A vehicle trajectory prediction method, electronic device, chip system and storage medium

Abstract

The application provides a vehicle trajectory prediction method, an electronic device, a chip system and a storage medium, relates to the Internet of Vehicles technical field, and can receive SRS signals periodically transmitted by a target vehicle through a base station, obtain motion parameters of the target vehicle and surrounding mobile scatterers through the SRS signals, match the motion parameters of the current period and historical trajectory information, obtain current trajectory information, and the current trajectory information can be used as historical trajectory information of the next period; through the current trajectory information and high-precision map data, the motion characteristics, interaction characteristics and map characteristics of each node are obtained, and each node includes the target vehicle and the surrounding mobile scatterers; the trajectory information of the target vehicle and the mobile scatterers is predicted through the three types of characteristics, and the accuracy of vehicle trajectory prediction can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of Internet of Vehicles, and in particular to a vehicle trajectory prediction method, an electronic device, a chip system and a storage medium. BACKGROUND

[0002] The vehicle trajectory prediction based on multiple targets can include two parts: obtaining trajectory information of multiple targets by associating and tracking multiple targets (including target vehicles) observed at different times; and performing subsequent trajectory prediction based on the trajectory information of the multiple targets. Therefore, the trajectory information of the multiple targets is crucial as the data basis for vehicle trajectory prediction.

[0003] Currently, the position, speed, and motion direction of multiple targets can be obtained by GPS positioning technology, and the trajectory information of the multiple targets can be obtained based on the motion parameters of the multiple targets at multiple time points, so as to perform trajectory prediction. However, the accuracy of GPS positioning technology is limited, resulting in low accuracy and precision of the trajectory information of the multiple targets, and thus the accuracy and precision of the trajectory prediction are also low. SUMMARY

[0004] The present application provides a vehicle trajectory prediction method, an electronic device, a chip system and a storage medium, which can accurately predict the trajectory information of a vehicle.

[0005] To achieve the above-mentioned purpose, the first aspect of the present application provides a vehicle trajectory prediction method, which adopts the following technical solution:

[0006] Obtaining SRS signals periodically transmitted by a target vehicle and received by a base station;

[0007] Obtaining motion parameters of each node based on the SRS signals, wherein the each node includes the target vehicle and scatterers located around the target vehicle, and the motion parameters include position, speed, and direction;

[0008] Obtaining current trajectory information including the motion parameters of each node in the current period based on the motion parameters of each node in the current period and historical trajectory information, wherein the current trajectory information is used as historical trajectory information in the next period;

[0009] Obtaining motion features, interaction features, and map features of each node based on the current trajectory information and map data;

[0010] Obtaining predicted trajectory information of each node in at least one subsequent period based on the motion features, interaction features, and map features of each node.

[0011] In the present application, the motion parameters of each node can be obtained through the SRS signals received by a single base station, the current trajectory information of each node can be obtained according to the motion parameters of each node and historical trajectory information, the motion characteristics of each node, the interaction characteristics between each node and the map characteristics can be obtained according to the current trajectory information and high-precision map data, and finally the prediction trajectory information of each node in the subsequent period can be predicted according to the characteristics of multiple angles; the present application can be applied in a single base station scene, so it can be applied in occasions where the base stations are sparse, and the application range is wider; the motion parameters of each node can be obtained according to the SRS information collected by the base station, and the accuracy of the motion data of each node is improved; in combination with high-precision map data, the map characteristics related to the motion parameters of the node can also be obtained, and the map characteristics can be used as parameters of the prediction trajectory information, so that the prediction can be more accurate and accurate; when predicting the trajectory information, not only the motion parameters of each node are considered, but also the interaction characteristics between each node are considered, so the accuracy and accuracy of the predicted trajectory information are higher.

[0012] As an implementation manner of the first aspect, the motion parameters of each node obtained according to the SRS signals comprise:

[0013] Separating the LOS path signals and the NLOS path signals of the SRS signals to obtain the first angle of arrival, the first time delay and the first Doppler frequency shift of the LOS path signals, and the second angle of arrival, the second time delay and the second Doppler frequency shift of each NLOS path signal;

[0014] Obtaining the first coordinate of the target vehicle according to the coordinates of the base station receiving the SRS signals, the first propagation distance of the LOS path signals and the first angle of arrival, wherein the first propagation distance of the LOS path signals is determined by the first time delay;

[0015] Obtaining the first driving direction of the target vehicle according to the first coordinate of the target vehicle and the road direction in the map data;

[0016] Obtaining the first driving speed of the target vehicle according to the first radial speed of the target vehicle, the first angle of arrival and the first driving angle of the target vehicle, wherein the first radial speed is the speed of the target vehicle in the direction towards the base station, the first radial speed is obtained according to the first Doppler frequency shift, and the first driving angle of the target vehicle is determined by the first driving direction.

[0017] In the present application, the correlation parameters of the LOS path signal directly reaching the base station from the target vehicle and the correlation parameters of the NLOS path signal reaching the base station through a scatterer can be obtained by separating the SRS signal; the coordinates of the target vehicle are determined according to the correlation parameters of the LOS path signal, and after the coordinates of the target vehicle are determined, the position of the target vehicle in the map can be determined, the roads in the map have a predetermined direction, so the driving direction of the target vehicle can be determined according to the position of the target vehicle in the map; after the driving direction of the target vehicle is determined, the driving speed of the target vehicle can be further calculated. By separating the LOS path signal in the manner of separating the SRS signal, and combining the map data to determine the driving direction and then calculating the driving speed, more accurate motion parameters of the target vehicle can be obtained.

