Abnormal sub-trajectory detection method and system based on space-time diagram convolutional network

Abstract

The invention discloses an abnormal sub-trajectory detection method and system based on a space-time diagram convolutional network, and the method specifically comprises the following steps: S1, obtaining the original movement trajectory data of a vehicle, and carrying out the preprocessing of the original movement trajectory data, and obtaining a movement trajectory data set; s2, acquiring a driving track data set of the vehicle in a passenger carrying state according to the moving track data set, and processing the driving track data set to obtain a target track data set; s3, according to the target trajectory data set, constructing a fine-grained online abnormal sub-trajectory detection model based on a space-time diagram convolutional network; and S4, detecting the target movement trajectory data of the vehicle by using the abnormal sub-trajectory detection model to obtain an abnormal sub-trajectory sequence.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation technology, and in particular to an abnormal sub-trajectory detection method based on spatiotemporal graph convolutional networks. Background Technology

[0002] With the rapid development of mobile internet and positioning technology, mobile trajectory big data is widely used in data-driven smart cities (intelligent transportation), urban computing, and other fields. Accurately identifying abnormal trajectories from large-scale trajectory data has significant practical implications and application value for urban traffic control, travel behavior analysis, and emergency warning.

[0003] In existing technologies, traditional detection methods mostly rely on statistical modeling or cluster analysis, which can detect abnormal trajectories to a certain extent. However, they can usually only make anomaly judgments based on the entire trajectory and are difficult to accurately identify local abnormal sub-trajectory sequences. At the same time, mobile trajectory big data generally suffers from problems such as noise, drift, or missing data, and most traditional identification methods perform poorly in terms of robustness.

[0004] With the development of deep learning, spatiotemporal graph neural network-based methods have been gradually applied to anomaly trajectory detection. These methods construct graph structures to model the spatial relationships between trajectory points and combine them with time series information for joint learning, achieving good results in practical traffic tasks such as traffic prediction and crowd flow analysis. However, in the application of anomaly trajectory detection, existing spatiotemporal graph neural network-based methods may still face the following challenges: (1) Some methods only focus on single spatial or temporal features and fail to fully integrate the complex dependencies between the two, thus limiting the detection accuracy; (2) Some methods have weak ability to identify local anomalies in the trajectory, resulting in accurate overall judgment but insufficient precision in fine-grained detection; at the same time, when facing sparse and imbalanced anomaly samples in real-world scenarios, the models often have insufficient generalization ability, resulting in a high false alarm rate or false negative rate; (3) Some methods have high computational complexity, making it difficult to meet the application requirements of real-time detection and online updates. Summary of the Invention

[0005] To address the problems of low accuracy, coarse granularity, and weak real-time performance in existing abnormal trajectory detection technologies, this invention proposes an abnormal sub-trajectory detection method based on spatiotemporal graph convolutional networks. This method can locate and interpret "when, when, and what type" of anomalies during continuous vehicle operation, significantly reducing false alarms and false negatives, improving detection accuracy, and possessing self-evolution capabilities to cope with dynamic changes in cities.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] An anomaly sub-trajectory detection method based on spatiotemporal graph convolutional networks is characterized by the following steps:

[0008] S1: Obtain the vehicle's original movement trajectory data, preprocess the original movement trajectory data, and obtain the movement trajectory dataset;

[0009] S2: Obtain the vehicle's driving trajectory dataset under passenger-carrying conditions based on the mobile trajectory dataset, and process the driving trajectory dataset to obtain the target trajectory dataset;

[0010] S3: Based on the target trajectory dataset, construct a fine-grained online abnormal sub-trajectory detection model based on a spatiotemporal graph convolutional network;

[0011] S4: An abnormal sub-trajectory detection model is used to detect the target movement trajectory data of the vehicle to obtain an abnormal sub-trajectory sequence.

[0012] Preferably, S1 includes:

[0013] S1-1: Collect the vehicle's original movement trajectory data, which includes at least the vehicle identifier, latitude and longitude coordinates, timestamp, and operating status;

[0014] S1-2: Preprocess the original movement trajectory data to obtain preprocessed data. Preprocessing includes missing values, drift point removal, coordinate format unification, timestamp synchronization, simple interpolation, and noise reduction.

