the value is the minimum, then this point is governed by the downstream boundary conditions and is in a congested flow state. According to the principle of the basic traffic triangle diagram, the macroscopic lane speed is obtained as follows: ； ； ； in, Indicates the location and time Vehicle density, This indicates the corresponding traffic flow. To represent a time interval, take , To represent spatial interval, take ; Macroscopic traffic speed is introduced as a constraint into the IDM intelligent driving model to reconstruct the micro-level car-following trajectory of individual vehicles. Based on the reconstructed micro-level car-following trajectory, lane-level real-time queuing status is identified and lane-level real-time traffic status is reconstructed, including: Perform macro-micro fusion vehicle status updates to reconstruct the micro-level car-following trajectory of individual vehicles: ； in, To obtain the final speed through a weighted fusion of macro and micro factors; To optimize weighting, the weights are dynamically adjusted based on the distance between the vehicle and the vehicle in front. This represents a macroscopic speed constraint. Indicates macroscopic speed; This indicates that the vehicle status is updated based on IDM. Based on the reconstructed micro-following trajectories of individual vehicles, lane-level real-time queuing status recognition and traffic status reconstruction are performed: ； in, This indicates the position of the vehicle at the end of the convoy. This indicates the stop line position.

2. The method for real-time traffic status reconstruction of multi-lane urban roads according to claim 1, characterized in that, In step S1, The path information includes: intersection approach lane numbers and corresponding passage time sequences; The vehicle license plate data includes: license plate detection equipment number, intersection number, vehicle passage time, license plate data, license plate type, exit from turning lane, and direction of approach lane.

3. The method for real-time traffic state reconstruction of multi-lane urban roads according to claim 2, characterized in that, In step S1, the steering history features include: Historical turning probability of individual vehicles at intersections; The distribution of vehicles turning at the intersection in the previous time period; For vehicles without historical turning records, the vehicle turning distribution at the intersection in the previous time period is used.

4. The method for real-time traffic state reconstruction of multi-lane urban roads according to claim 1, characterized in that, In step S2, training soft labels are generated through a soft label fusion mechanism. The calculation process is as follows: Soft tags are generated by weighting the route information and steering history characteristics of individual vehicles: ； in, A soft tag for individual vehicles. A hard label for individual vehicles. For the steering history characteristics of individual vehicles, This refers to hard label weights.

5. The method for real-time traffic state reconstruction of multi-lane urban roads according to claim 4, characterized in that, In step S2, the turning probability of individual vehicles is predicted using an LSTM (Long Short-Term Memory) network. The turning probabilities are then accumulated to obtain the macroscopic turning flow rate. The calculation process is as follows: The discrete index of the upstream inlet channel is converted into a dense vector representation through the word embedding layer; The output of the word embedding layer is concatenated with the temporal feature information according to the feature dimension, generating a dimension of... T×d The path time vector is input into the LSTM layer to extract the hidden temporal correlations of the path information. T The time step length, d For feature dimensions: Computing the forget gate in LSTM long short-term memory networks : ； in, for , The number of hidden layer units. Here is the weight matrix for the forget gate. For the offset of the forget gate, This is the hidden state from the previous moment. Let the input vector be the input vector at the current time. for function; Computing the input gate in a Long Short-Term Memory (LSTM) network : ； in, This represents the weight matrix of the input gate. This represents the bias term of the input gate; Generate candidate cell state information: ； in, For activation function, w c The weight matrix represents the candidate cells. b c The bias term representing candidate cells; Update cell status: ； Calculate the output gate of an LSTM (Long Short-Term Memory) network. : ； in, w o This represents the weight matrix of the output gate; This represents the bias term of the output gate; Output the hidden state at the current moment. : ( )； The hidden state and historical features of the last time step output of the LSTM Long Short-Term Memory network are input into the fully connected layer. The vehicle historical turning data is encoded as [1,0], and the vehicle turning data of the previous time step at the intersection is encoded as [0,1]. The output of the fully connected layer is transformed into normalized turning probabilities through a Softmax layer. 　； in, For category; This is the output of the fully connected layer; Total number of categories; Calculate the loss during training using KL divergence: ； in, The total number of samples; This represents the true probability distribution; To predict the probability distribution; and Represent the true distribution and the predicted distribution in the category, respectively. The probability of it; The predicted turning probabilities of all individual vehicles within the study period are summed to obtain the macroscopic turning flow.

