Vehicle trajectory tracing method and system based on highway gantry data

Abstract

This invention discloses a vehicle trajectory tracing method and system based on highway gantry data, belonging to the fields of intelligent transportation and big data analysis technology. By constructing a spatiotemporal association memory network of the road network, the road network is abstracted as a directed graph. A deep neural network model is trained using a graph attention network and a temporal Transformer encoder. Real-time incomplete observation sequences are input into the model, and conditional probability diffusion generation is performed with the observation gantry as a forced anchor point, resulting in the emergence of candidate complete paths and the calculation of confidence scores. Based on the confidence score threshold, the final trajectory is output or anomalies are marked. This invention transforms trajectory tracing from "explicit path search" to "implicit pattern emergence," proactively sensing abnormal behaviors such as toll evasion through confidence scores. It also possesses continuous learning and road network adaptive capabilities, significantly improving the robustness of trajectory reconstruction, the proactivity of anomaly detection, and the system's adaptive capabilities.

Description

Technical Field

[0001] This invention relates to the fields of intelligent transportation and big data analysis technology, and in particular to a method and system for tracing vehicle trajectories based on highway gantry data. Background Technology

[0002] With the widespread adoption of ETC (Electronic Toll Collection) systems on highways, roadside gantry systems have achieved full network coverage. Each ETC gantry automatically records information such as license plate number, passage time, and vehicle characteristics as a vehicle passes through, generating massive amounts of vehicle traffic data. Tracing vehicle trajectories based on this gantry data—that is, reconstructing the actual travel path of a vehicle from the entrance to the exit—has become a core technology for refined highway operation, with significant application value in scenarios such as toll collection auditing, accident handling, and road network optimization.

[0003] Currently, vehicle trajectory tracing methods based on gantry data mainly include: 1) Explicit Reconstruction Based on Graph Theory and Shortest Path Algorithms: This type of method abstracts the highway network as a directed graph and uses Dijkstra's algorithm, A's algorithm, or K's shortest path algorithm to find the optimal path between the starting point and the ending point, and uses this path to fill in the missing gantries in the observation data. For example, existing research published "Exploration of the Application of Highway Gantry Capture Data Traffic Analysis" uses A's algorithm with mandatory points for trajectory reconstruction. However, its limitations are: when the gantry miss rate reaches more than 20%, the path completed by simply relying on the shortest path deviates significantly from the actual path; and the algorithm is sensitive to the scale of the road network, with low computational efficiency on provincial road networks (thousands of gantries).

[0004] 2) Intelligent Tracing Based on Multi-Source Data Fusion: This type of method integrates multi-source information such as ETC gantry data, toll station capture data, and GPS / BeiDou data, and reconstructs the trajectory through spatiotemporal logic verification and machine learning models. For example, CN119672965A discloses: A Method and System for Reconstructing the Trajectory of Key Vehicles, whose technical solution adopts mask self-supervised multimodal fusion governance technology, improving the trajectory reconstruction rate to 98%. However, its limitations are: it is highly dependent on the quality of multi-source data, and there is still a risk of failure in scenarios where ETC signals are completely lost or GPS data is missing; moreover, existing methods are mostly passive reconstruction paradigms of "collecting data first and then splicing paths", which lacks sensitivity to abnormal trajectories (such as toll evasion) and requires additional configuration of audit rules.

[0005] In summary, existing tracing methods generally suffer from insufficient accuracy in path completion when data is missing, and they also face technical challenges such as the trade-off between computational efficiency and road network scale, and a lack of proactive detection capabilities for abnormal trajectories. To address these issues, a vehicle trajectory tracing method and system based on highway gantry data is proposed. Summary of the Invention

[0006] The main objective of this invention is to provide a vehicle trajectory tracing method and system based on highway gantry data, which can effectively solve the problems in the background art.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: In a first aspect, the present invention provides a vehicle trajectory tracing method based on highway gantry data, comprising the following steps: Step S1: Construct a road network spatiotemporal association memory network The highway network is abstracted as a directed graph, where nodes represent gantries or toll stations and edges represent adjacent reachable road segments; Acquire historical gantry data and extract complete vehicle passage trajectories as training samples; A deep neural network model was constructed and trained using a graph attention network and a temporal Transformer encoder, serving as a spatiotemporal association memory network for the road network. The network model takes as input a partially observed gantry sequence and its timestamps and environmental feature vectors as input, and outputs as the probability distribution and emergence confidence of the complete gantry sequence.