[0018] As an implementation form of the first aspect, the first driving speed of the target vehicle is obtained according to the first radial speed of the target vehicle, the first angle of arrival and the first driving angle of the target vehicle, and includes:

[0019] ;

[0020] wherein, the first driving speed of the target vehicle, denotes the speed of light, the first radial speed of the target vehicle, the carrier frequency, the first driving angle of the target vehicle; the first angle of arrival of the target vehicle.

[0021] In the present application, the driving speed of the target vehicle can be calculated according to the mutual positions among the target vehicle, the base station and the antenna direction (horizontal direction in the coordinate system), and the radial speed obtained according to the Doppler shift.

[0022] As an implementation form of the first aspect, the motion parameters of each node are obtained according to the SRS signal, and include:

[0023] The second coordinates of the pth scatterer are obtained according to the coordinates of the base station receiving the SRS signal, the second propagation distance of the pth NLOS path signal and the second angle of arrival of the pth NLOS path signal, wherein the second propagation distance of the pth NLOS path signal is determined by the second time delay of the pth NLOS path signal, and the maximum value of p is related to the number of NLOS path signals;

[0024] The second driving direction of the pth scatterer is obtained according to the second coordinates of the pth scatterer and the road direction in the map data;

[0025] The second Doppler shift of the pth NLOS path signal, the second angle of arrival, and the second angle of travel of the target vehicle are used to obtain the second travel speed of the pth scatterer, and the second angle of travel is determined by the second travel direction.

[0026] In the present application, since the scatterer reflects the SRS signal of the target vehicle, the calculation process of the travel speed of the scatterer cannot only consider the position and travel direction of the scatterer, but also needs to consider the position, travel speed and travel direction of the target vehicle; the travel speed of the scatterer can be obtained in combination with the motion parameters of the target vehicle.

[0027] As an implementation form of the first aspect, the second Doppler shift of the pth NLOS path signal includes:

[0028] A third Doppler shift generated by the pth scatterer, and the third Doppler shift is related to the first travel speed, the first angle of travel of the target vehicle, and the angle between the target vehicle and the pth scatterer;

[0029] A fourth Doppler shift generated when the pth scatterer reflects the signal, and the fourth Doppler shift is related to the third Doppler shift, the second travel speed of the pth scatterer, the second angle of travel, and the angle between the target vehicle and the pth scatterer;

[0030] And a fifth Doppler shift generated when the base station receives the SRS signal, and the fifth Doppler shift is related to the fourth Doppler shift, the first travel speed of the target vehicle, the second angle of travel and the second angle of arrival of the pth scatterer.

[0031] In the present application, the NLOS path signal is information scattered by the scatterer, so the Doppler shift corresponding to the NLOS path signal includes the Doppler shift generated by the scatterer, the Doppler shift generated when the signal is reflected, and the Doppler shift generated when the base station receives the signal.

[0032] As an implementation form of the first aspect, the second Doppler shift of the pth NLOS path signal, the second angle of arrival, and the second angle of travel of the target vehicle are used to obtain the second travel speed of the pth scatterer, and the second angle of travel is determined by the second travel direction.

[0033] ;

[0034] ;

[0035] ;

[0036] ;

[0037] wherein, a second Doppler shift of the pth scatterer, a third Doppler shift, a fourth Doppler shift, a fifth Doppler shift, a carrier frequency, a light speed, a first driving speed of the target vehicle calculated, a first driving angle of the target vehicle, an angle between the pth scatterer and the target vehicle; a second driving speed of the pth scatterer calculated, a second driving angle of the pth scatterer, a second angle of arrival of the pth NLOS path signal.

[0038] In the present application, the driving speed of a scatterer can be obtained according to multiple Doppler shifts and motion parameters of a target vehicle.

[0039] As an implementation form of the first aspect, the obtaining, according to motion parameters of each node in a current period and historical trajectory information, of current trajectory information containing the motion parameters of each node in the current period comprises:

[0040] generating a motion feature graph according to the motion parameters of each node in the current period, wherein a node in the motion feature graph is a vertex i;

[0041] generating a trajectory feature graph according to the historical trajectory information, wherein a node in the trajectory feature graph is a vertex j;

[0042] recursively enhancing an initial vertex feature of the vertex i in the motion feature graph to obtain an enhanced vertex feature of the vertex i, wherein the initial vertex feature of the vertex i is the motion parameter of the vertex i in the current period;

[0043] recursively enhancing an initial vertex feature of the vertex j in the trajectory feature graph to obtain an enhanced vertex feature of the vertex j, wherein the initial vertex feature of the vertex j is a historical motion parameter of the vertex j;

[0044] after obtaining the enhanced vertex feature of each vertex i in the motion feature graph, obtaining an edge feature in the motion feature graph according to the vertex features of any two vertices in the motion feature graph;

[0045] after obtaining the enhanced vertex feature of each vertex j in the trajectory feature graph, obtaining an edge feature in the trajectory feature graph according to the vertex features of any two vertices in the trajectory feature graph;

[0046] According to the enhanced vertex features and edge features in the motion feature graph and the enhanced vertex features and edge features in the trajectory feature graph, the motion feature graph and the trajectory feature graph are matched to obtain the current trajectory information.