[0015] S1-3: Based on vehicle identification and time sequence, the preprocessed data is classified and sorted by length to obtain the movement trajectory dataset.

[0016] Preferably, in steps S1-3, trajectory data with the same starting point and ending point after preprocessing are grouped into the same category.

[0017] Preferably, S2 includes:

[0018] S2-1: In the mobile trajectory dataset, sort according to vehicle ID, find the trajectory data with the same vehicle ID and continuous operating status of 111...10, and use it as the first trajectory data; passenger status is recorded as 1, and empty status is recorded as 0;

[0019] S2-2: Remove data from the first trajectory data, retain the vehicle trajectory data under continuous passenger-carrying conditions, and record the latitude, longitude and timestamp corresponding to each trajectory as the second trajectory data;

[0020] S2-3: Perform map matching based on the second trajectory data to obtain the gridded trajectory sequence of the vehicle on the actual road network, which serves as the third trajectory data;

[0021] S2-4: The road network is gridded based on the third trajectory data to obtain grid sequence trajectories and noise labels.

[0022] Preferably, S2-4 includes:

[0023] S2-4-1: Divide the third trajectory data into multiple OD pairs, and then further divide each OD pair into 24 time slots, corresponding to 24 hours in a day; according to the start time of the journey, assign each trajectory to the time slot corresponding to its OD pair;

[0024] S2-4-2: Calculate the transition ratio of all tracks within each time slot group from one sub-track to another:

[0025]

[0026] Among them, P ij Represents the i-th sub-trajectory e i to the j-th sub-trajectory e j The transition probability of N; ij Represents the i-th sub-trajectory e i to the j-th sub-trajectory e j The number of trajectories; N total This indicates the total number of trajectories within that time slot;

[0027] S2-4-3: Calculate the noise label for each sub-trajectory:

[0028]

[0029] Among them, NoiseLabel(e i ) represents the i-th sub-trajectory e i Noise label; O i Represents the i-th sub-trajectory e i The original point; α represents the threshold;

[0030] S2-4-4: The driving trajectory data with the same starting point are identified as trajectory pairs. The driving trajectory in the trajectory pair is enhanced to obtain a grid mapping trajectory represented by a continuous grid sequence.

[0031] S2-4-5: Group the trajectories according to the original point or grid starting point, extract / combine the trajectories with the same starting point and perform data augmentation on them to obtain the target trajectory dataset.

[0032] Preferably, in S2-4-5, data augmentation includes time scale transformation, sequence resampling, local pruning / interpolation, perturbation, or random masking.

[0033] Preferably, S3 includes:

[0034] S3-1: Input the road segment sequence and time slot code into the time feature network to extract the time feature sequence;

[0035] S3-2: Construct a time-varying graph structure based on OD pairs, define the adjacency matrix A and the three-dimensional feature tensor X, and extract the spatial feature sequence;

[0036] S3-3: An attention mechanism is used to weight and fuse the temporal and spatial feature sequences to obtain a spatiotemporal joint representation, and an abnormal sub-trajectory detection model is constructed.

[0037] Preferably, S3-2 includes:

[0038] S3-2-1: Construct a time-varying graph structure based on OD pairs, and define the adjacency matrix A and the three-dimensional feature tensor X(s,d,t), where X(s,d,t) represents the transition probability from the starting point O to the ending point D under interval t;

[0039] S3-2-2: Extracting spatial features using Graph Convolutional Networks (GCN):

[0040] F s =GCN(A,X) (3)

[0041] Among them, F s A represents spatial features; A represents the adjacency matrix; X represents the three-dimensional feature tensor.

[0042] Preferably, S3-3 includes:

[0043] S3-3-1: Weighted fusion of temporal and spatial feature sequences yields a joint spatiotemporal representation:

[0044] F st =α t ·F t +α s ·F s (4)

[0045] Among them, F st Represents the spatiotemporal joint representation; α t The attention weights of the time feature sequence; F t Represents a time-feature sequence; F s Represents a spatial feature sequence; α s Attention weights representing spatial feature sequences;

[0046] S3-3-2: Pre-train the model using the generated spatiotemporal joint representation; then train DENet; then use DENet's predictions as labels to back-train the model, iterating alternately until the validation set index converges, and construct an abnormal sub-trajectory detection model.

