A user trajectory prediction and abnormal trajectory detection method and system

Abstract

The application discloses a user trajectory prediction and abnormal trajectory detection method and system, and the method comprises the following steps: S100, acquiring road network data, geographic information data and high-risk space-time information data; S200, based on a road network calculation engine and the road network data, preprocessing original trajectory data, including removing abnormal values, filling missing items, road network matching and period division; S300, based on the preprocessed trajectory data, identifying all landing points / station points of the user trajectory as a landing point / station point candidate set; S400, based on the road network data, the geographic information data and the landing point / station point candidate set, extracting a road network structure feature vector, a trajectory structure feature vector and a landing point / station point semantic feature vector in the trajectory data. The application can consider the prediction of user short-term trajectory trend and trajectory station point, has higher accuracy and better prediction effect.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of spatio-temporal data analysis, in particular to a user trajectory prediction and abnormal trajectory detection method and system. BACKGROUND

[0002] The existing user trajectory prediction technology is generally limited to predicting one of the user's short-term trajectory trend or trajectory anchor point, and cannot simultaneously predict both. The existing abnormal trajectory detection method mainly measures abnormality based on the deviation of local trajectory compared with normal trajectory, without considering the connection between local trajectory and anchor point or spatio-temporal risk point. Therefore, there is an urgent need for a technical solution that can simultaneously predict user short-term trajectory trend and trajectory anchor point, and can also predict trajectory abnormality. SUMMARY

[0003] In view of the defects in the prior art, the purpose of the present application is to provide a user trajectory prediction and abnormal trajectory detection method and system, which can simultaneously predict user short-term trajectory trend and trajectory anchor point, has higher prediction accuracy and better prediction effect.

[0004] To achieve the above purpose, the technical solution adopted by the present application is as follows:

[0005] A user trajectory prediction and abnormal trajectory detection method, comprising:

[0006] S100, acquiring road network data, geographic information data and high-risk spatio-temporal information data;

[0007] S200, based on a road network calculation engine and the road network data, pre-processing original trajectory data, including removing outliers, filling missing items, road network matching and period division;

[0008] S300, based on the pre-processed trajectory data, identifying all landing points / anchor points of user trajectory as a landing point / anchor point candidate set;

[0009] S400, based on the road network data, the geographic information data and the landing point / anchor point candidate set, extracting road network structure feature vectors, trajectory structure feature vectors and landing point / anchor point semantic feature vectors in the trajectory data;

[0010] S500, performing feature fusion on the road network structure feature vectors, trajectory structure feature vectors and landing point / anchor point semantic feature vectors in the trajectory data to obtain a fused current trajectory feature vector, and performing historical feature fusion on the fused current trajectory feature vector and historical trajectory feature vectors in the same period to obtain a historical feature vector;

[0011] S600, generating a corresponding trajectory sequence based on the fused current trajectory feature vector and the historical feature vector, the trajectory sequence including a predicted road segment feature vector and a predicted landing point / parking point feature vector;

[0012] S700, based on the candidate road segment set at the next time point of the current road segment and the predicted road segment feature vector, regularization is performed to obtain a predicted road segment, and based on the candidate set of landing points / parking points and the predicted landing point / parking point feature vector, regularization is performed to obtain a predicted landing point / parking point;

[0013] S800, based on the predicted road segment, the predicted landing point / parking point, the high-risk spatiotemporal information data, and the candidate route set between the current road segment and the predicted parking point / landing point, obtaining two indicators of route deviation indicators and potential risk indicators as abnormal state measurement indicators, performing state detection to obtain a detection result.

[0014] Further, the user trajectory prediction and abnormal trajectory detection method as described above, in S200, the original trajectory data is matched with the road network, including:

[0015] The trajectory data of the user is matched with the road network data by the road network calculation engine, and the road segment corresponding to each time point of the user trajectory is obtained from the road network data.

[0016] Further, the user trajectory prediction and abnormal trajectory detection method as described above, S300 includes:

[0017] After obtaining the road segment corresponding to each time point of the user trajectory from the road network data, the trajectory data in the self-defined period is integrated, and the landing point / parking point in the period is obtained by the landing point / parking point cluster as the candidate set of landing points / parking points.