[0047] In the present application, feature graphs are generated from motion parameters and historical trajectory information, thereby converting into a graph matching problem. By solving the graph matching problem, each node in the motion parameters and the historical trajectory information is corresponded, so that the current trajectory information containing the current motion parameters can be obtained according to the historical trajectory information and the current motion parameters. When performing graph matching, the features can be enhanced multiple times, so that the features of each node are highlighted, and the graph matching is more accurate. When performing graph matching, not only the vertex features are considered, but also the edge features between vertices are considered. The graph matching is performed through the vertex features and the edge features, thereby improving the accuracy of node matching.

[0048] As another implementation of the first aspect, the cyclically enhancing the initial vertex feature of the vertex i in the motion feature graph to obtain the enhanced vertex feature of the vertex i comprises:

[0049] In each cyclic enhancement process of the vertex feature of the vertex i in the motion feature graph, the enhanced vertex feature of the vertex i is:

[0050] ;

[0051] wherein, denotes the enhanced vertex feature of the vertex i obtained in the last cycle, denotes the enhanced vertex feature of the vertex i obtained in the current cycle, and MLP denotes a multi-layer perception machine, and the aggregated neighborhood feature of the vertex i denotes the geometric feature cosine similarity between the vertex i and the vertex j. .

[0052] In the present application, by taking the similarity between the current vertex and the adjacent vertex as the enhanced feature, the accuracy of graph matching can be further improved.

[0053] As an implementation of the first aspect, the current trajectory information comprises a motion sequence set of multiple nodes, and the matching of the motion feature graph and the trajectory feature graph according to the enhanced vertex features and edge features in the motion feature graph and the enhanced vertex features and edge features in the trajectory feature graph to obtain the current trajectory information comprises:

[0054] calculating vertex similarities between the vertex features in the motion feature graph and the vertex features in the trajectory feature graph, and edge similarities between the edge features in the motion feature graph and the edge features in the trajectory feature graph;

[0055] maximizing the vertex similarity and the edge similarity to obtain a matching score of a vertex in the motion feature graph and a vertex in the trajectory feature graph, the vertex in the motion feature graph being a motion parameter of a node, the vertex in the trajectory feature graph being a motion sequence of each node, wherein a first vertex in the motion feature graph is matched with a motion sequence having the highest matching score with the first vertex, the first vertex being any vertex in the motion feature graph;

[0056] if the first vertex in the motion feature graph does not exist a matched state sequence, the first vertex is a newly appeared node, and a state sequence of the first vertex is generated;

[0057] if the first vertex in the motion feature graph exists a matched state sequence, and a Mahalanobis distance between a position of the first vertex and a predicted position of the matched state sequence is greater than a distance threshold, the first vertex is deleted;

[0058] if the first vertex in the motion feature graph exists a matched state sequence, and a Mahalanobis distance between a position of the first vertex and a predicted position of the matched state sequence is less than or equal to the distance threshold, a motion parameter of the first vertex is added to the matched state sequence;

[0059] if there are a plurality of first state sequences in the trajectory feature graph, each of which does not add a motion parameter of a first vertex in a plurality of periods, the first state sequences are deleted.

[0060] In the present application, after matching each node in the motion feature graph and each node in the trajectory feature graph, the node j in the trajectory feature graph with the highest matching score with the node i in the motion feature graph can be taken as the state sequence matched with the node i. If the vertex i does not have a matched state sequence, it means that the vertex i has no historical trajectory information and is a node newly appearing in the current period. Therefore, the state sequence of the vertex is generated. If the vertex i has a matched state sequence, and the Mahalanobis distance between the position of the vertex i and the predicted position of the matched state sequence j is greater than the distance threshold, it means that the vertex i deviates from the motion trajectory of j and may not be the same node as j. The motion parameters of the vertex i are deleted. If the vertex i has a matched state sequence, and the Mahalanobis distance between the position of the vertex i and the predicted position of the matched state sequence j is less than or equal to the distance threshold, it means that the vertex i is within the motion trajectory of j and is the same node as j. The motion parameters of the vertex i are added to the matched state sequence j, and the current trajectory information of the vertex is generated. Of course, if the motion parameters of a state sequence are not updated for several periods, it means that the node corresponding to the state sequence has disappeared from the current environment, and the current state sequence can be deleted. In this way, the nodes newly appearing in the current environment can be added to the current trajectory information, the state sequences of the nodes disappearing in the current environment can be deleted, and the motion parameters of the nodes still in the current environment can be connected to the historical trajectory information to generate the current trajectory information.

[0061] As an implementation form of the first aspect, the maximum of the vertex similarity and the edge similarity is used to obtain the matching score of the vertex in the motion feature graph and the vertex in the trajectory feature graph, including:

[0062] A quadratic programming model related to the vertex similarity and the edge similarity is trained to obtain a trained quadratic programming model;

[0063] After the training is completed, the matching score of the vertex in the motion feature graph and the vertex in the trajectory feature graph is obtained;

[0064] The quadratic programming model is:

[0065] ;

[0066] The solution of the quadratic programming model is:

[0067] , ;

[0068] wherein d is the dimension of the feature vector of the edge weight, , is a weighted adjacency tensor of the motion feature graph, , is a weighted adjacency tensor of the motion feature graph, B denotes vertex similarity, B is a vertex similarity matrix, and elements of B are , X denotes constraints after a double random matrix, M denotes edge similarity, M is a symmetric quadratic similarity matrix of edge features; , , is a cosine similarity matrix of edge features, ; , , , ; denotes the motion parameters of the vertices in the motion feature map, denotes the relationship between the vertices in the motion feature map, denotes the state sequence of the vertices in the trajectory feature map, denotes the relationship between the vertices in the trajectory feature map;

[0069] The loss function used to train the quadratic programming model is:

[0070] ;

[0071] wherein, ; is the predicted matching score between the vertices in the motion feature map and the vertices in the trajectory feature map, is the true label value, is a balance factor, is a temperature parameter, is the last matching score.