[0047] This invention also provides an anomaly sub-trajectory detection system based on spatiotemporal graph convolutional networks, comprising:

[0048] The trajectory acquisition module is used to acquire the vehicle's raw movement trajectory data and preprocess it to obtain a movement trajectory dataset;

[0049] The trajectory processing module is used to extract the driving trajectory dataset based on the passenger status obtained from the mobile trajectory dataset, and to perform map matching, gridding, and trajectory enhancement processing to obtain the target trajectory dataset.

[0050] The model building module is used to build an abnormal sub-trajectory detection model based on a spatiotemporal graph convolutional network based on the target trajectory dataset;

[0051] The anomaly detection module is used to detect the target trajectory using an anomaly sub-trajectory detection model and output anomaly sub-trajectory sequences.

[0052] In summary, by adopting the above technical solution, the present invention has at least the following beneficial effects compared with the prior art:

[0053] By constructing an anomaly sub-trajectory detection model based on spatiotemporal graph convolutional networks, and combining noise self-supervised label generation, weakly supervised iterative training, and online incremental update mechanisms, fine-grained real-time detection of anomaly sub-trajectories is achieved. It requires no manual annotation, has online inference capabilities, and features fine detection granularity, low latency, and strong adaptability. It effectively solves the problems of coarse anomaly detection granularity, inability to process in real time online, and reliance on a large number of manual labels in existing technologies. Attached image description:

[0054] Figure 1 This is a schematic diagram of an anomaly sub-trajectory detection method based on a spatiotemporal graph convolutional network according to an exemplary embodiment of the present invention.

[0055] Figure 2 This is a schematic diagram of the preprocessing flow of raw movement trajectory data according to an exemplary embodiment of the present invention.

[0056] Figure 3 This is a schematic diagram illustrating the processing of a driving trajectory dataset according to an exemplary embodiment of the present invention.

[0057] Figure 4 This is a schematic diagram illustrating the construction process of an abnormal sub-trajectory detection model according to an exemplary embodiment of the present invention.

[0058] Figure 5 This is a schematic diagram of the abnormal sub-trajectory detection result output process according to an exemplary embodiment of the present invention.

[0059] Figure 6 This is a schematic diagram of an anomaly sub-trajectory detection system based on a spatiotemporal graph convolutional network according to an exemplary embodiment of the present invention. Detailed Implementation

[0060] The present invention will be further described in detail below with reference to embodiments and specific implementation methods. However, it should not be construed as the scope of the above-mentioned subject matter of the present invention being limited to the following embodiments. All technologies implemented based on the content of the present invention are within the scope of the present invention.

[0061] In the description of this invention, it should be understood that the terms "longitudinal," "lateral," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0062] like Figure 1 As shown, the present invention provides an anomaly sub-trajectory detection method based on spatiotemporal graph convolutional networks, which specifically includes the following steps.

[0063] S1: Obtain the vehicle's original movement trajectory data, preprocess the original movement trajectory data, and obtain the movement trajectory dataset.

[0064] In this embodiment, as Figure 2 As shown, S1 specifically includes the following steps:

[0065] S1-1: Collect the vehicle's original movement trajectory data, which includes at least the vehicle identifier, latitude and longitude coordinates, timestamp, and operating status (carrying passengers / empty vehicle).

[0066] In this embodiment, the operating status includes a passenger-carrying status (recorded as 1) and an empty vehicle status (recorded as 0).

[0067] S1-2: Preprocess the original trajectory data to obtain preprocessed data. Preprocessing includes removing missing values / drift points, unifying coordinate formats, synchronizing timestamps, and performing simple interpolation or denoising to ensure the continuity and reliability of trajectory points.

[0068] S1-3: Based on vehicle identification and time sequence, the preprocessed data is classified (those with the same starting point O and ending point D are classified into the same category) and sorted by length (the extracted trajectory data is sorted according to short, medium and long) to obtain the movement trajectory dataset.