[0018] Further, the user trajectory prediction and abnormal trajectory detection method as described above, in S400, the road network structure feature vector in the trajectory data is extracted, including:

[0019] The road segment obtained by matching the trajectory data with the road network is taken as the 30-minute isochronal circle as the receptive field, all road segments in the receptive field are obtained by the receptive field local filter, and the road network structure feature vector is extracted as the road network structure feature vector in the trajectory data;

[0020] The trajectory structure feature vector in the trajectory data is extracted, including:

[0021] All road segments of the user trajectory in the trajectory data in the road network data are obtained, and the graph embedding algorithm is used to perform graph embedding on each road segment by anonymous walking to obtain the trajectory structure feature vector as the trajectory structure feature vector in the trajectory data.

[0022] extracting a landing point / stay point semantic feature vector in the trajectory data, including:

[0023] obtaining information associated with the landing point / stay point in the landing point / stay point candidate set from the geographic information data, and screening attributes of the information as the landing point / stay point semantic feature vector in the trajectory data through feature engineering.

[0024] Further, the user trajectory prediction and abnormal trajectory detection method as described above, in S500, the road network structure feature vector, the trajectory structure feature vector and the landing point / stay point semantic feature vector in the trajectory data are fused, including:

[0025] The road network structure feature vector, the trajectory structure feature vector and the landing point / stay point semantic feature vector in the trajectory data are unified into the same dimension through vector embedding;

[0026] The road network structure feature vector, the trajectory structure feature vector and the landing point / stay point semantic feature vector in the trajectory data are fused through an Attention operator, a full connection operator or a summation operator.

[0027] Further, the user trajectory prediction and abnormal trajectory detection method as described above, S600 includes:

[0028] The fused current trajectory feature vector and the historical feature vector, and the motion state of the current trajectory and whether the current trajectory is at a historical landing point / stay point are input into the neural network-based seq2seq model, and the predicted road segment feature vector and the predicted landing point / stay point feature vector are output.

[0029] Further, the user trajectory prediction and abnormal trajectory detection method as described above, S700 includes:

[0030] The candidate road segments at the next time point of the current road segment are obtained by the road network calculation engine as the candidate road segment set;

[0031] The structure feature vector of the candidate road segment set is dot product operated with the predicted road segment feature vector, and the road segment with the maximum dot product value is taken as the predicted road segment;

[0032] The semantic feature vector of the landing point / stay point candidate set is dot product operated with the predicted landing point / stay point feature vector, and the landing point / stay point with the maximum dot product value is taken as the predicted landing point / stay point.

[0033] Further, in the user trajectory prediction and abnormal trajectory detection method as described above, in S800, based on the predicted road section, the predicted landing point / station point, the high-risk spatiotemporal information data, and a candidate route set between the current road section and the predicted landing point / station point, a route deviation index and a potential risk index are obtained, including:

[0034] all potential shortest road sections from the current road section to the predicted landing point / station point are obtained as the candidate route set by the road network calculation engine;

[0035] a probability of the predicted road section appearing in the candidate route set is calculated as the route deviation index;

[0036] a coverage rate of the candidate route set on the high-risk spatiotemporal information data is calculated as the potential risk index.

[0037] Further, in the user trajectory prediction and abnormal trajectory detection method as described above, in S800, a state detection is performed on the abnormal state measurement index by a detection classification model based on logistic regression.

[0038] In the embodiments of the present application, a user trajectory prediction and abnormal trajectory detection system is also provided, including:

[0039] a data acquisition module, configured to acquire road network data, geographic information data, and high-risk spatiotemporal information data;

[0040] a preprocessing module, configured to preprocess original trajectory data based on a road network calculation engine and the road network data, including abnormal value removal, missing item filling, road network matching, and period division;

[0041] a landing point / station point identification module, configured to identify all landing points / station points of user trajectories as a landing point / station point candidate set based on the preprocessed trajectory data;

[0042] a feature extraction module, configured to extract a road network structure feature vector, a trajectory structure feature vector, and a landing point / station point semantic feature vector in the trajectory data based on the road network data, the geographic information data, and the landing point / station point candidate set;

