A two-stage highway vehicle discrete point trajectory reconstruction method and system based on deep learning

Abstract

The present invention discloses a two-stage highway vehicle discrete point trajectory reconstruction method and system based on deep learning, including using density-based outlier detection algorithm to clean abnormal discrete trajectory points after feature extraction based on millimeter radar wave data; establishing a highway network topology map, considering the Betweenness Centrality method based on graph theory to identify key trajectory nodes; using the CA-LWR method to predict the full-time and space traffic flow trajectory of the cleaned millimeter radar wave multi-source discrete traffic flow trajectory points; embedding the traffic flow trajectory obtained by the discrete trajectory point prediction into the encoder of deep learning and inputting it into the Transformer as the historical vehicle trajectory to enhance the accuracy of target vehicle trajectory reconstruction. The present invention can improve the accuracy and real-time performance of trajectory reconstruction, help optimize the traffic system, improve traffic efficiency, provide data support for highway managers, and at the same time provide a safer path for vehicle cruise control in the field of autonomous driving.

Description

Technical Field

[0001] The present invention belongs to the field of communications, and further relates to a two-stage highway vehicle discrete point trajectory reconstruction method and system based on deep learning in the technical field of traffic flow simulation and traffic data analysis in the transportation field. Background Art

[0002] With the continuous development of urban traffic and the increasing demand for traffic management, highway traffic flow trajectory reconstruction has become one of the important technologies for optimizing the traffic system and improving traffic efficiency. As an advanced sensing technology, millimeter radar waves have the characteristics of high precision and high resolution. They can penetrate adverse weather conditions such as rain and snow and obtain the motion information of target objects in real time. In the field of highway traffic management, the trajectory reconstruction method based on millimeter radar waves accurately monitors the tiny movements of objects such as vehicle engines, wheels and bodies. However, considering the detector wave frequency, how to more effectively reconstruct discrete data into a traffic flow that is more in line with actual needs is an urgent problem to be solved under the rapid development of intelligent transportation.

[0003] In terms of traffic big data application, traditional traffic flow trajectory reconstruction equipment mainly includes magnetic sensors, cameras, radars, etc., and traditional methods usually use signal processing and calculation methods based on single or combined sensor data. These devices and methods have more or less shortcomings: magnetic sensors are suitable for low-speed scenes, but for occasions such as highways that require high-precision trajectories, there may be problems with insufficient accuracy. Cameras are affected by factors such as weather and lighting, and their performance may degrade under harsh conditions. Therefore, traffic flow trajectory reconstruction methods based on advanced technologies such as millimeter-wave radars have emerged to overcome the shortcomings of these traditional methods and improve the accuracy and real-time performance of trajectory reconstruction. Summary of the invention

[0004] Purpose of the invention: The purpose of the present invention is to provide a two-stage highway vehicle discrete point trajectory reconstruction method and system based on deep learning, which can improve the accuracy and real-time performance of trajectory reconstruction, help optimize the traffic system, improve traffic efficiency, provide data support for highway managers, and provide a safer path for vehicle cruise control in the field of autonomous driving.

[0005] Technical solution: To achieve the above-mentioned invention object, the technical solution adopted by the present invention is:

[0006] A two-stage highway vehicle discrete point trajectory reconstruction method based on deep learning, comprising the following steps:

[0007] After extracting features from millimeter radar wave data, a density-based outlier detection algorithm is used to clean abnormal discrete trajectory points;

[0008] Establish a highway network topology map and consider the Betweenness Centrality method based on graph theory to identify key trajectory nodes;

[0009] The CA-LWR method is used to predict the full-time and space traffic flow trajectory of the cleaned millimeter radar wave discrete traffic flow trajectory points. The LWR model is used to update the state of the cellular automaton. The domain cell vehicle speed and flow control the vehicle flow density in each cell. The value range of the vehicle speed in the cellular automaton is controlled according to the maximum speed in the LWR model. For key trajectory nodes, the interaction of surrounding vehicles is considered and the safety distance between vehicles is set.