[0008] Specifically, the environmental feature vector includes at least the following dimensions of real-time information: the hour and day of the week encoding of the current time, the weather type encoding, the holiday markers, and the real-time congestion index of the road segment where the observation gantry is located.

[0009] Specifically, the training process of the road network spatiotemporal association memory network includes: For complete trajectory samples, 20%-30% of the intermediate gantry is randomly masked to construct incomplete observation sequences as model input. The original complete trajectory is used as the supervision signal, and training is performed using a negative log-likelihood loss function. Simultaneously, a contrastive learning loss function is introduced, using the true trajectory and randomly sampled erroneous paths as negative sample pairs to enhance the model's discriminative ability. The total loss function is: ,in For negative log-likelihood loss, For anchor point constraint loss, To compare learning loss, , These are the weighting coefficients.

[0010] Specifically, the model architecture of the road network spatiotemporal correlation memory network includes: One or more graph attention network layers are used to learn the static spatial embeddings of nodes from the road network topology; A temporal Transformer encoder is used to encode the time intervals and environmental features of the observation sequence to obtain a context vector; A GRU-based conditional path generation decoder autoregressively generates complete gantry sequences based on context vectors and static spatial embeddings.

[0011] Step S2: Online Trajectory Emergence Obtain a real-time incomplete observation sequence of the vehicle to be traced, which includes at least two observation gantries and their elapsed time; The observed sequence and current environmental features are input into a trained road network spatiotemporal association memory network to perform a conditional probability diffusion generation process; Using the observed gantry as the forced anchor point, a continuous gantry sequence is generated hop by hop, resulting in one or more candidate complete paths, and the emergence confidence of each candidate path is calculated.

[0012] Specifically, a beam search algorithm is used with a beamwidth of B. In each generation step, the conditional probability of transitioning to each neighboring node is calculated based on the current node and the context vector.

[0013] Specifically, conditional probability is defined as: ,in For the current node, As a candidate next node, For context vectors, For the generated partial path, It is a learnable scoring function; Specifically, when the minimum number of remaining steps is equal to 1, the next observation node is forcibly selected as the generated result.

[0014] Specifically, the emergence confidence score is used to quantify the reliability of the generated path by the model. A confidence score close to 1 indicates that the path is highly reliable, while a confidence score close to 0 indicates that the path is abnormal. Specifically, the emergence confidence level is defined as: ,in For the Sigmoid function, Let τ be the mean log-likelihood of path τ. For the expected log-likelihood baseline of a uniform random path, , This is an adjustable parameter.

[0015] Step S3: Result Determination and Output The emergence confidence is compared with a preset threshold, and the path output strategy is determined based on the comparison result, wherein the first threshold is preferably 0.85 and the second threshold is preferably 0.6; The specific path output strategy is as follows: When the emergence confidence is greater than or equal to 0.85, the path is output directly; When the emergence confidence level is between 0.6 and 0.85, output the path and mark it as a recommended manual sampling. When the emergence confidence level is less than 0.6, it is marked as an abnormal trajectory candidate, the Top-K candidate paths are output and an alarm is triggered.

[0016] Step S4: Continuous learning and model updates.

[0017] We collect real trajectory samples confirmed by manual review and abnormal trajectory samples confirmed by the audit system. Using an experience replay mechanism, we mix new samples with samples from the historical buffer pool to incrementally train the road network spatiotemporal association memory network. Simultaneously, we employ an elastic weight consolidation regularization method, adding parameter importance weights to the loss function to prevent catastrophic forgetting. The model is updated daily or weekly.

[0018] Furthermore, when the road network structure changes, including the opening of new road sections, the addition or removal of gantries, the following adaptive update process is executed: Reconstruct the directed graph structure and update the node and edge sets; The static embedding layer parameters of the original model are retained as initialization, and only the newly added gantry nodes are randomly initialized; Use at least one week of new data to make local fine-tuning on the new road network structure to enable the model to quickly adapt to changes in the road network.