[0072] As an implementation manner of the first aspect, the motion feature, the interaction feature, and the map feature of each node are obtained according to the current trajectory information and the map data, including:

[0073] The motion feature is: , ;

[0074] wherein, denotes a history encoder shared by all target nodes of type K, K denotes the type of node i, K=1 denotes a high-speed moving type, and K=2 denotes a low-speed moving type, denotes the dynamic feature of node i, denotes the state sequence in the current trajectory information.

[0075] In the present application, nodes are divided into high-speed and low-speed types according to changes in driving speed or position of the nodes, different types of nodes use different encoders, and the accuracy of the motion features of the nodes is improved in this way.

[0076] As another implementation of the first aspect, the motion feature, the interaction feature and the map feature of each node are obtained according to the current trajectory information and the map data, including:

[0077] A directed edge heterogeneous feature graph is constructed by the motion feature and the directed edge feature of each node, and the directed edge feature includes an edge weight and an edge type.

[0078] The initial node feature of each node in the directed edge heterogeneous feature graph is extracted.

[0079] The type of the initial node feature of node i is projected into a feature space shared by each type to obtain the type-converted initial node feature of node i.

[0080] The edge weight and the edge type of node i are converted respectively to obtain the converted edge feature of node i, and the converted edge feature of node i includes a converted edge weight and a converted edge type.

[0081] The converted edge feature of node i and the converted initial node feature of node i are spliced into a feature vector of node i.

[0082] The attention coefficient of node i is obtained according to the feature vector of node i and the converted initial node feature.

[0083] The feature vectors of each node in the neighborhood of node i are combined by taking the attention coefficient of node i as a weight to obtain a neighborhood feature vector of node i.

[0084] The neighborhood feature vector of node i is updated by using a sigmoid function to obtain a target node feature of node i.

[0085] The target node features of each node are combined to obtain the interaction feature.

[0086] In the present application, when calculating the interaction feature of a node, the motion feature of each node alone needs to be considered, and the edge feature between nodes also needs to be considered to reflect the interaction characteristics. Therefore, a directed edge heterogeneous feature graph is constructed to improve the accuracy and robustness. When obtaining the interaction feature, the initial node feature is converted according to the type of the node, and the edge weight and the edge type of the node are both considered as parameters of the interaction feature, so that the obtained interaction feature is more comprehensive and can reflect the mutual influence between each node.

[0087] As an implementation manner of the first aspect, the attention coefficient of the node i is obtained according to the feature vector of the node i and the converted initial node feature, and the obtaining includes:

[0088]

[0089] is an angle of a control negative slope; is the feature vector of the node i, is the converted initial node feature of the node i; represents a single-layer feedforward network.

[0090] As another implementation manner of the first aspect, the target node feature of the node i is obtained by updating the neighborhood feature vector of the node i using a sigmoid function, and the updating includes:

[0091]

[0092] is a sigmoid function, represents the neighborhood feature vector of the node i, and contains the converted edge weight and the converted initial node feature, and does not contain the converted edge type.

[0093] As an implementation manner of the first aspect, the motion feature, the interaction feature and the map feature of each node are obtained according to the current trajectory information and the map data, and the obtaining includes:

[0094]

[0095] represents the map feature of the node i, represents a selection gate, and the value is between 0 and 1, represents a Hadamard product, is a feature of the map data extracted using a CNN, is a sigmoid function, is a weight matrix, represents a combination of a position vector and a velocity vector of the node i at the t moment, is an offset.

[0096] In the present application, when the map feature is extracted, all nodes share the same map data, which saves storage space, and the motion parameters of the nodes are also considered, so that the extracted map feature and the current motion state are more matched, and the finally predicted trajectory information is more accurate.

[0097] ​​​​​​​​In a second aspect, an electronic device is provided, including a processor configured to invoke a computer program stored in a memory to implement the steps of the method of any one of the first aspects of this application.

[0098] Thirdly, a chip system is provided, including a processor coupled to a memory, the processor executing a computer program stored in the memory to cause an electronic device to implement the steps of the method of any one of the first aspects of this application.

[0099] Fourthly, a computer-readable storage medium is provided, which stores a computer program that, when executed on an electronic device, causes the electronic device to implement the steps of the method of any one of the first aspects of this application.

[0100] Fifthly, a computer program product is provided that, when run on a device, causes an electronic device to perform the steps of the method of any one of the first aspects of this application.

[0101] It is understood that the beneficial effects of the second to fifth aspects mentioned above can be found in the relevant descriptions in the first aspect mentioned above, and will not be repeated here. Attached Figure Description

[0102] Figure 1 A schematic diagram of the hardware structure of an electronic device provided in an embodiment of this application;

[0103] Figure 2 This is a schematic diagram illustrating an application scenario of a vehicle trajectory prediction method provided in an embodiment of this application.

[0104] Figure 3 This application provides a schematic diagram of an information interaction process as an embodiment of the present application.