[0010] The traffic flow trajectory obtained by discrete trajectory point prediction is embedded in the deep learning encoder and input into the Transformer as the historical vehicle trajectory to enhance the accuracy of target vehicle trajectory reconstruction; the Transformer network model is a vehicle trajectory reconstruction model of the Transformer network based on multi-head attention.

[0011] Furthermore, the method of extracting features from the millimeter radar wave data and then using a density-based outlier detection method to clean abnormal discrete trajectory points includes:

[0012] The motion characteristics of the vehicle are extracted. According to the actual situation and the principles of classical mechanics, when the vehicle is detected by the millimeter radar wave while driving on the highway, the vehicle speed should be within the speed limit of the highway, the acceleration should be between 0.1g and 0.5g (g is the local gravity acceleration), and the direction angle should be within the road line shape;

[0013] Extract features from traffic flow data, including vehicle longitude, latitude, and time; calculate the Euclidean distance between points, use the average distance as the cluster radius, set the minimum cluster density to twice the dimension, input the discrete vehicle trajectory point set into the DBSCAN algorithm training, delete the outlier data marked as noise points, and retain the boundary points and core points.

[0014] Furthermore, the highway network topology map is established, and the Betweenness Centrality method based on graph theory is considered to identify key trajectory nodes, including:

[0015] The highway network is represented as a graph, where nodes represent high-density points, ramps, bottlenecks and toll gates in the density-based outlier detection algorithm, and edges represent roads connecting these locations. The high-density points represent points whose data points contained in the neighborhood are at least the minimum cluster density; for each pair of nodes, the Dijkstra algorithm is used to calculate the number of shortest paths passing through the node and the betweenness centrality, and the nodes with a preset proportion in the betweenness centrality are marked as key trajectory nodes in the highway network. The expression of betweenness centrality is:

[0016]

[0017] where σ st is the number of shortest paths from node s to node t, σ st (v) represents the number of shortest paths passing through node v on the shortest path.

[0018] Furthermore, the method of using the CA-LWR method to predict the full-time and space traffic flow trajectory of the cleaned millimeter radar wave multi-source discrete traffic flow trajectory points includes:

[0019] According to the established highway topology, the road is divided into several cells. The key trajectory nodes should be included in independent cells.

[0020] Initialize the state of the cellular automaton according to the current traffic flow and vehicle position at each point, as well as the layout of the traffic network, the length of the road and the number of lanes. According to the LWR model, the update of the next cell should make the state of the current cell satisfy the formula q = ρ·v as much as possible, where q is the flow, ρ is the traffic density, and v is the vehicle speed; the speed and flow of the neighborhood cells control the traffic density in each cell, and the range of the vehicle speed in the cellular automaton is controlled according to the maximum speed in the LWR model;

[0021] For key trajectory nodes, the interaction of surrounding vehicles is taken into consideration. Key nodes are areas with large traffic volume, such as weaving areas, which are prone to accidents. Therefore, a safe distance between vehicles is set to control the update rules of the cellular automaton. The above method is used to predict the trajectory of all vehicle discrete trajectory data to obtain the full-time and space traffic flow trajectory.

[0022] The LWR model predicts the vehicle position at the next moment based on the information of the number of vehicles on each cell on the road, the road network information, the length of each cell, and the time step of the model; wherein the update rule of vehicle density satisfies:

[0023]

[0024] where ρ i (t) is the vehicle density at position i at time t, q i+1 / 2 (t) and q i-1 / 2 (t) are the traffic flows at positions i+1 / 2 and i-1 / 2, respectively. Δt and Δx are the time step and space step, respectively.