6. The method for real-time traffic state reconstruction of multi-lane urban roads according to claim 1, characterized in that, In step S3, using the matched vehicle as a reference point, the cumulative lane arrival flow from the reference point to the current time is estimated using a deep neural network (DNN), including: Matching vehicles by matching license plate data at upstream and downstream intersections Using the baseline, calculate the cumulative arrival flow of the lane at the current time: ； ； in, for Time Lane The cumulative arrival flow, Number the downstream lane. Number the upstream inlet channel; To match vehicles The corresponding cumulative number of arriving vehicles; This refers to the number of upstream import channels; This is a mapping function between feature information and lane arrival flow. for and Between, leaving the upstream entrance road Traffic; For upstream import channels and lane Nonlinear mapping function between; for and Within a given time period, the turning flow obtained from the aggregation of turning probabilities; For upstream import channels In the departing traffic, lanes are allocated Traffic on the internet For lane The corresponding downstream shift, For upstream import channels The departing traffic, For upstream import channels In the departing flow, Steering ratio in direction, For upstream import channels The resulting turning flow In the middle, assigned to lane Traffic share on; Using a deep neural network (DNN) to fit a nonlinear mapping relationship between features and lane arrival flow. The arrival flow rate within the lane is obtained. The DNN deep neural network uses the ReLU activation function and sets the dropout parameter, with the root mean square error (RMSE) as the loss function. 　； in, The total number of samples, For the first The estimated arrival flow for each sample This corresponds to the actual arrival traffic.

7. The method for real-time traffic state reconstruction of multi-lane urban roads according to claim 6, characterized in that, In step S3, the corresponding lane cumulative arrival curve is estimated, including: The original cumulative arrival curve is smoothed using the longest increasing subsequence of LIS, by defining... and Given two arrays, iterate through the sequence and update the state for each element in the sequence. Using binary search in the array Find the largest , making If such a thing exists, it is true. Then update Assign the value to the current index. At the same time, the predecessor index of the current element is... Assigned value via array Obtain the last index of the longest increasing subsequence, and then use the array... By tracing the entire subsequence in reverse, the smoothed monotonically increasing cumulative arrival curve can be obtained; in, Used to record a length of The index of the smallest last element of the increasing subsequence in the original sequence. Record the predecessor index of each element for subsequent reconstruction of the increasing subsequence. The length of the subsequence. To use the current element index as the length The last element of an increasing subsequence.

8. A device for real-time traffic state reconstruction of multi-lane urban roads, employing the method for real-time traffic state reconstruction of multi-lane urban roads as described in any one of claims 1 to 7, characterized in that, include: The data acquisition and matching module is used to acquire vehicle license plate data and road network topology information, match the vehicle license plate data with the road network topology information, and extract the path information of individual vehicles in the road network and the turning history characteristics at each intersection. The turning flow prediction module is used to generate training soft labels based on the path information and turning history characteristics of individual vehicles through a soft label fusion mechanism, predict the turning probability of individual vehicles through an LSTM long short-term memory network, and accumulate the turning probabilities to obtain the macroscopic turning flow. The lane cumulative arrival curve estimation module uses the matched vehicle as a reference point and a deep neural network (DNN) to estimate the lane cumulative arrival flow from the reference point to the current time, while also estimating the corresponding lane cumulative arrival curve. The macro-micro fusion traffic state reconstruction module is used to reconstruct the lane's all-time macro traffic speed based on the downstream cumulative departure curve and the estimated lane cumulative arrival curve, using the Newell simplified motion wave model. The macro traffic speed is introduced as a constraint into the IDM intelligent driving model to reconstruct the micro car-following trajectory of individual vehicles. Based on the reconstructed micro car-following trajectory, the real-time queuing status at the lane level is identified and the real-time traffic status at the lane level is reconstructed.

Images