[0105] Figure 4 A flowchart illustrating a vehicle trajectory prediction method provided in an embodiment of this application;

[0106] Figure 5 Provided for the embodiments of this application Figure 4 A schematic diagram of step (1) in the process of calculating the motion parameters of the target vehicle;

[0107] Figure 6 Provided for the embodiments of this application Figure 4 A schematic diagram of step (1) in the process of calculating the motion parameters of the scatterer;

[0108] Figure 7 A schematic diagram illustrating the positional relationship between a base station, a target vehicle, and a scattering object, provided in an embodiment of this application;

[0109] Figure 8 Provided for the embodiments of this applicationFigure 4 The process diagram of step (2) in FIG. 3 is shown in FIG. 4.

[0110] Figure 9 The process diagram of step (2) in FIG. 3 is shown in FIG. 4. Figure 8 The process diagram of step (2) in FIG. 3 is shown in FIG. 4.

[0111] Figure 10 The process diagram of step (2) in FIG. 3 is shown in FIG. 4. Figure 8 The process diagram of step (2) in FIG. 3 is shown in FIG. 4.

[0112] Figure 11 The process diagram of step (2) in FIG. 3 is shown in FIG. 4. DETAILED DESCRIPTION

[0113] In the following description, for the purpose of explanation and not limitation, specific details are set forth, such as particular system configurations, techniques, etc., in order to provide a thorough understanding of the embodiments of the present application. However, it will be apparent to those skilled in the art that the present application can be practiced in other embodiments that depart from these specific details.

[0114] It should be understood that, when used in the specification and the appended claims of the present application, the term "comprising" indicates the presence of the described features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0115] It should also be understood that, in the embodiments of the present application, "one or more" means one, two, or more than two; "and / or" describes the association relationship of the associated objects, which means that there can be three relationships; for example, A and / or B can mean that A exists alone, A and B exist together, and B exists alone, where A and B can be singular or plural. The character " / " generally represents an "or" relationship between the associated objects.

[0116] In addition, in the description of the present application and the appended claims, the terms "first", "second", "third", "fourth", etc. are only used for differentiation and cannot be understood as indicating or implying relative importance.

[0117] Reference within the specification to "one embodiment" or "an embodiment" or "some embodiments" means that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment of the application. The appearances of the phrase "in one embodiment" or "in some embodiments" in various places within specified

[0118] Figure 1 A structural schematic diagram of an electronic device for performing a vehicle trajectory prediction method provided by the application is shown in FIG. 1. The electronic device includes a processor, a communications interface, a memory, and a bus. The processor, the communications interface, and the memory communicate with each other via the bus. Figure 1

[0119] The communications interface is configured to receive and output information.

[0120] The processor is configured to execute a computer program.

[0121] The memory is configured to store the computer program.

[0122] Specifically, the computer program can include program code, and the program code can include operation instructions of the electronic device.

[0123] The processor can be a central processing unit (CPU) or an application specific integrated circuit (ASIC) or one or more integrated circuits configured to implement embodiments of the application.

[0124] The memory is configured to store the computer program. The memory can include a random access memory (RAM) and can also include a non-volatile memory such as at least one disk memory.

[0125] The electronic device can execute the computer program stored in the memory by using the processor, thereby performing the vehicle trajectory prediction method provided by embodiments of the application. ​

[0126] The electronic device can be a server or a base station. When the electronic device is a base station, the communication interface can be an antenna.

[0127] As an example, a base station receives a Sounding Reference Signal (SRS) through an antenna, and a processor in the base station executes a computer program stored in a memory, so as to obtain trajectory information of a target vehicle according to the received SRS signal.

[0128] As another example, a server receives a SRS signal collected by a base station through a communication interface, and a processor in the server executes a computer program stored in a memory, so as to obtain trajectory information of a target vehicle according to the received SRS signal.

[0129] Detailed contents can be referred to the description of subsequent embodiments. The embodiments of the present application do not particularly limit the specific structure of the execution subject of the vehicle trajectory prediction method, as long as the code of the vehicle trajectory prediction method recorded in the embodiments of the present application can be run to communicate according to the vehicle trajectory prediction method provided by the embodiments of the present application. For example, the execution subject of the vehicle trajectory prediction method provided by the embodiments of the present application can be a functional module capable of calling and executing programs in an electronic device, or a communication device applied in an electronic device, for example, a chip.

[0130] 5G-V2X is a technology that uses 5G network to realize Vehicle-to-Everything (V2X) communication. Other entities can be vehicles, infrastructure, pedestrians, etc. 5G-V2X can be applied in the field of Internet of Vehicles, such as intelligent driving, intelligent transportation, etc.

[0131] V2X messages are transmitted between vehicles and other entities through a channel. The channel is a channel for signal transmission in a communication system, and is a transmission medium through which the signal is transmitted from the transmitting end to the receiving end. The transmission condition of the signal (for example, success, delay, failure, etc.) is closely related to the state of the channel (for example, payload, interference signal, etc.). Therefore, the base station needs to monitor the channel state.

[0132] When a vehicle is driving on a road, the position of the vehicle is changing in real time, and the state of the channel used by the vehicle is also changing in real time. Moreover, when the vehicle speed is high, the state of the channel used by the vehicle also changes relatively fast. Therefore, the base station needs to monitor the channel state at a high frequency, which results in a large overhead of the base station and complex beam management.