[0043] a feature fusion module, configured to perform feature fusion on the road network structure feature vector, the trajectory structure feature vector, and the landing point / station point semantic feature vector in the trajectory data to obtain a fused current trajectory feature vector, and perform history feature fusion on the fused current trajectory feature vector and a history trajectory feature vector in the same period to obtain a history feature vector;

[0044] a trajectory sequence generation module configured to generate a corresponding trajectory sequence based on the fused current trajectory feature vector and the historical feature vector, the trajectory sequence including a predicted road segment feature vector and a predicted landing point / stop point feature vector;

[0045] a prediction result regularization module configured to regularize a predicted road segment based on a candidate road segment set at a next time point of a current road segment and the predicted road segment feature vector, and regularize a predicted landing point / stop point based on a candidate landing point / stop point set and the predicted landing point / stop point feature vector;

[0046] a state detection module configured to obtain two indexes, a route deviation index and a potential risk index, as abnormal state measurement indexes based on the predicted road segment, the predicted landing point / stop point, the high-risk spatiotemporal information data, and a candidate route set between the current road segment and the predicted landing point / stop point, perform state detection, and obtain a detection result.

[0047] The method and system provided by the application can take into account the prediction of short-term trajectory trends and trajectory stop points of users, and can also analyze the abnormality of a current trajectory by combining local trajectory trends, future potential stop points, and spatiotemporal risk points defined in different scenarios, so that the prediction accuracy is higher and the prediction effect is better. BRIEF DESCRIPTION OF DRAWINGS

[0048] Figure 1 FIG. 1 is a flowchart of a user trajectory prediction and abnormal trajectory detection method provided in an embodiment of the application;

[0049] Figure 2 FIG. 2 is a structural diagram of a user trajectory prediction and abnormal trajectory detection system provided in an embodiment of the application;

[0050] Figure 3 FIG. 3 is a flow framework diagram of a user trajectory prediction and abnormal trajectory detection system provided in an embodiment of the application. DETAILED DESCRIPTION

[0051] The application will be further described in detail below in combination with the accompanying drawings and specific embodiments.

[0052] As shown in FIG. 1, a user trajectory prediction and abnormal trajectory detection method includes the following steps. Figure 1

[0053] S100, obtaining road network data, geographic information data, and high-risk spatiotemporal information data;

[0054] ​In this embodiment, the data most necessary for trajectory prediction is road network data, which can be designed in different storage modes according to different selected road network computing engines. In order to improve the prediction accuracy, the semantic characteristics of the trajectory need to be obtained, and the public POI (Point of Information) data is selected to assist in representing the semantic characteristics of the trajectory. The geographic database can be used for storage, or the graph database can be used for storage after matching with the road network. The high-risk spatiotemporal information data is data for assisting in representing the abnormality of the trajectory, which is obtained and defined according to the business requirements, such as the key places in the fire field and the large activity gathering places and key places in the public security field. The same storage mode as the POI data is used.

[0055] S200, preprocessing the original trajectory data based on the road network computing engine and the road network data, including removing abnormal values, filling missing items, road network matching and period division;

[0056] In this embodiment, the original trajectory data has a lot of missing information or abnormal values, such as large-scale deviation of positioning in a short period, missing of positioning information, etc. First, the abnormal values need to be removed, smoothed and subjected to other abnormal processing operations; then, the missing values are filled, the missing items are filled by interpolation, resampling and other methods for small-scale missing, and a suitable interpolation scheme is selected for processing according to the scene for large-scale missing; the trajectory data needs to be matched with the road network, and the trajectory points are matched with the road network by the road network computing engine combined with the road network data to obtain the road section corresponding to each trajectory; finally, the trajectory data needs to be divided into at least one fixed period, such as hour, day, week, month, etc., so as to comprehensively predict the related information in the period.

[0057] S300, identifying all landing points / stay points of the user trajectory as a landing point / stay point candidate set based on the preprocessed trajectory data;

[0058] In this embodiment, the landing point is the accommodation place of the user, and the stay point is the place where the user stays for a long time. After obtaining the road section corresponding to each time point of the trajectory, the trajectory records in the self-defined period are comprehensively obtained, and the spatiotemporal clustering algorithm is used to obtain the landing points / stay points in the period to determine the potential destination of the user's trip, thereby forming the landing point / stay point candidate set.