[0019] Secondly, the present invention provides a vehicle trajectory tracing system based on highway gantry data for implementing the above method, comprising: Data preprocessing module: used to obtain the gantry topology of the highway network, collect historical gantry data and perform cleaning, stroke segmentation and integrity labeling, and generate training sample set; Model training module: used to construct a deep neural network model composed of a graph attention network and a temporal Transformer encoder, and to train it using the training sample set to obtain a road network spatiotemporal association memory network; Online inference module: It is used to receive the real-time incomplete observation sequence of the vehicle to be traced, input the observation sequence and the current environmental features into the trained road network spatiotemporal association memory network, execute the conditional probability diffusion generation process, use the observation gantry as the forced anchor point, generate a continuous gantry sequence hop by hop, generate one or more candidate complete paths, and calculate the emergence confidence of each candidate path. Result determination module: used to compare the emergence confidence with a preset threshold, and output the final trajectory or mark abnormal trajectory candidates based on the comparison result; The continuous learning module is used to collect real trajectory samples and abnormal trajectory samples after manual review. It adopts an experience replay mechanism and elastic weight consolidation regularization method to incrementally train and dynamically update the road network spatiotemporal association memory network.

[0020] Technical effect Compared with the prior art, the present invention has the following beneficial effects: This invention transforms the trajectory tracing problem into a "maximum likelihood emergence problem that satisfies observation constraints" by constructing an implicit spatiotemporal associated memory network. Even if the gantry miss rate reaches 20%-30%, it can still maintain high confidence in trajectory reconstruction, which is superior to existing methods based on explicit shortest path search. When abnormal behaviors such as toll evasion and U / J-shaped driving occur, they are essentially low-probability paths. In the method of this invention, low emergence confidence will be generated naturally, and proactive alarms can be achieved without additional configuration of audit rules. After the model training in this invention is completed, the computational complexity of a single trajectory emergence is close to constant level O(1), which is far superior to the performance of search algorithms such as Dijkstra on large-scale road networks and meets the requirements of real-time tracing. Through a continuous learning mechanism, the model can adapt to changes in road network structure (new road openings, gantry additions and subtractions) and traffic flow patterns (holidays, construction, etc.) without the need for manual parameter readjustment. This invention no longer views trajectory tracing as a path search problem, but rather as a pattern completion problem. The technical solution represents a technological leap from data recording to data cognition. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating the vehicle trajectory tracing method based on highway gantry data according to the present invention. Figure 2 This is a structural block diagram of the vehicle trajectory tracing system based on highway gantry data according to the present invention; Figure 3 This is a schematic diagram of the training process of the road network spatiotemporal correlation memory network model of the present invention; Figure 4 This is a schematic diagram of the online trajectory emergence process of the present invention; Figure 5 This is a schematic diagram of the confidence level determination and result output process of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0023] Example 1: In this embodiment, a highway network is used as an example. This network has N=1,247 gantry nodes and E=3,856 directed edges. Historical gantry data for six months, from January to June 2024, is obtained, resulting in approximately 1.2 billion original records. Based on the above, the implementation process of the vehicle trajectory tracing method based on highway gantry data provided by this invention is described, as follows: Figure 1 , Figure 3 , Figure 4 , Figure 5 As shown, it includes the following steps: Step 1, Data Preparation and Preprocessing The collected data undergoes preprocessing, and the preprocessing process is as follows: The original records are grouped by license plate number, and each group is sorted in ascending order by timestamp. Records with out-of-order timestamps or obvious license plate recognition errors (such as length not conforming to the verification rules) will be removed. Based on a threshold of more than 30 minutes between adjacent records, the daily passage records of the same vehicle are divided into independent trips; For each independent trip, if the distance between adjacent observation gantry in the topology map is greater than 1 (i.e. there is an unobserved gantry in between), it is marked as a missed shot; A total of approximately 8.5 million valid travel samples were obtained, of which about 65% were complete trajectories (without any missed shots), and 35% had varying degrees of missed shots.