[0133] If a relatively accurate vehicle trajectory prediction method is used to predict the motion state (for example, position, driving speed and driving direction, etc.) of the vehicle, the base station can adjust the beam direction in advance, reduce the frequency of monitoring the channel state, and thereby save the overhead.

[0134] Therefore, the vehicle trajectory prediction method provided in the embodiments of the present application can be applied in a single base station scenario, and in combination with high-precision map data and deep learning technology, the motion state of the target vehicle can be predicted more accurately.

[0135] Of course, if the vehicle trajectory prediction method provided in the embodiments of the present application is used, the base station can trigger the adjustment of the beam direction, the reasonable allocation of the frequency spectrum resources, and the frequency of monitoring the channel state in advance according to the predicted motion state of the target vehicle, thereby reducing the overhead of the base station.

[0136] In a wireless communication system (for example, 5G-New Radio, which can also be referred to as 5G-NR), if the channel needs to be understood, a signal can be sent to probe the channel first. For example, the signal for probing the uplink channel state of the user equipment to the base station is called a sounding reference signal (SRS), and the signal for probing the downlink channel state of the base station to the user equipment is called a channel state information reference signal (CSI-RS).

[0137] In a frequency division duplex (FDD) system, the SRS signal can only measure the uplink channel state. In a time division duplex (TDD) system, since the uplink and downlink channels have reciprocity, the downlink channel state can be inferred from the uplink channel, so the SRS signal can also be used to measure the downlink channel state in the TDD system.

[0138] In actual applications, the ways in which the user equipment sends the SRS signal include periodic, semi-static, and aperiodic. The user equipment selects the way in which the SRS signal is sent according to the received resource configuration information. In the embodiments of the present application, the base station can send the resource configuration information to the user equipment: periodic sending mode, so the user equipment can periodically send the SRS signal.

[0139] Referring to FIG. 1, Figure 2 As an application scenario of an embodiment of the present application, the vehicles driving on the road include vehicle A and vehicle B, and there can also be pedestrians and the like on the road. Of course, there are also some fixed-position objects, such as large trees, beside the road.

[0140] When one of the vehicles (e.g., vehicle A) is targeted, the trajectory information of the vehicle A is predicted, the vehicle A is the target vehicle, the vehicle B, the pedestrian and the tree are scatterers in the environment of the target vehicle, and the vehicle A can periodically send SRS signals during the driving of the vehicle A. The SRS signals sent in one period can reach the base station after being scattered by multiple scatterers, and the base station receives the SRS signals by using a multiple-input, multiple-output (MIMO) antenna array.

[0141] The SRS signals are used to determine the motion parameters of the target vehicle and the scatterers. The motion parameters of the target vehicle and the scatterers in multiple continuous periods (the motion parameters in multiple continuous periods form the trajectory information) can be determined by the SRS signals in multiple continuous periods. The predicted trajectory information of the target vehicle can be obtained by the historical trajectory information.

[0142] Referring to (a) in FIG. 1, Figure 3 After the base station receives the SRS signals, the base station calculates the predicted trajectory information of the target vehicle according to the SRS signals. Correspondingly, after the base station calculates the predicted trajectory information of the target vehicle, the base station can adjust the beam direction, reasonably allocate the spectrum resources, and keep the current relatively low frequency of monitoring the channel state according to the predicted trajectory information of the target vehicle.

[0143] Referring to (b) in FIG. 1, Figure 3 After the base station receives the SRS signals, the base station can send the SRS signals to the server. The server calculates the predicted trajectory information of the target vehicle according to the SRS signals. The server sends the calculated predicted trajectory information of the target vehicle to the base station. The base station can adjust the beam direction, reasonably allocate the spectrum resources, and keep the current relatively low frequency of monitoring the channel state according to the predicted trajectory information of the target vehicle. Of course, in actual applications, the base station can first process the SRS signals, send the processed data to the server, and then send the predicted trajectory information of the target vehicle to the base station after the server performs the subsequent steps.

[0144] In the embodiments of the present application, the process of obtaining the predicted trajectory information of the target vehicle according to the SRS signals includes multiple steps. In actual applications, the base station can determine whether to perform, perform part of or perform all of the steps according to the execution capability of the base station itself. Of course, if the base station does not perform or performs part of the steps, the server needs to participate in the execution of other steps.

[0145] The subsequent embodiments of the present application describe the specific steps in the process of obtaining the predicted trajectory information of the target vehicle according to the SRS signals. In the description process, the execution subject of each step is not limited.

[0146] Referring to FIG. 1, Figure 4The main steps of the vehicle trajectory prediction method provided by the embodiments of the present application are shown in the schematic diagram.

[0147] (1) Extracting motion parameters

[0148] The motion parameters of each node, including the target vehicle and each scatterer located around the target vehicle, are obtained according to the SRS signals periodically transmitted by the target vehicle and received by the base station, and the motion parameters include position, speed and direction.

[0149] In this step, the motion parameters of the target vehicle and the motion parameters of the scatterers need to be calculated respectively.

[0150] (2) Obtain current trajectory information through graph matching

[0151] According to the motion parameters of each node in the current period and the historical trajectory information, the current trajectory information containing the motion parameters of each node in the current period is obtained, and the current trajectory information is used as the historical trajectory information of the next period.

[0152] In this step, the motion feature map of each node in the current period T is generated, the trajectory information of each node determined in the last period (T-1) is used as the historical trajectory information of the current period T, the trajectory feature map is generated, and the motion feature map and the trajectory feature map are matched through graph matching to obtain the current trajectory information containing the motion parameters of each node in the current period T, and the current trajectory information is used as the historical trajectory information of the next period.