[0059] S400, extracting the road network structure feature vector, the trajectory structure feature vector and the landing point / stay point semantic feature vector in the trajectory data based on the road network data, the geographic information data and the landing point / stay point candidate set;

[0060] For the road network structure feature vector, the features of the road segment relative to the road segments in the receptive field need to be calculated, the receptive field is defined according to the scene (the default is a 30-minute isochronal circle), the preset road network data is obtained, and the road segment set in the corresponding receptive field is obtained from each road segment as the starting point; the topological features of the road network are extracted by inputting the road segment set obtained from the receptive field, and the feature vector of the road network is output. Different fields can select different indicators for different feature extraction methods, such as the integration degree and the selection degree in spatial syntax, the proximity and the intermediate degree in spatial design network analysis, and the embedding feature vector of each road segment in the receptive field can also be obtained by using the graph neural network.

[0061] For the trajectory structure feature vector, a part of the user trajectory data can be selected according to the business needs (the default is one week), the road segment subgraph of the user trajectory in the road network is obtained, and the frequency of jumping between the user road segments is set as the weight between the road segments. The graph embedding algorithm is used to perform anonymous walking on each road segment to obtain the trajectory structure feature. For the road segments outside the road segment subgraph, the trajectory feature can be initialized as a zero vector or a random vector according to needs.

[0062] For the landing point / stop point semantic feature vector, for each landing point / stop point, the attributes of the geographic semantic information data are selected as the semantic features of the landing point / stop point through feature engineering. The semantic features of the road segments to which the non-landing point / stop point belongs are zero vectors by default.

[0063] S500, the road network structure feature vector, the trajectory structure feature vector, and the landing point / stop point semantic feature vector in the trajectory data are fused to obtain a fused current trajectory feature vector, and the fused current trajectory feature vector and the historical trajectory feature vector in the same period are fused to obtain a historical feature vector;

[0064] In this embodiment, the three features need to be fused in the feature fusion stage, and the fusion method is to unify them into the same dimension first, and then fuse the three feature vectors by using an Attention operator, a full connection operator or a summation operator. The information of a single trajectory record cannot reflect the user's intention, and only the trajectory containing the context can reflect the user's real intention. The fused current trajectory feature vector and the historical trajectory feature vector in the same period are input into the historical feature fusion device, and a summation operator or a GRU neural network model is used to represent the historical feature vector.

[0065] S600, based on the fused current trajectory feature vector and the historical feature vector, a corresponding trajectory sequence is generated, and the trajectory sequence includes a predicted road segment feature vector and a predicted landing point / stop point feature vector;

[0066] In this embodiment, in order to comprehensively predict the short-term trend and the medium and long-term stationary point, the trajectory sequence generation stage needs to input the current trajectory motion state (whether to stay), the comprehensive features of the current trajectory (the current trajectory feature vector and the historical feature vector fused in S500), whether the current trajectory is at the historical landing point / stationary point into the seq2seq model of the neural network, and output a vector composed of two parts, one part is the predicted road segment feature vector (i.e. the road segment feature vector at the next time point), and the other part is the predicted landing point / stationary point feature vector (i.e. the landing point / stationary point feature vector at the next time point).

[0067] S700, based on the candidate road segment set at the next time point of the current road segment and the predicted road segment feature vector, regularization is performed to obtain a predicted road segment, and based on the landing point / stationary point candidate set and the predicted landing point / stationary point feature vector, regularization is performed to obtain a predicted landing point / stationary point;

[0068] In this embodiment, the candidate road segment set is composed of the candidate road segments at the next time point obtained by the road network calculation engine based on the current road segment. The structural feature vector of the candidate road segment set is multiplied by the predicted road segment feature vector output by S600, and the candidate road segment with the maximum dot product value is taken as the final predicted road segment. Similarly, the semantic feature vector of the landing point / stationary point candidate set obtained by S300 is multiplied by the predicted landing point / stationary point feature vector output by S600, and the landing point / stationary point with the maximum dot product value is taken as the final predicted landing point / stationary point.