[0025] Furthermore, embedding the traffic flow trajectory obtained by discrete trajectory point prediction into the deep learning encoder and inputting it into the Transformer as the historical vehicle trajectory to enhance the accuracy of target vehicle trajectory reconstruction includes:

[0026] The full-time and space traffic flow trajectory predicted by CA-LWR is set as the relevant vehicles around the target vehicle, and each coordinate point is mapped to an embedded vector to form a time series, which is represented by Trajectory near =[(x 1t ,y 1t ),(x 2t ,y 2t ),...(x nt ,y nt )], where (x 1t ,y 1t )、(x 2t ,y 2t )...(x nt ,y nt ) is the coordinate data of vehicle 1, vehicle 2, ..., vehicle n at time t. The trajectory data is processed by Positional Embedding to retain the time series information of the trajectory points;

[0027] After embedding the encoder with the predicted trajectory as the historical trajectory, multiple Self-Attention Layer layers are used to extract the surrounding vehicle information in the sequence, and the output is used as the input of the decoder;

[0028] The decoder receives the output of the encoder and dynamically weights the self-attention mechanism in the interaction modeling module to capture the spatial dependency between each vehicle and the target vehicle in the scene. After assigning different attention weights, the multi-head attention mechanism is input to learn the vehicle interaction effects from multiple different feature subspaces, and the interactive effects of different feature spaces are summarized and aggregated through the weights of the fully connected layer, thereby enhancing the model's fitting ability and adaptability to different scenarios, and generating predictions of future trajectories through the feedforward neural network layer.

[0029] A two-stage highway vehicle discrete point trajectory reconstruction system based on deep learning, comprising:

[0030] The preprocessing module is used to extract features from millimeter radar wave data and then use a density-based outlier detection algorithm to clean abnormal discrete trajectory points;

[0031] The key trajectory node identification module is used to establish the highway network topology map and consider the Betweenness Centrality method based on graph theory to identify the key trajectory nodes;

[0032] The first stage trajectory prediction module is used to use the CA-LWR method to predict the full-time and space traffic flow trajectory of the cleaned millimeter radar wave discrete traffic flow trajectory points. The LWR model is used to update the state of the cellular automaton CA. The neighborhood cell speed and flow control the traffic density in each cell. The value range of the vehicle speed in the cellular automaton is controlled according to the maximum speed in the LWR model. For key trajectory nodes, the interaction of surrounding vehicles is considered and the safety distance between vehicles is set;

[0033] The second-stage trajectory prediction module is used to embed the traffic flow trajectory obtained by discrete trajectory point prediction into the deep learning encoder and input it into the Transformer network model as the historical vehicle trajectory to enhance the accuracy of target vehicle trajectory reconstruction; the Transformer network model is a vehicle trajectory reconstruction model based on the Transformer network with multi-head attention.

[0034] A computer system includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded into the processor, the steps of a two-stage highway vehicle discrete point trajectory reconstruction method based on deep learning are implemented.

[0035] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of a two-stage highway vehicle discrete point trajectory reconstruction method based on deep learning.

[0036] A computer program product includes a computer program / instruction, which, when executed by a processor, implements the steps of a two-stage highway vehicle discrete point trajectory reconstruction method based on deep learning.

[0037] Beneficial effects: The present invention uses an innovative two-stage prediction, using the CA-LWR algorithm to predict the full-time and space traffic flow trajectory as historical vehicle trajectory data embedded in the Transformer to improve the prediction accuracy, and then considers vehicle interaction. The Transformer algorithm based on the self-attention mechanism is used to output the reconstructed trajectory, which is helpful to optimize the traffic system, improve traffic efficiency, and provide data support for highway managers. At the same time, in the field of autonomous driving, it can provide a safer path for vehicle cruise control. Compared with the prior art, the present invention has the following advantages: the present invention can reconstruct vehicle trajectories in combination with discrete trajectory data obtained by millimeter radar wave measurement, and can also reconstruct vehicle trajectories without relying on historical traffic flow trajectory data. Secondly, the LWR model is combined on the basis of the CA prediction method, so that the prediction is more suitable for the traffic field, and has certain practical significance for highway traffic detection and control. At the same time, the two-stage trajectory prediction aims to generate multiple possible and safe trajectories for target vehicles in heterogeneous, complex and highly dynamic environments, which can provide a more optimized and safe path for the vehicle's autonomous driving technology, and more in-depth research can be applied to the future autonomous driving technology of vehicles. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 A flow chart of a two-stage highway vehicle discrete point trajectory reconstruction method based on deep learning according to the present invention;