[0069] S2: Obtain the vehicle's driving trajectory dataset under passenger-carrying conditions based on the mobile trajectory dataset, and process the driving trajectory dataset to obtain the target trajectory dataset.

[0070] In this embodiment, data processing includes trajectory map matching, gridding, extraction of trajectory pairs with the same starting point, and trajectory enhancement processing.

[0071] In this embodiment, as Figure 3 As shown, S2 specifically includes the following steps:

[0072] S2-1: Sort the mobile trajectory dataset according to vehicle ID, find the trajectory data with the same vehicle ID and continuous operating status of 111...10, and use them as the first trajectory data.

[0073] The operational status includes passenger-carrying status and empty vehicle status, with passenger-carrying status recorded as 1 and empty vehicle status recorded as 0.

[0074] S2-2: Remove data from the first trajectory data, retain the trajectory data of vehicles in passenger-carrying state (i.e., remove 0 and keep only 1), and record the latitude, longitude and timestamp corresponding to each trajectory as the second trajectory data T.

[0075] S2-3: Perform map matching based on the second trajectory data T to obtain the gridded trajectory sequence T of the vehicle on the actual road network. r , as the third trajectory data.

[0076] S2-3-1: Divide the entire map into W×H grids of equal size, where W and H are obtained from the latitude and longitude of the map edge and the grid size parameters, W represents the width of the grid and H represents the height of the grid.

[0077] S2-3-2: For the trajectory point data P in the second trajectory data T i Mapping to obtain Grid i Thus, the gridded trajectory sequence T is obtained. r ={Grid1,Grid2,…,Grid m ,Pad}, which serves as the third trajectory data; where m is the actual length of the trajectory sequence.

[0078] In this embodiment, to ensure the consistency of trajectory sequence length during model training, Pad is used as a padding character for the gridded trajectory sequence to align the data. The aligned trajectory data T r ∈R n , where n is the maximum length of the trajectory.

[0079] S2-4: Based on the third trajectory data, the road network is gridded to obtain grid sequence trajectories and noise labels, specifically including:

[0080] S2-4-1: Divide the third trajectory data into multiple OD pairs (Origin-Destination); then further divide each OD pair into 24 time slots, corresponding to 24 hours in a day; finally, according to the start time of the journey, assign each trajectory to the time slot corresponding to its OD pair.

[0081] S2-4-2: Calculate the transition ratio from one sub-track to another for all tracks within each time slot group, where sub-track e is the transition ratio. i to e j The transition probability P ij The calculation is as follows:

[0082]

[0083] Among them, P ij Represents the i-th sub-trajectory e i to the j-th sub-trajectory e j The transition probability of N; ij Represents the i-th sub-trajectory e i to the j-th sub-trajectory e j The number of trajectories; N total This indicates the total number of trajectories within that time slot.

[0084] In this embodiment, each trajectory is divided into continuous sub-trajectories based on road segments.

[0085] S2-4-3: Calculate the noise label for each sub-trajectory.

[0086] In this embodiment, the noise label of the sub-trajectory is identified by a threshold parameter α. If the original point O of the sub-trajectory... i If the value is less than the threshold α, the sub-trajectory is marked as noise (1); otherwise, it is marked as noise-free (0). The specific rule is defined as follows:

[0087]

[0088] Among them, NoiseLabel(e i ) represents the i-th sub-trajectory e i Noise label; O i Represents the i-th sub-trajectory e i The original point; α represents the threshold.

[0089] S2-4-4: The driving trajectory data with the same starting point are identified as trajectory pairs. The driving trajectory in the trajectory pair is enhanced to obtain a grid mapping trajectory represented by a continuous grid sequence.

[0090] S2-4-5: Group the trajectories according to the original point or grid starting point, extract / combine trajectories with the same starting point and perform data augmentation on them (such as time scale transformation, sequence resampling, local pruning / interpolation, perturbation or random masking, etc.) to enrich the training samples and obtain the target trajectory dataset.

[0091] S3: Based on the target trajectory dataset, construct a fine-grained online abnormal sub-trajectory detection model based on a spatiotemporal graph convolutional network, such as... Figure 4 As shown, it specifically includes:

[0092] S3-1: Input the road segment sequence and time slot encoding into the time feature network to extract the time feature sequence, including:

[0093] S3-1-1: Initialize the embedding layer of the time feature network at the road segment level; pre-trained road segment vectors such as Toast can be used.