[0024] Step 2: Construct a road network spatiotemporal association memory network Specifically, the highway network is abstracted as a directed graph, where nodes represent gantries or toll stations and edges represent adjacent reachable road segments. At least three months of historical gantry data are acquired, and complete vehicle travel trajectories are extracted as training samples. A deep neural network model is constructed and trained using a graph attention network and a temporal Transformer encoder as the spatiotemporal association memory network of the road network. The input of the network model consists of partially observed gantry sequences and their timestamps and environmental feature vectors, and the output consists of the probability distribution and emergence confidence of the complete gantry sequence.

[0025] The training process of the road network spatiotemporal association memory network includes: For complete trajectory samples, 20%-30% of the intermediate gantry is randomly masked to construct incomplete observation sequences as model input. The original complete trajectory is used as the supervision signal, and training is performed using a negative log-likelihood loss function. Simultaneously, a contrastive learning loss function is introduced, using the true trajectory and randomly sampled erroneous paths as negative sample pairs to enhance the model's discriminative ability. The total loss function is: ,in For negative log-likelihood loss, For anchor point constraint loss, To compare learning loss, , These are the weighting coefficients.

[0026] The specific model architecture of the road network spatiotemporal correlation memory network includes: One or more graph attention network layers are used to learn the static spatial embeddings of nodes from the road network topology; A temporal Transformer encoder is used to encode the time intervals and environmental features of the observation sequence to obtain a context vector; A GRU-based conditional path generation decoder autoregressively generates complete gantry sequences based on context vectors and static spatial embeddings.

[0027] In this embodiment, as Figure 2 As shown, the model architecture specifically includes: 2.1) Graph Attention Network Layer (GAT): Initial feature vector of nodes Includes: gantry type encoding (interchange hub / provincial border / regular gantry, one-hot encoding), number of lanes (normalized), longitude coordinates (normalized), and latitude coordinates (normalized). A 2-layer GAT is used, with 128 hidden dimensions per layer, and 4 multi-head attention heads. The output is the static spatial embedding of the nodes. .

[0028] 2.2) Timing Transformer Encoder For each observation point in the observation sequence Construct dynamic feature vectors : ,in, This is a weather type code, with values ​​of 0, 1, 2, and 3, corresponding to sunny, cloudy, rainy, and snowy days, respectively. This is a holiday indicator, with a value of 0 or 1, corresponding to holidays and weekdays respectively. The value is between 0 and 1 to observe the real-time congestion index of the road section where the gantry is located.

[0029] Embedding static space With dynamic feature vectors After splicing, a linear transformation is performed to obtain The Transformer encoder uses four layers, with eight attention heads per layer, and outputs a context vector. .

[0030] 2.3) Conditional Path Generation Decoder The decoder is based on GRU, with a hidden state dimension of 256. At each step t, the input is the static embedding of the current node. and context vector Update hidden status Then calculate the score for moving to each neighbor node. : ,in , All of these are learnable parameters.

[0031] Model Training: In this embodiment, the Adam optimizer is used with an initial learning rate of 0.001, a batch size of 256, and a training duration of 50 epochs. The training set consists of 6 million run samples, the validation set of 1.5 million, and the test set of 1 million. The loss function weights are λ1=0.5 and λ2=0.1. Training is performed on four NVIDIA V100 GPUs.

[0032] Step 3, Online Trajectories Emerge Specifically, the real-time incomplete observation sequence of the vehicle to be traced is obtained, which includes at least two observation gantries and their elapsed times; the observation sequence and current environmental features are input into a trained road network spatiotemporal association memory network, and a conditional probability diffusion generation process is performed; with the observation gantries as forced anchor points, a continuous gantry sequence is generated hop by hop, one or more candidate complete paths emerge, and the emergence confidence of each candidate path is calculated.

[0033] The conditional probability is calculated using a beam search algorithm with a beamwidth of B. In each generation step, the conditional probability of transitioning to each neighboring node is calculated based on the current node and the context vector. The conditional probability is defined as: ,in For the current node, As a candidate next node, For context vectors, For the generated partial path, It is a learnable scoring function; when the minimum number of remaining steps is equal to 1, the next observation node is forced to be selected as the generated result.

[0034] Emergent confidence is used to quantify the reliability of the generated path by the model. A confidence score close to 1 indicates a highly reliable path, while a confidence score close to 0 indicates an abnormal path. Emergent confidence is defined as: ,in For the Sigmoid function, Let τ be the mean log-likelihood of path τ. For the expected log-likelihood baseline of a uniform random path, , This is an adjustable parameter.