[0039] Figure 2 A schematic diagram of cell division and state transition in step S3 of a two-stage highway vehicle discrete point trajectory reconstruction method based on deep learning of the present invention;

[0040] Figure 3 Schematic diagram of the vehicle trajectory prediction model structure in step S4 of a two-stage highway vehicle discrete point trajectory reconstruction method based on deep learning of the present invention. DETAILED DESCRIPTION

[0041] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0042] like Figure 1 As shown, a two-stage highway vehicle discrete point trajectory reconstruction method based on deep learning is disclosed in an embodiment of the present invention, and the specific implementation steps are as follows:

[0043] S1. After extracting features from millimeter radar wave data, a density-based outlier detection algorithm is used to clean abnormal discrete trajectory points.

[0044] In this embodiment, the acquired millimeter radar wave data includes physical information such as time, flow rate, vehicle speed, lane number, acceleration, and coordinate information.

[0045] The motion characteristics of the vehicle are extracted. According to the actual situation and the principles of classical mechanics, the vehicle speed when the vehicle is detected by the millimeter radar wave on the highway should be within the speed limit of the highway, and the acceleration should be between 0.1g and 0.5g (g is the gravity acceleration of the highway). According to the actual longitude and latitude combined with the radius of the earth, the acceleration range is [0.978m / s 2 ,4.892m / s 2 ], direction angle where x n (t),y n (t) is the horizontal and vertical coordinates of vehicle n at time t, Δt is the time span, which should be within the road line shape. According to the actual road line shape, the range of road direction angle is

[0046] Density-based outlier detection algorithm: Extract features from traffic flow data, including vehicle longitude, latitude, and time. Calculate the Euclidean distance between discrete vehicle trajectory points where x p ,x q is the horizontal coordinate distance between vehicle p and vehicle q, y p ,y q is the ordinate distance between vehicle p and vehicle q. As the cluster radius, the minimum cluster density is set to twice the dimension, where D(p i ,q i ) is the calculated Euclidean distance, and n represents the number of selected samples. The discrete vehicle trajectory point set is input into the DBSCAN algorithm training, and the outlier data marked as noise points are deleted, and the boundary points and core points are retained.

[0047] S2. Establish a highway network topology map, consider the Betweenness Centrality method based on graph theory to identify key trajectory nodes, and then introduce the CA-LWR traffic flow trajectory reconstruction method to optimize vehicle state transfer. Including:

[0048] The highway network is represented as a graph, where nodes represent high-density points in the density-based outlier detection algorithm (if the data points contained in the ∈ neighborhood of a point P have at least the minimum cluster density mentioned above, then this point P is defined as a high-density point), ramps, bottlenecks, toll gates, and edges represent roads connecting these locations. For each pair of nodes, the Dijkstra algorithm is used to calculate the number of shortest paths passing through the node. The formula is:

[0049] dist[v]=min(dist[v],dist[u]+weight(u,v))

[0050] The meaning of this formula is that if the currently known shortest distance d(v) from the source node to node v is greater than the distance from the source node through node u to node v dist[v],dist[u]+weight(u,v), then d(v) is updated to min(dist[v],dist[u]+weight(u,v)), indicating that the distance to node v through node u is shorter.

[0051] Input the topological graph into the operations research computing tool to obtain all the shortest paths in the topological graph.