[0094] S3-1-2: The road segment sequence and time slot encoding are fed into an xLSTM model (Extended Long Short-Term Memory, a novel recurrent neural network architecture) that supports variable-length input to extract the time feature sequence F. t .

[0095] S3-2: Construct a time-varying graph structure based on OD pairs, define the adjacency matrix A and the three-dimensional feature tensor X, and extract the spatial feature sequence F. s ,include:

[0096] S3-2-1: Construct a time-varying graph structure based on OD pairs, defining the adjacency matrix A and the three-dimensional feature tensor X(s,d,t), where X(s,d,t) represents the transition probability from the starting point O(o) to the ending point D(d) in time period t.

[0097] S3-2-2: Spatial features are extracted using a Graph Convolutional Network (GCN), with normalized adjacency and single-layer graph convolution, represented as:

[0098] F s =GCN(A,X) (3)

[0099] Among them, F s A represents spatial features; A represents the adjacency matrix; X represents the three-dimensional feature tensor.

[0100] S3-3: An attention mechanism is used to weight and fuse the temporal and spatial feature sequences to obtain a spatiotemporal joint representation, and an abnormal sub-trajectory detection model is constructed.

[0101] Specifically, it includes:

[0102] S3-3-1: Weighted fusion of temporal and spatial feature sequences yields a joint spatiotemporal representation:

[0103] F st =α t ·F t +α s ·F s (4)

[0104] Among them, F st Represents the spatiotemporal joint representation; α t The attention weights of the time feature sequence; F t Represents a time-feature sequence; F s Represents a spatial feature sequence; α s The attention weights represent the spatial feature sequences.

[0105] S3-3-2: Weak supervision and iterative labeling process: The abnormal sub-trajectory detection model is pre-trained using the generated spatiotemporal joint representation; then DENet is trained; finally, the model is back-trained using the predictions of DENet as labels, alternating and iterating until the validation set index converges, thus constructing the abnormal sub-trajectory detection model.

[0106] S4: An abnormal sub-trajectory detection model is used to detect the target movement trajectory data of the vehicle, resulting in an abnormal sub-trajectory sequence. For example... Figure 5 As shown, it specifically includes:

[0107] S4-1: When inspecting a certain road segment e i At that time, its state s i Defined as:

[0108] s i =[z i ; v(e i-1 .l)] (5)

[0109] Among them, s i Indicates road segment e i state; z i Represents the spatiotemporal eigenvector; v(e i-1 .l) represents e i-1 The road segment label embedding vector indicates whether the previous road segment is normal (0) or abnormal (1);

[0110] In this embodiment, the status of a road segment integrates the spatiotemporal characteristics of the current road segment and the label information of the previous moment to capture the continuity and pattern of the trajectory.

[0111] S4-2: An abnormal sub-trajectory detection model is used to construct an online decision state, detect the state of each road segment, and thus give a road segment label (0 for normal and 1 for abnormal).

[0112] S4-3: Assemble the continuous sub-trajectory segments marked as abnormal into an abnormal sub-trajectory sequence, and return or alarm immediately.

[0113] S4-4: Trigger lightweight incremental fine-tuning or periodic batch retraining on demand for online detection results, or add high-confidence detections as pseudo-labels to the training pool.

[0114] S4-5: Set the range of values ​​for online delay budget, sliding window length, pseudo-label update frequency, confidence threshold β, and conservative factors and verification loops to avoid pseudo-label deterioration, so as to ensure detection stability and interpretability.

[0115] In this embodiment, a reward function is also designed, consisting of two parts: a local reward and a global reward. These local and global rewards improve the stability and accuracy of detection. The local reward ensures label continuity and prevents frequent switching between normal and abnormal labels; the global reward evaluates the overall label quality after DENet relabeling.

[0116] In this embodiment, a weakly supervised and iterative labeling process is adopted. First, the abnormal sub-trajectory detection model is pre-trained using the generated noise labels; then, DENet is trained; then, the model is fine-tuned by using the predictions of DENet as labels, and the process is iterated alternately until the validation set index converges.