[0035] The emergence confidence level is compared with a preset threshold, and the path output strategy is determined based on the comparison result, where the first threshold is 0.85 and the second threshold is 0.6; the specific path output strategy is as follows: When the emergence confidence is greater than or equal to 0.85, the path is output directly; When the emergence confidence level is between 0.6 and 0.85, output the path and mark it as a recommended manual sampling. When the emergence confidence level is less than 0.6, it is marked as an abnormal trajectory candidate, the Top-K candidate paths are output and an alarm is triggered.

[0036] In this embodiment, taking a vehicle whose actual driving path is A→B→C→D→E (gantry number) as an example, it is assumed that the system observes the incomplete sequence: A (08:05:22), C (08:23:15), E (08:52:40), where B and D are missed.

[0037] Step 3.1) Encode the observation sequence to obtain the current environmental characteristics (Wednesday morning at 8:00 AM, sunny, light congestion on AC section).

[0038] Step 3.2) Perform beam search (beamwidth B=5): Starting from A, calculate the transition probability to all neighboring nodes, and keep the top-5. Generate hop by hop, and force the user to pass through C when approaching C; It continues to generate from C, eventually terminating at E.

[0039] Step 3.3) The model generates 3 candidate paths: Path 1: A→B→C→D→E (Conditional probability 0.72, confidence level 0.91) Path 2: A→B'→C→D→E (Conditional probability 0.15, confidence level 0.52) Path 3: A→B→C→D'→E (Conditional probability 0.08, confidence level 0.38) Step 3.4) The confidence level of path 1 is 0.91 > 0.85, so it is directly output as the final trajectory, which is completely consistent with the actual path of the vehicle.

[0040] Step 4, Anomaly Detection Case Another vehicle attempted to evade tolls, with an actual travel path of A→B→C→X→Y→C→D→E (a U-shaped route, turning back between C and D), and the observed sequence was A, C, D, E. Since U-shaped travel has an extremely low probability in a normal road network, the confidence level of the optimal path A→B→C→D→E emerging from the model was only 0.31 < 0.6. The system automatically marked it as an abnormal trajectory candidate and issued an alarm. After manual review, it was confirmed that the vehicle had evaded tolls.

[0041] Step 5, Performance Evaluation The evaluation was conducted on a test set (1 million trips, manually labeled with real routes), and the results are shown in the table below:

[0042] The P99 latency for a single trajectory tracing is 47ms, which meets the real-time business requirements.

[0043] Step 6, Continuous Learning and Adaptation After three months of deployment and operation, approximately 20,000 manually reviewed samples were collected, including 15,000 corrected samples (low model confidence but actually correct) and 5,000 outlier samples. Incremental training (once a week, with fine-tuning for 2 epochs each time) was employed, combined with experience replay (buffer pool capacity of 100,000 samples) and EWC regularization (λ). EWC =1000).

[0044] Tests show that the model maintains an accuracy of over 94% on the new data, and there is no catastrophic forgetting—the accuracy on the old test set three months ago only dropped by 0.3 percentage points.

[0045] When a new hub interchange is added to the road network (adding 5 gantries and 12 directed edges), an adaptive update is performed: the directed graph is reconstructed, retaining the original model parameters, and the static embedding of the new gantry is initialized to the average value of its neighboring nodes. Fine-tuning is then performed using one week's worth of new data (approximately 80,000 trips). After the update, the trajectory reconstruction accuracy for the new road segment reaches 89%, and improves to 93% after one week as data accumulates.

[0046] Example 2: Comparison with other advanced methods On the same dataset, the method of this invention is compared with the following methods: Method A: The A* algorithm with required points is used. Method B: A multi-source data fusion method (GPS assistance required) Method C: Map matching method based on Hidden Markov Model

[0047] *Note: Method B has an accuracy of 94.8% when GPS is available, but drops to 82.3% when GPS is unavailable.

[0048] Industrial applicability The method provided by this invention can be deployed on servers or cloud computing platforms in highway operation and management centers, serving as a core functional module of toll collection and auditing systems, road network monitoring systems, and emergency response systems. Specific application scenarios include: Toll collection audit: Automatically identifies toll evasion behaviors such as card swapping, U / J-shaped driving, and buying short-distance tickets for long-distance trips, outputs abnormal alarms and confidence levels, and assists auditors in quickly locating suspicious vehicles.