[0052] Furthermore, we calculate the betweenness centrality using the NetworkX library

[0053] import networkx as nx

[0054] G = nx.Graph()

[0055] betweenness_centrality=nx.betweenness_centrality(G)

[0056] key_nodes=[node for node, centrality in betweenness_centrality.items()if centrality>threshold]

[0057] The top 20% of nodes in betweenness centrality are marked as key nodes in the highway network. The expression of betweenness centrality is:

[0058] where σ st is the number of shortest paths from node s to node t, σ st (v) represents the number of shortest paths passing through node v on the shortest path.

[0059] S3. The CA-LWR method is used to predict the full-time and space traffic flow trajectory of the cleaned millimeter radar wave discrete traffic flow trajectory points. The LWR model is used to update the state of the cellular automaton CA. The neighborhood cell speed and flow control the traffic density in each cell. The value range of the vehicle speed in the cellular automaton is controlled according to the maximum speed in the LWR model. For key trajectory nodes, the interaction of surrounding vehicles is considered and the safety distance between vehicles is set. Including:

[0060] According to the established highway topology, the road is divided into several cells. Key nodes should be included in independent cells.

[0061] like Figure 2The interaction between the vehicles shown in the figure is based on the current traffic flow and vehicle position of each discrete trajectory point, as well as the layout of the traffic network, the length of the road and the number of lanes. The state of the cellular automaton is initialized. According to the LWR model, the update of the next cell should make the state of the current cell satisfy the formula q = ρ·v as much as possible, where q is the flow, ρ is the traffic density, and v is the speed. At the same time, the speed and flow of the neighborhood cells control the traffic density in each cell, and the range of the vehicle speed in the cellular automaton is controlled according to the maximum speed in the LWR model;

[0062] The model assumes that traffic flow Q is the product of traffic flow density k and vehicle speed v, that is, Q = K·V. Among them, vehicle speed v is usually assumed to be a function of traffic flow density, while traffic flow density k is a function of position x and time t. The LWR model can deduce the evolution of the macroscopic state of traffic flow, such as changes in traffic flow, traffic density and vehicle speed. The basic assumption of the model is the conservation of traffic flow. Without considering the influence of sources and sinks, the traffic inflow and outflow of any section of road are equal.

[0063] The LWR model outputs the predicted vehicle position at the next moment based on the information of the number of vehicles on each unit cell (or cellular) on the input road, the road network information, the length of each cellular, the time step of the model, etc. The update rule of vehicle density should satisfy:

[0064]

[0065] where ρ i (t) is the vehicle density at position i at time t, q i+1 / 2 (t) and q i-1 / 2 (t) are the traffic flows at positions i+1 / 2 and i-1 / 2, respectively. Δt and Δx are the time step and space step, respectively.

[0066] The model combines the discrete characteristics of cellular automata and the continuous characteristics of the LWR model to simulate and analyze vehicle trajectories. This model is good at simulating large-scale traffic networks and can handle various actual traffic rules and driving behaviors.

[0067] Furthermore, for key trajectory nodes, the interaction with surrounding vehicles is considered. Key nodes are areas with high traffic volume, such as weaving areas, which are prone to accidents. Therefore, it is considered to set a safe distance between vehicles to control the update rules of the cellular automaton. The above method is used to predict the trajectory of all vehicle discrete trajectory data to obtain the full-time and space traffic flow trajectory, thereby building the first-stage full-time and space macro traffic flow trajectory interaction environment.

[0068] S4. The traffic flow trajectory obtained by discrete trajectory point prediction is embedded in the deep learning encoder and input into the Transformer as the historical vehicle trajectory to enhance the accuracy of target vehicle trajectory reconstruction. Including:

[0069] The full-time and space traffic flow trajectory predicted by CA-LWR is set as the relevant vehicles around the target vehicle, and each coordinate point is mapped to an embedded vector to form a time series of length n, which is represented by Trajectory near =[(x 1t ,y 1t ),(x 2t ,y 2t ),...(x nt ,y nt )], where (x 1t ,y 1t )、(x 2t ,y 2t )...(x nt ,y nt ) is the coordinate data of vehicle 1, vehicle 2, ..., vehicle n at time t. The trajectory data is processed by Positional Embedding to retain the time series information of the trajectory points.