[0117] In this embodiment, TFSR, SFSR, and DENet are jointly trained: During the pre-training phase, the three networks are pre-trained using trajectory data with noise labels. TFSR and SFSR are responsible for extracting feature representations for each road segment; DENet uses these features and noise labels to learn a policy network to determine whether a road segment is abnormal.

[0118] Based on the above method, the present invention also provides an abnormal sub-trajectory detection system based on a spatiotemporal graph convolutional network, such as... Figure 6 As shown, it includes a trajectory acquisition module, a trajectory processing module, a model building module, and an anomaly detection module. The output of the trajectory acquisition module is connected to the input of the trajectory processing module, the output of the trajectory processing module is connected to the input of the model building module, and the output of the model building module is connected to the input of the anomaly detection module.

[0119] in,

[0120] The trajectory acquisition module is used to acquire the vehicle's raw movement trajectory data and preprocess it to obtain a movement trajectory dataset.

[0121] The trajectory processing module is used to extract the driving trajectory dataset based on the passenger status obtained from the mobile trajectory dataset, and to perform map matching, gridding, and trajectory enhancement processing to obtain the target trajectory dataset.

[0122] The model building module is used to construct an abnormal sub-trajectory detection model based on a spatiotemporal graph convolutional network from the target trajectory dataset.

[0123] The anomaly detection module is used to detect the target trajectory using an anomaly sub-trajectory detection model and output anomaly sub-trajectory sequences.

[0124] Those skilled in the art will understand that the above embodiments are specific examples of implementing the present invention, and in practical applications, various changes can be made in form and detail without departing from the content and scope of the present invention.

Claims

1. An anomaly sub-trajectory detection method based on spatiotemporal graph convolutional networks, characterized in that, Specifically, the following steps are included: S1: Obtain the vehicle's original movement trajectory data, preprocess the original movement trajectory data, and obtain the movement trajectory dataset; S2: Obtain the vehicle's driving trajectory dataset under passenger-carrying conditions based on the mobile trajectory dataset, and process the driving trajectory dataset to obtain the target trajectory dataset; S3: Based on the target trajectory dataset, construct a fine-grained online abnormal sub-trajectory detection model based on a spatiotemporal graph convolutional network; S4: An abnormal sub-trajectory detection model is used to detect the target movement trajectory data of the vehicle to obtain an abnormal sub-trajectory sequence.

2. The abnormal sub-trajectory detection method based on spatiotemporal graph convolutional networks as described in claim 1, characterized in that, S1 includes: S1-1: Collect the vehicle's original movement trajectory data, which includes at least the vehicle identifier, latitude and longitude coordinates, timestamp, and operating status; S1-2: Preprocess the original movement trajectory data to obtain preprocessed data. Preprocessing includes missing values, drift point removal, coordinate format unification, timestamp synchronization, simple interpolation, and noise reduction. S1-3: Based on vehicle identification and time sequence, the preprocessed data is classified and sorted by length to obtain the movement trajectory dataset.

3. The abnormal sub-trajectory detection method based on spatiotemporal graph convolutional networks as described in claim 2, characterized in that, In S1-3, trajectory data with the same starting point and ending point after preprocessing are grouped into the same category.

4. The abnormal sub-trajectory detection method based on spatiotemporal graph convolutional networks as described in claim 1, characterized in that, S2 includes: S2-1: Sort the mobile trajectory dataset according to vehicle ID, find the trajectory data with the same vehicle ID and continuous operating status of 111...10, and use it as the first trajectory data; the passenger status is recorded as 1, and the empty vehicle status is recorded as 0. S2-2: Remove data from the first trajectory data, retain the vehicle trajectory data under continuous passenger-carrying conditions, and record the latitude, longitude and timestamp corresponding to each vehicle trajectory data as the second trajectory data; S2-3: Perform map matching based on the second trajectory data to obtain the gridded trajectory sequence of the vehicle on the actual road network, which serves as the third trajectory data; S2-4: The road network is gridded based on the third trajectory data to obtain grid sequence trajectories and noise labels.