[0049] Accident tracing: After a traffic accident occurs, quickly locate the entire route of the vehicle involved to provide a complete chain of evidence in terms of time and space for determining liability.

[0050] Road network operation monitoring: Through large-scale trajectory reconstruction, analyze the distribution of traffic flow paths, identify congested bottleneck sections and abnormal traffic patterns, and provide decision support for easing congestion and ensuring smooth traffic flow.

[0051] Key vehicle tracking: Real-time trajectory monitoring of key vehicles such as hazardous chemical transport vehicles, passenger vehicles and hazardous materials transport vehicles, and proactive warnings when they deviate from the predetermined route or stay for an extended period of time.

[0052] This invention does not rely on GPS or other auxiliary positioning devices, and can operate using only existing ETC gantry data. It has the advantages of low deployment cost, wide coverage, and easy promotion.

[0053] Example 2: like Figure 2 As shown, the present invention also provides a vehicle trajectory tracing system based on highway gantry data for implementing the above method, comprising: Data preprocessing module: used to obtain the gantry topology of the highway network, collect historical gantry data and perform cleaning, stroke segmentation and integrity labeling, and generate training sample set; Model training module: used to construct a deep neural network model composed of a graph attention network and a temporal Transformer encoder, and to train it using a training sample set to obtain a road network spatiotemporal association memory network; Online inference module: It is used to receive the real-time incomplete observation sequence of the vehicle to be traced, input the observation sequence and the current environmental features into the trained road network spatiotemporal association memory network, execute the conditional probability diffusion generation process, use the observation gantry as the forced anchor point, generate a continuous gantry sequence hop by hop, generate one or more candidate complete paths, and calculate the emergence confidence of each candidate path. Result determination module: used to compare the emergence confidence with a preset threshold, and output the final trajectory or mark abnormal trajectory candidates based on the comparison result; The continuous learning module is used to collect real trajectory samples and abnormal trajectory samples after manual review. It adopts an experience replay mechanism and elastic weight consolidation regularization method to incrementally train and dynamically update the road network spatiotemporal association memory network.

[0054] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A vehicle trajectory tracing method based on highway gantry data, characterized in that, Includes the following steps: S1: The highway network is abstracted as a directed graph with gantries or toll stations as nodes and adjacent reachable road segments as edges; historical gantry data is acquired, and complete vehicle travel trajectories are extracted as training samples; a deep neural network model is constructed and trained as a spatiotemporal association memory network of the road network using a graph attention network and a temporal Transformer encoder; the input of the network model is the partially observed gantry sequence and its timestamps and environmental feature vectors, and the output is the probability distribution and emergence confidence of the complete gantry sequence; S2: Obtain the real-time incomplete observation sequence of the vehicle to be traced, the sequence containing at least two observation gantries and the elapsed time; The observed sequence and current environmental features are input into the trained road network spatiotemporal association memory network, and the conditional probability diffusion generation process is executed. The observed gantries are used as forced anchor points to generate continuous gantry sequences hop by hop, resulting in at least one candidate complete path. The emergence confidence of each candidate path is calculated. S3: Compare the emergence confidence with a preset threshold and determine the path output strategy based on the comparison result.

2. The vehicle trajectory tracing method based on highway gantry data according to claim 1, characterized in that, The training process of the road network spatiotemporal association memory network includes: For the complete trajectory samples, 20%-30% of the intermediate gantry is randomly masked to construct incomplete observation sequences as model input. The original complete trajectory is used as the supervision signal, and the negative log-likelihood loss function is used for training. A contrastive learning loss function is introduced, which uses the real trajectory and the randomly sampled error path as negative sample pairs; Total loss function Represented as: ,in For negative log-likelihood loss, For anchor point constraint loss, To compare learning loss, , These are the weighting coefficients.