[0153] (3) Extract motion features, interaction features and map features

[0154] According to the current trajectory information (motion sequence set) and map data, the motion features, interaction features and map features of each node are obtained.

[0155] In this step, a single module is used to respectively perform the following steps: extracting the motion features of each node from the trajectory information; obtaining the interaction features according to the trajectory information and the extracted motion features of each node; and obtaining the map features according to the trajectory information and the map data.

[0156] (4) Jointly predict trajectory information by multiple features

[0157] According to the motion features, interaction features and map features of each node, the predicted trajectory information of each node in at least one subsequent period is obtained. In this step, the three features are jointly used to predict the predicted trajectory information of each node in at least one subsequent period.

[0158] In the embodiments of the present application, the motion parameters of the nodes in multiple consecutive periods are obtained from the SRS signals of multiple consecutive periods obtained by a base station; since the scatterers in the nodes determined in different periods can be different, for example, as time changes, the scatterers in the environment of the target vehicle also change, in the case of taking the scatterers in the previous period as the reference, new scatterers can appear in the next period, and the scatterers can also disappear, and the number of scatterers can also change, therefore, the motion feature map generated by matching the motion parameters of the current period and the trajectory feature map generated by the historical trajectory information are matched, and the trajectory information (motion sequence with time continuity) of the nodes after re-matching is determined, the way of obtaining the trajectory information by the graph matching method is clear and intuitive, has high flexibility, and has higher robustness; after obtaining the motion sequence with time continuity, the motion features, trajectory features and map features can be obtained according to the motion sequence with time continuity and the high-precision map information, and the predicted trajectory information is obtained by the multi-feature joint prediction method, since the features used are multi-dimensional features, not only the motion features of the nodes themselves are considered, but also the interaction features between the nodes are considered, and the information of each road in the high-precision map data is also considered, therefore, the prediction is more accurate. Of course, the predicted trajectory information can also be used to correct the motion parameters of the nodes calculated in the subsequent period.

[0159] The process of predicting the trajectory information of each node according to the SRS signal is described in detail below.

[0160] Referring to FIGS. 1 to 3, Figure 5 and Figure 6 the process of obtaining the motion parameters of each node according to the SRS signal in step (1) in Figure 4 is described taking one period as an example.

[0161] In S101, the line-of-sight (LOS path) signal and the non-line-of-sight (NLOS path) signal of the SRS signal are separated by using the compression sensing technology, and the time delay, Doppler shift and angle of arrival of each path are obtained.

[0162] The LOS path refers to the path in which the signal propagates directly from the transmitting end to the receiving end without passing through any obstacles. In this path, the signal propagation loss is small, the signal strength is high, and the multipath effect is not obvious, so the signal quality is good.

[0163] The NLOS path refers to the path in which the signal encounters obstacles during propagation and needs to reach the receiving end through reflection, scattering or diffraction. In this path, the signal will undergo multiple reflections and attenuation, resulting in a decrease in signal strength, a significant multipath effect, and possible signal time delay and fading.

[0164] After the target vehicle sends out the SRS signal, the signal is scattered by the scatterer and reaches the base station. For the base station, the time-domain channel impulse response of the received signal can be generally expressed as:

[0165] ;

[0166] wherein, denotes the time-domain channel impulse response at the observation time t and the time delay , denotes the number of paths (including the LOS path and the NLOS path), is the complex attenuation coefficient of the path , denotes the phase change term, and j denotes the imaginary part of the complex number, is the Doppler shift of the path , is the receiving antenna array pattern, is the angle of arrival of the path , is the impulse function, is the time delay of the path , denotes the dimension of the time-domain channel impulse response , i.e., a complex number vector, and N denotes the number of antennas of the base station receiving the signal.

[0167] The channel measurement signal in the time domain is the convolution of the time-domain channel impulse response and the time-domain waveform, and the frequency-domain waveform adopts denotes, and the signal received by the base station is expressed as an OFDM subcarrier domain in the discrete frequency domain:

[0168] ;

[0169] wherein k denotes a sparse sampling point, i.e., the kth sampling point, is an additive white Gaussian noise.

[0170] After high sparse sampling and introduction of a sparse dictionary matrix , the signal received by the base station is expressed as:

[0171] ;

[0172] wherein, denotes the time-domain channel impulse response of the path , is the time delay of the path , is the Doppler shift of the path , is the angle of arrival of the path .

[0173] Since the real time delay, Doppler shift and angle of arrival are continuous, in order to apply compressed sensing, the possible value range of these parameters is discretized into a grid, thus converting into a sparse representation based on compressed sensing:

[0174] ;

[0175] wherein, There are only a few sparse bases and non-zero, corresponding to the sparsity of the multipath propagation path.

[0176] The separation of the LOS path and the NLOS path can be achieved by a sparse signal recovery technique, which includes but is not limited to a greedy algorithm and a convex optimization method. The greedy algorithm can be an orthogonal matching pursuit (OMP), and the convex optimization method can be a Lasso algorithm (Least absolute shrinkage and selection operator).

[0177] After separation, a sparse dictionary matrix is obtained , which includes the time delay, Doppler shift and angle of arrival of each path.

[0178] S102, according to the angle of arrival of the LOS path signal, the propagation distance and the Cartesian coordinates of the base station, the Cartesian coordinates of the target vehicle are obtained.