[0070] After the encoder is embedded with the predicted trajectory as the historical trajectory, the surrounding vehicle information in the sequence is extracted by an independent layer through multiple Self-AttentionLayer layers, and the output is used as the input of the decoder.

[0071] The decoder receives the output of the encoder and dynamically weights the self-attention mechanism in the interactive modeling module.

[0072] Capture the spatial dependency between each vehicle in the scene and the target vehicle, where Q, K, and V are input vectors, and d k is the scale factor, and the weight is obtained through the soft function.

[0073] Furthermore, different attention weights are assigned and then input into the multi-head attention mechanism Multi-Head(Q,K,V)=Concat(head 1 ,...head h )W o Learning vehicle interactions from multiple different feature subspaces, the formula represents concatenating the outputs of multiple attention heads and then passing through a linear transformation W oGet the final multi-head attention output, enhance the model's fitting ability and adaptability to different scenarios, and use the Transformer model classTransformerModel(nn.Module) in PyTorch;

[0074] input_size = 2 coordinate dimensions

[0075] hidden_size=64 #Hidden layer dimensions

[0076] num_layers = 4 # Transformer layers

[0077] num_heads = 8 # Number of attention heads

[0078] output_size=2#output dimension, i.e. predicted coordinate dimension

[0079] Feed-forward neural network layers generate predictions of future trajectories.

[0080] In summary, the technical solution further uses the deep learning Transformer algorithm to perform the second-stage trajectory reconstruction of the target vehicle on the basis of the technology of constructing historical vehicle trajectories by the CA-LWR algorithm. Compared with the traditional vehicle trajectory reconstruction, the two-stage prediction adopted by the present invention is more adaptable to the reconstruction of discrete vehicle trajectory points. At the same time, it can reduce the prediction error to a certain extent when there is a lack of historical traffic flow trajectory data. It can make up for the inherent defects of the more traditional detectors, help optimize the traffic system, improve traffic efficiency, provide data support for highway managers, and provide a safer path for vehicle cruise control in the field of autonomous driving.

[0081] Based on the same inventive concept, an embodiment of the present invention discloses a two-stage highway vehicle discrete point trajectory reconstruction system based on deep learning, including: a preprocessing module, used to extract features based on millimeter radar wave data and then use a density-based outlier detection algorithm to clean abnormal discrete trajectory points; a key trajectory node identification module, used to establish a highway network topology map, considering the Betweenness Centrality method based on graph theory to identify key trajectory nodes; a first-stage trajectory prediction module, used to use the CA-LWR method to predict the full-time and space traffic flow trajectory for the cleaned millimeter radar wave discrete traffic flow trajectory points; a second-stage trajectory prediction module, used to embed the traffic flow trajectory predicted by the discrete trajectory points into a deep learning encoder and input it into the Transformer network model as a historical vehicle trajectory to enhance the accuracy of target vehicle trajectory reconstruction.

[0082] Based on the same inventive concept, an embodiment of the present invention discloses a computer system, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded into the processor, the steps of the two-stage highway vehicle discrete point trajectory reconstruction method based on deep learning are implemented.

[0083] Based on the same inventive concept, an embodiment of the present invention discloses a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, the steps of the two-stage highway vehicle discrete point trajectory reconstruction method based on deep learning are implemented.

[0084] Based on the same inventive concept, an embodiment of the present invention discloses a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of a two-stage highway vehicle discrete point trajectory reconstruction method based on deep learning.