3. The vehicle trajectory tracing method based on highway gantry data according to claim 1, characterized in that, The conditional probability diffusion generation process specifically includes: employing a beam search algorithm with a beamwidth of B, and in each generation step, calculating the conditional probability of transitioning to each neighboring node based on the current node and the context vector; the conditional probability is defined as: ,in For the current node, As a candidate next node, For context vectors, For the generated partial path, It is a learnable scoring function; When the minimum number of remaining steps is 1, the next observation node is forcibly selected as the result.

4. The vehicle trajectory tracing method based on highway gantry data according to claim 1, characterized in that, The emergence confidence score is used to quantify the reliability of the generated path by the model. A confidence score close to 1 indicates that the path is highly reliable, and a confidence score close to 0 indicates that the path is abnormal. It is defined as follows: ,in For the Sigmoid function, Let τ be the mean log-likelihood of path τ. For the expected log-likelihood baseline of a uniform random path, , This is an adjustable parameter.

5. The vehicle trajectory tracing method based on highway gantry data according to claim 1, characterized in that, The environmental feature vector includes at least the following real-time information dimensions: the hour and day of the week encoding of the current time, the weather type encoding, the holiday marker, and the real-time congestion index of the road segment where the observation gantry is located.

6. The vehicle trajectory tracing method based on highway gantry data according to claim 1, characterized in that, Also includes: Collect real trajectory samples confirmed by manual review, as well as abnormal trajectory samples confirmed by the audit system; An experience replay mechanism is used to mix new samples with samples in the historical buffer pool to incrementally train the road network spatiotemporal association memory network. Meanwhile, an elastic weighted regularization method is adopted, and a parameter importance weighting term is added to the loss function to prevent catastrophic forgetting; The model is updated daily or weekly.

7. The vehicle trajectory tracing method based on highway gantry data according to claim 1, characterized in that, The first threshold is 0.85, and the second threshold is 0.6; The specific steps for determining the path output strategy based on the comparison results are as follows: When the emergence confidence is greater than or equal to 0.85, the path is output directly; When the emergence confidence level is between 0.6 and 0.85, output the path and mark it as a recommended manual sampling. When the emergence confidence level is less than 0.6, it is marked as an abnormal trajectory candidate, the Top-K candidate paths are output and an alarm is triggered.

8. The vehicle trajectory tracing method based on highway gantry data according to claim 1, characterized in that, The model architecture of the road network spatiotemporal correlation memory network includes: One or more graph attention network layers are used to learn the static spatial embeddings of nodes from the road network topology; A temporal Transformer encoder is used to encode the time intervals and environmental features of the observation sequence to obtain a context vector; A GRU-based conditional path generation decoder autoregressively generates complete gantry sequences based on context vectors and static spatial embeddings.

9. The vehicle trajectory tracing method based on highway gantry data according to claim 1, characterized in that, When the road network structure changes, including the opening of new road sections, the addition or removal of gantries, the following adaptive update process is executed: Reconstruct the directed graph structure and update the node and edge sets; The static embedding layer parameters of the original model are retained as initialization, and only the newly added gantry nodes are randomly initialized; Use at least one week of new data to make local fine-tuning on the new road network structure to enable the model to quickly adapt to changes in the road network.

10. A vehicle trajectory tracing system based on highway gantry data, used to implement the method as described in any one of claims 1-9, characterized in that, include: The data preprocessing module is used to obtain the gantry topology of the highway network, collect historical gantry data and perform cleaning, stroke segmentation and integrity labeling to generate a training sample set; The model training module is used to construct a deep neural network model composed of a graph attention network and a temporal Transformer encoder, and to train it using the training sample set to obtain a road network spatiotemporal association memory network. The online inference module is used to receive the real-time incomplete observation sequence of the vehicle to be traced, input the observation sequence and the current environmental features into the trained road network spatiotemporal association memory network, execute the conditional probability diffusion generation process, use the observation gantry as the forced anchor point, generate a continuous gantry sequence hop by hop, generate one or more candidate complete paths, and calculate the emergence confidence of each candidate path. The result determination module is used to compare the emergence confidence with a preset threshold and output the final trajectory or mark abnormal trajectory candidates based on the comparison result. The continuous learning module is used to collect real trajectory samples and abnormal trajectory samples after manual review. It adopts an experience replay mechanism and elastic weight consolidation regularization method to incrementally train and dynamically update the road network spatiotemporal association memory network.

Images