5. The abnormal sub-trajectory detection method based on spatiotemporal graph convolutional networks as described in claim 1, characterized in that, S2-4 includes: S2-4-1: Divide the third trajectory data into K OD pairs, and then further divide each OD pair into 24 time slots; according to the start time of the journey, assign each trajectory to the time slot corresponding to the OD pair; S2-4-2: Calculate the transition ratio from one sub-track to another among all tracks within each time slot group: Among them, P ij Represents the i-th sub-trajectory e i to the j-th sub-trajectory e j The transition probability of N; ij Represents the i-th sub-trajectory e i to the j-th sub-trajectory e j The number of trajectories; N total This indicates the total number of trajectories within that time slot; S2-4-3: Calculate the noise label for each sub-trajectory: Among them, NoiseLabel(e i ) represents the i-th sub-trajectory e i Noise label; O i Represents the i-th sub-trajectory e i The original point; α represents the threshold; S2-4-4: The driving trajectory data with the same starting point are identified as trajectory pairs. The driving trajectory in the trajectory pair is enhanced to obtain a grid mapping trajectory represented by a continuous grid sequence. S2-4-5: Group the trajectories according to the original point or grid starting point, extract / combine trajectories with the same starting point and perform data augmentation on them to obtain the target trajectory dataset.

6. The abnormal sub-trajectory detection method based on spatiotemporal graph convolutional networks as described in claim 5, characterized in that, In S2-4-5, data augmentation includes time scale transformation, sequence resampling, local pruning / interpolation, perturbation, or random masking.

7. The abnormal sub-trajectory detection method based on spatiotemporal graph convolutional networks as described in claim 1, characterized in that, S3 includes: S3-1: Input the road segment sequence and time slot code into the time feature network to extract the time feature sequence; S3-2: Construct a time-varying graph structure based on OD pairs, define the adjacency matrix A and the three-dimensional feature tensor X, and extract the spatial feature sequence; S3-3: An attention mechanism is used to weight and fuse the temporal and spatial feature sequences to obtain a spatiotemporal joint representation, and an abnormal sub-trajectory detection model is constructed.

8. The abnormal sub-trajectory detection method based on spatiotemporal graph convolutional networks as described in claim 7, characterized in that, S3-2 includes: S3-2-1: Construct a time-varying graph structure based on OD pairs, and define the adjacency matrix A and the three-dimensional feature tensor X(s,d,t), where X(s,d,t) represents the transition probability from the starting point O to the ending point D under interval t; S3-2-2: Extracting spatial features using Graph Convolutional Networks (GCN): F s =GCN(A,X) (3) Among them, F s A represents spatial features; A represents the adjacency matrix; X represents the three-dimensional feature tensor.

9. The abnormal sub-trajectory detection method based on spatiotemporal graph convolutional networks as described in claim 7, characterized in that, S3-3 includes: S3-3-1: Weighted fusion of temporal and spatial feature sequences yields a joint spatiotemporal representation: F st =a t ·F t +a s ·F s (4) Among them, F st Represents the spatiotemporal joint representation; α t The attention weights of the time feature sequence; F t Represents a time-feature sequence; F s Represents a spatial feature sequence; α s Attention weights representing spatial feature sequences; S3-3-2: Pre-train the model using the generated spatiotemporal joint representation; then train DENet; and use DENet's predictions as labels to back-train the model, iterating alternately until the validation set index converges, thus constructing an abnormal sub-trajectory detection model.

10. An anomaly sub-trajectory detection system based on a spatiotemporal graph convolutional network according to the method of any one of claims 1-9, characterized in that, include: The trajectory acquisition module is used to acquire the vehicle's raw movement trajectory data and preprocess it to obtain a movement trajectory dataset; The trajectory processing module is used to extract the driving trajectory dataset based on the passenger status obtained from the mobile trajectory dataset, and to perform map matching, gridding, and trajectory enhancement processing to obtain the target trajectory dataset. The model building module is used to build an abnormal sub-trajectory detection model based on a spatiotemporal graph convolutional network based on the target trajectory dataset; The anomaly detection module is used to detect the target trajectory using an anomaly sub-trajectory detection model and output anomaly sub-trajectory sequences.

Images