[0069] S800, based on the predicted road segment, the predicted landing point / stationary point, the high-risk spatiotemporal information data, and the candidate route set between the current road segment and the predicted landing point / stationary point, two indexes of route deviation index and potential risk index are obtained as abnormal state measurement indexes, state detection is performed, and a detection result is obtained.

[0070] In this embodiment, all potential shortest road segments from the current road segment to the predicted landing point / stationary point output by S700 are obtained as the candidate route set by the road network calculation engine, the probability of the predicted road segment output by S700 appearing in the candidate route set is calculated as the route deviation index, and the coverage rate of the candidate route set to the high-risk spatiotemporal information data is calculated as the potential risk index. The two indexes are used as abnormal state measurement indexes, and the abnormal state measurement indexes are detected by a detection classification model based on logistic regression to detect whether the current trajectory state is normal.

[0071] As shown in FIG. Figure 2 A user trajectory prediction and abnormal trajectory detection system, comprising:

[0072] A data acquisition module 100 is configured to acquire road network data, geographic information data, and high-risk spatiotemporal information data.

[0073] The preprocessing module 200 is configured to preprocess original trajectory data based on the road network calculation engine 900 and road network data, including removing outliers, filling missing items, road network matching, and period division.

[0074] The landing point / stay point identification module 300 is configured to identify all landing points / stay points of user trajectories as a landing point / stay point candidate set based on the preprocessed trajectory data.

[0075] The feature extraction module 400 is configured to extract a road network structure feature vector, a trajectory structure feature vector, and a landing point / stay point semantic feature vector in the trajectory data based on road network data, geographic information data, and the landing point / stay point candidate set.

[0076] The feature fusion module 500 is configured to perform feature fusion on the road network structure feature vector, the trajectory structure feature vector, and the landing point / stay point semantic feature vector in the trajectory data to obtain a fused current trajectory feature vector, and perform history feature fusion on the fused current trajectory feature vector and a history trajectory feature vector in the same period to obtain a history feature vector.

[0077] The trajectory sequence generation module 600 is configured to generate a corresponding trajectory sequence based on the fused current trajectory feature vector and the history feature vector, where the trajectory sequence includes a predicted road segment feature vector and a predicted landing point / stay point feature vector.

[0078] The prediction result regularization module 700 is configured to perform regularization based on a candidate road segment set at a next time point of a current road segment and the predicted road segment feature vector to obtain a predicted road segment, and perform regularization based on the landing point / stay point candidate set and the predicted landing point / stay point feature vector to obtain a predicted landing point / stay point.

[0079] The state detection module 800 is configured to obtain two indexes, a route deviation index and a potential risk index, as abnormal state measurement indexes based on the predicted road segment, the predicted landing point / stay point, high-risk spatiotemporal information data, and a candidate route set between the current road segment and the predicted landing point / stay point, perform state detection, and obtain a detection result.

[0080] In this embodiment, as shown in FIG. 1, the system 100 includes a preprocessing module 200, a landing point / stay point identification module 300, a feature extraction module 400, a feature fusion module 500, a trajectory sequence generation module 600, a prediction result regularization module 700, and a state detection module 800. Figure 3As shown, in the preprocessing module 200, the trajectory data is first subjected to outlier processing, and then subjected to missing item interpolation; the trajectory data after interpolation will be subjected to road network matching, and the road network calculation engine 900 is called to calculate the pre-stored road network data to obtain relevant information; for the matched trajectory data, the data within a certain period of time will be input into the landing point / station point clusterer for landing point / station point identification, and for the newly appearing landing point / station point, it will be added to the candidate station point / landing point set; for the matched trajectory data, the data within a certain period of time will be subjected to trajectory structure feature extraction; through the semantic feature extractor, the semantic features will be extracted according to the information associated with the landing point / station point in the geographic information library, and for the non-landing point / station point, it will be defaulted as a zero vector or a random vector; the road segments in the road network will obtain all the road segments within their receptive fields through the receptive field local filter and perform road network feature extraction; the trajectory structure feature vector, the road network structure feature vector, and the trajectory semantic feature vector are subjected to feature fusion; the fused trajectory feature vector will be input into the historical feature fusioner together with the historical trajectory feature vector within the same period for historical feature fusion; the trajectory feature, the historical feature, the state of the current trajectory, and whether the current trajectory is in the historical landing point / station point will be input into the trajectory sequence generation module 600 for trajectory prediction, and finally the future station point / landing point feature vector and the road segment feature vector are output; the prediction result is regularized through the prediction result regularization module 700 to obtain the predicted station point / landing point and the predicted road segment; the state detection module 800 will comprehensively consider the predicted road segment, the predicted station point / landing point, and the candidate route set between the current road segment and the predicted station point / landing point, and output whether the current trajectory is in an abnormal state in combination with the self-defined high-risk spatiotemporal information library.