Claims

1. A two-stage highway vehicle discrete point trajectory reconstruction method based on deep learning, characterized in that: The following steps are involved: After extracting features from millimeter radar wave data, a density-based outlier detection algorithm is used to clean abnormal discrete trajectory points; Establish a highway network topology map and consider the Betweenness Centrality method based on graph theory to identify key trajectory nodes; The CA-LWR method is used to predict the full-time and space traffic flow trajectory of the cleaned millimeter radar wave discrete traffic flow trajectory points. The LWR model is used to update the state of the cellular automaton CA. The neighborhood cell speed and flow control the traffic density in each cell. The value range of the vehicle speed in the cellular automaton is controlled according to the maximum speed in the LWR model. For key trajectory nodes, the interaction of surrounding vehicles is considered and the safety distance between vehicles is set. The traffic flow trajectory obtained by discrete trajectory point prediction is embedded in the deep learning encoder and input into the Transformer network model as the historical vehicle trajectory to enhance the accuracy of target vehicle trajectory reconstruction; the Transformer network model is a vehicle trajectory reconstruction model of the Transformer network based on multi-head attention.

2. The two-stage highway vehicle discrete point trajectory reconstruction method based on deep learning according to claim 1 is characterized in that: The method of extracting features from the millimeter radar wave data and then using a density-based outlier detection algorithm to clean abnormal discrete trajectory points includes: The motion characteristics of the vehicle are extracted. According to the actual situation and the principles of classical mechanics, when the vehicle is detected by the millimeter radar wave while driving on the highway, the vehicle speed should be within the speed limit range of the highway, the acceleration should be between 0.1g and 0.5g, where g is the local gravity acceleration, and the direction angle should be within the road line shape; Extract features from traffic flow data, including vehicle longitude, latitude, and time; calculate the Euclidean distance between discrete vehicle trajectory points, use the average distance as the cluster radius, set the minimum cluster density to twice the dimension, input the discrete vehicle trajectory point set into the DBSCAN algorithm training, delete the outlier data marked as noise points, and retain the boundary points and core points.

3. The two-stage highway vehicle discrete point trajectory reconstruction method based on deep learning according to claim 1 is characterized in that: The highway network topology map is established, and the Betweenness Centrality method based on graph theory is considered to identify key trajectory nodes, including: The highway network is represented as a graph, where nodes represent high-density points, ramps, bottlenecks, and toll gates in the density-based outlier detection algorithm, and edges represent roads connecting these locations. The high-density points represent points whose data points contained in the neighborhood are at least the minimum cluster density; for each pair of nodes, the number of shortest paths and betweenness centrality are calculated. where σ st is the number of shortest paths from node s to node t, σ st (v) represents the number of shortest paths passing through node v on the shortest path; nodes with a preset proportion in betweenness centrality are marked as key trajectory nodes in the highway network.

4. According to claim 1, a two-stage highway vehicle discrete point trajectory reconstruction method based on deep learning is characterized in that: The method of using the CA-LWR method to predict the full-time and space traffic flow trajectory of the cleaned millimeter radar wave discrete traffic flow trajectory points includes: According to the established highway network topology, the road is divided into several cells. The key trajectory nodes should be included in independent cells. Initialize the state of the cellular automaton according to the traffic flow and vehicle position near each discrete trajectory point, as well as the layout of the traffic network, the length of the road, and the number of lanes. According to the LWR model, the update of the next cell should make the state of the current cell satisfy the formula q = ρ v as much as possible, where q is the flow, ρ is the traffic density, and v is the vehicle speed. The speed and flow of the neighborhood cells control the traffic density in each cell, and the range of the vehicle speed in the cellular automaton is controlled according to the maximum speed in the LWR model. For key trajectory nodes, the interaction between surrounding vehicles is considered, and the safety distance between vehicles is set to control the update rules of the cellular automaton. The trajectory prediction is performed on all the discrete trajectory data of vehicles to obtain the full-time and space traffic flow trajectory.