[0081] The user trajectory prediction and abnormal trajectory detection method and system provided by the application can consider the prediction of user short-term trajectory trend and trajectory station point, and can analyze the abnormality of the current trajectory by combining local trajectory trend, future potential station point and spatiotemporal risk points defined in different scenarios, so that the prediction accuracy is higher and the prediction effect is better.

[0082] Obviously, those skilled in the art can make various modifications and variations to the application without departing from the spirit and scope of the application. Thus, if these modifications and variations of the application fall within the scope of the claims of the application and their equivalents, the application also intends to include these modifications and variations.

Claims

1. A method for user trajectory prediction and abnormal trajectory detection, characterized in that, The method comprises the following steps: S100, acquiring road network data, geographic information data and high-risk spatiotemporal information data; S200, preprocessing original trajectory data based on a road network calculation engine and the road network data, including removing outliers, filling missing items, road network matching and period division; S300, identifying all landing points / stationary points of a user trajectory as a landing point / stationary point candidate set based on the preprocessed trajectory data; S400, extracting a road network structure feature vector, a trajectory structure feature vector and a landing point / stationary point semantic feature vector in the trajectory data based on the road network data, the geographic information data and the landing point / stationary point candidate set; S500, performing feature fusion on the road network structure feature vector, the trajectory structure feature vector and the landing point / stationary point semantic feature vector in the trajectory data to obtain a fused current trajectory feature vector, performing history feature fusion on the fused current trajectory feature vector and a history trajectory feature vector in the same period to obtain a history feature vector; S600, generating a corresponding trajectory sequence based on the fused current trajectory feature vector and the history feature vector, the trajectory sequence comprising a predicted road segment feature vector and a predicted landing point / stationary point feature vector; S700, performing regularization based on a candidate road segment set at a next time point of a current road segment and the predicted road segment feature vector to obtain a predicted road segment, and performing regularization based on the landing point / stationary point candidate set and the predicted landing point / stationary point feature vector to obtain a predicted landing point / stationary point; S800, acquiring a route deviation index and a potential risk index as abnormal state measurement indexes based on the predicted road segment, the predicted landing point / stationary point, the high-risk spatiotemporal information data, and a candidate route set between the current road segment and the predicted landing point / stationary point, performing state detection to obtain a detection result; S600 comprises: inputting the fused current trajectory feature vector and the history feature vector, a motion state of a current trajectory and whether the current trajectory is at a history landing point / stationary point into a neural network-based seq2seq model, and outputting the predicted road segment feature vector and the predicted landing point / stationary point feature vector; In S800, acquiring a route deviation index and a potential risk index based on the predicted road segment, the predicted landing point / stationary point, the high-risk spatiotemporal information data, and a candidate route set between the current road segment and the predicted landing point / stationary point, comprises: acquiring all potential shortest road segments from the current road segment to the predicted landing point / stationary point as the candidate route set through the road network calculation engine; calculating a probability of the predicted road segment appearing in the candidate route set as the route deviation index; calculating a coverage rate of the candidate route set to the high-risk spatiotemporal information data as the potential risk index. 2.The user trajectory prediction and abnormal trajectory detection method of claim 1, wherein, In S200, the road network matching of the original trajectory data comprises: matching a user trajectory in the trajectory data with the road network data through a road network calculation engine to acquire a road segment corresponding to each time point of the user trajectory from the road network data. 3.The user trajectory prediction and abnormal trajectory detection method of claim 2, wherein, S300 comprises: After obtaining the road segments corresponding to each time point of the user trajectory from the road network data, the trajectory data in a self-defined period is integrated, and the anchor points / parking points in the period are obtained by an anchor point / parking point clusterer as an anchor point / parking point candidate set. 4.The user trajectory prediction and abnormal trajectory detection method of claim 2, wherein, In S400, a road network structure feature vector in the trajectory data is extracted, including: Taking the road segment obtained by road network matching of the trajectory data as a starting point, a 30-minute isochronal circle of walking as a receptive field, and all road segments in the receptive field being obtained by a receptive field local filter, a road network structure feature vector is extracted as the road network structure feature vector in the trajectory data; A trajectory structure feature vector in the trajectory data is extracted, including: All road segments of the user trajectory in the trajectory data in the road network data are obtained, and a graph embedding algorithm is used to perform graph embedding on each road segment in an anonymous walk manner, so as to obtain a trajectory structure feature vector as the trajectory structure feature vector in the trajectory data; An anchor point / parking point semantic feature vector in the trajectory data is extracted, including: Information associated with the anchor points / parking points in the anchor point / parking point candidate set is obtained from the geographic information data, and an attribute of the information is selected by feature engineering as the anchor point / parking point semantic feature vector in the trajectory data. 5.The user trajectory prediction and abnormal trajectory detection method of claim 1, wherein, In S500, the road network structure feature vector, the trajectory structure feature vector, and the anchor point / parking point semantic feature vector in the trajectory data are fused, including: The road network structure feature vector, the trajectory structure feature vector, and the anchor point / parking point semantic feature vector in the trajectory data are unified into the same dimension by vector embedding; The road network structure feature vector, the trajectory structure feature vector, and the anchor point / parking point semantic feature vector in the trajectory data are fused by an Attention operator, a fully connected operator, or a summation operator. 6.The user trajectory prediction and abnormal trajectory detection method of claim 1, wherein, S700 includes: A candidate road segment set is obtained by the road network computing engine as the candidate road segment set; A dot product operation is performed between the structure feature vector of the candidate road segment set and the predicted road segment feature vector, and a road segment with the maximum dot product value is taken as the predicted road segment; A dot product operation is performed between the semantic feature vector of the anchor point / parking point candidate set and the predicted anchor point / parking point feature vector, and an anchor point / parking point with the maximum dot product value is taken as the predicted anchor point / parking point. 7.The user trajectory prediction and abnormal trajectory detection method of claim 1, wherein, In S800, a state of the abnormal state measurement index is detected by a detection classification model based on logistic regression.