5. According to claim 1, a two-stage highway vehicle discrete point trajectory reconstruction method based on deep learning is characterized in that: The LWR model predicts the vehicle position at the next moment based on the information of the number of vehicles on each cell on the road, the road network information, the length of each cell and the time step of the model; wherein the update rule of vehicle density satisfies: where ρ i (t) is the vehicle density at position i at time t, q i+1 / 2 (t) and q i-1 / 2 (t) are the traffic flows at positions i+1 / 2 and i-1 / 2, respectively. Δt and Δx are the time step and space step, respectively.

6. According to claim 1, a two-stage highway vehicle discrete point trajectory reconstruction method based on deep learning is characterized in that: The method embeds the traffic flow trajectory obtained by discrete trajectory point prediction into a deep learning encoder and inputs the historical vehicle trajectory into the Transformer network model to enhance the accuracy of target vehicle trajectory reconstruction, including: The full-time and space traffic flow trajectory predicted by CA-LWR is set as the relevant vehicles around the target vehicle, and each coordinate point is mapped to an embedded vector to form a time series, which is represented by Trajectory near =[(x 1t ,y 1t ),(x 2t ,y 2t ),…(x nt ,y nt )], where (x 1t ,y 1t )、(x 2t ,y 2t )…(x nt ,y nt ) is the coordinate data of vehicle 1, vehicle 2, ..., vehicle n at time t. The trajectory data is processed by Positional Embedding to retain the time series information of the trajectory points; After embedding the encoder with the predicted trajectory as the historical trajectory, multiple Self-Attention Layer layers are used to extract the surrounding vehicle information in the sequence, and the output is used as the input of the decoder; The decoder receives the output of the encoder and dynamically weights the self-attention mechanism in the interaction modeling module to capture the spatial dependency between each vehicle and the target vehicle in the scene. It assigns different attention weights and then inputs the multi-head attention mechanism to learn the vehicle interaction effects from multiple different feature subspaces. The interactive effects of different feature spaces are summarized and aggregated through the weights of the fully connected layer, thereby enhancing the model's fitting ability and adaptability to different scenarios, and generating predictions of future trajectories through the feedforward neural network layer.

7. A two-stage highway vehicle discrete point trajectory reconstruction system based on deep learning, characterized in that: include: The preprocessing module is used to extract features from millimeter radar wave data and then use a density-based outlier detection algorithm to clean abnormal discrete trajectory points; The key trajectory node identification module is used to establish the highway network topology map and consider the Betweenness Centrality method based on graph theory to identify the key trajectory nodes; The first stage trajectory prediction module is used to use the CA-LWR method to predict the full-time and space traffic flow trajectory of the cleaned millimeter radar wave discrete traffic flow trajectory points. The LWR model is used to update the state of the cellular automaton CA. The neighborhood cell speed and flow control the traffic density in each cell. The value range of the vehicle speed in the cellular automaton is controlled according to the maximum speed in the LWR model. For key trajectory nodes, the interaction of surrounding vehicles is considered and the safety distance between vehicles is set; The second-stage trajectory prediction module is used to embed the traffic flow trajectory obtained by discrete trajectory point prediction into the deep learning encoder and input it into the Transformer network model as the historical vehicle trajectory to enhance the accuracy of target vehicle trajectory reconstruction; the Transformer network model is a vehicle trajectory reconstruction model based on the Transformer network with multi-head attention.

8. A computer system comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the computer program is loaded into the processor, the steps of a two-stage highway vehicle discrete point trajectory reconstruction method based on deep learning are implemented according to any one of claims 1 to 6.

9. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of a two-stage highway vehicle discrete point trajectory reconstruction method based on deep learning are implemented according to any one of claims 1 to 6.

10. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the steps of a two-stage highway vehicle discrete point trajectory reconstruction method based on deep learning are implemented according to any one of claims 1-6.

Images