8. A user trajectory prediction and abnormal trajectory detection system, characterized in that, It includes: A data acquisition module is configured to acquire road network data, geographic information data, and high-risk spatiotemporal information data; A preprocessing module is configured to preprocess original trajectory data based on a road network computing engine and the road network data, including removing outliers, filling missing items, road network matching, and period division; An anchor point / parking point identification module is configured to identify all anchor points / parking points of a user trajectory based on the preprocessed trajectory data as an anchor point / parking point candidate set; a feature extraction module configured to extract a road network structure feature vector, a trajectory structure feature vector, and a landing point / stay point semantic feature vector in the trajectory data based on the road network data, the geographic information data, and the landing point / stay point candidate set; a feature fusion module configured to perform feature fusion on the road network structure feature vector, the trajectory structure feature vector, and the landing point / stay point semantic feature vector in the trajectory data to obtain a fused current trajectory feature vector, perform history feature fusion on the fused current trajectory feature vector and a history trajectory feature vector in the same period to obtain a history feature vector; a trajectory sequence generation module configured to generate a corresponding trajectory sequence based on the fused current trajectory feature vector and the history feature vector, the trajectory sequence including a predicted road segment feature vector and a predicted landing point / stay point feature vector; a prediction result regularization module configured to perform regularization on a candidate road segment set at a next time point of a current road segment and the predicted road segment feature vector to obtain a predicted road segment, and perform regularization on the landing point / stay point candidate set and the predicted landing point / stay point feature vector to obtain a predicted landing point / stay point; a state detection module configured to obtain two indexes, a route deviation index and a potential risk index, as abnormal state measurement indexes based on the predicted road segment, the predicted landing point / stay point, the high-risk spatiotemporal information data, and a candidate route set between the current road segment and the predicted landing point / stay point, perform state detection, and obtain a detection result.

Images