[0179] In the embodiments of the present application, the angle of arrival of the LOS path signal has been estimated, and the Cartesian coordinates of the base station are known, which are represented as: .

[0180] The propagation distance of the LOS signal represents the distance from the target vehicle to the base station, and the time delay of the LOS path has also been obtained. The propagation distance of the LOS signal can be obtained according to the time delay of the LOS path signal and the speed of the LOS path signal (about the speed of light).

[0181] Before calculating the Cartesian coordinates of the target vehicle, first describe the position relationship between the base station, the target vehicle and the scatterer, and the various parameters that may be involved.

[0182] Referring to Figure 7 , the position relationship between the base station, the target vehicle and the scatterer, and the various parameters that may be involved are provided in the embodiments of the present application.

[0183] Base station coordinates ;

[0184] Target vehicle coordinates ;

[0185] Radial velocity of target vehicle (velocity on the line connecting target vehicle and base station) ;

[0186] Angle of arrival of target vehicle relative to base station ;

[0187] Angle of moving direction of target vehicle in coordinate system ;

[0188] Moving scatterer coordinate ;

[0189] Radial velocity of moving scatterer (velocity on the line connecting scatterer and base station) ;

[0190] Angle of arrival of moving scatterer relative to base station ;

[0191] Angle of moving direction of moving scatterer in coordinate system ;

[0192] Angle of moving scatterer p relative to target vehicle k .

[0193] According to Figure 7 , the coordinates of base station are known, and the Cartesian coordinates of target vehicle are calculated by the following formula: Specifically, the Cartesian coordinates of target vehicle are calculated by the following formula:

[0194] ;

[0195] ;

[0196] wherein, is the propagation distance of LOS path signal, is the angle of arrival of LOS path signal, is the Cartesian coordinates of base station.

[0197] In S103, the driving direction of target vehicle is obtained according to the Cartesian coordinates of target vehicle and map information.

[0198] In the current urban environment, vehicles usually drive along the established road planning direction in a short time, and the driving direction of target vehicle (denoted as node k) can be obtained by using the coordinates of target vehicle and the road information in high-precision map data, for example, the coordinates of target vehicle determine a specific road in map information, and the direction of the road is the driving direction of target vehicle. The driving direction has an angle with the x-axis in the coordinate system, so the driving direction can be represented by the angle with the x-axis (antenna direction represents the x-axis), that is, the driving angle.

[0199] S104, obtaining the moving speed of the target vehicle k according to the radial speed of the target vehicle, the angle of arrival of the LOS path signal and the moving angle of the target vehicle.

[0200] After obtaining the moving angle of the target vehicle, the actual moving speed of the target vehicle is calculated:

[0201] ;

[0202] wherein, is the actual moving speed of the target vehicle k, denotes the speed of light, is the radial speed of the target vehicle (the speed component on the line connecting the base station and the target vehicle, which can be obtained according to the Doppler shift of the LOS path signal), is the carrier frequency, is the angle of the moving direction of the target vehicle in the coordinate system (or with the base station antenna), i.e. the moving angle; is the angle of arrival of the LOS path signal.

[0203] Since the scatterer reflects the SRS signal of the target vehicle to the base station, the above method for calculating the Cartesian coordinates, moving direction and moving speed of the target vehicle is no longer applicable to calculating the motion parameters of the scatterer. In the embodiments of the present application, the motion parameters of the scatterer are calculated by other methods.

[0204] Referring to Figure 6 as an example of calculating the motion parameters of the scatterer.

[0205] S201, calculating the position of the scatterer p according to the coordinates of the base station receiving the SRS signal, the propagation distance of the pth NLOS path signal and the angle of arrival of the pth NLOS path signal.

[0206] In the above process of calculating the motion parameters of the target vehicle, the related parameters of the target vehicle represented by the LOS index are used. In this step, the target vehicle is denoted as node k, and any scatterer is denoted as node p. In order to more clearly describe the parameters, the parameters of the target vehicle represented by the LOS in the above embodiments can be replaced by the index k.

[0207] As described above, the Cartesian coordinates of the base station are The Cartesian coordinates of the target vehicle k have been calculated in the above steps, which are denoted as: The estimated parameters corresponding to the pth NLOS path signal are , denotes the time delay of the pth NLOS path signal, denotes the Doppler shift of the pth NLOS path signal, represents the angle of arrival of the p-th NLOS path signal, i.e. the angle of scatterer p relative to the base station, the maximum value of p is related to the number of NLOS path signals.

[0208] Cartesian coordinates of scatterer p are:

[0209] ;

[0210] ;

[0211] wherein, represents the propagation distance of the p-th NLOS path signal, which can be determined according to the time delay of the p-th NLOS path signal and the propagation speed, represents the angle of arrival of the p-th NLOS path signal.

[0212] S202, calculate the relevant angle according to the Cartesian coordinates of the scatterer: the angle of scatterer p relative to the target vehicle k :

[0213] ;

[0214] wherein, represents the Cartesian coordinates of scatterer p, represents the Cartesian coordinates of target vehicle k.

[0215] S203, obtain the driving direction of scatterer p according to the Cartesian coordinates of scatterer p and the map information.

[0216] The road where the scatterer is located can be determined through the map information and the scatterer position, and the driving direction of the scatterer is determined according to the road direction, and the angle of the driving direction in the coordinate system (or with the antenna direction) is , i.e. the driving angle.

[0217] S204, obtain the driving speed of the p-th scatterer according to the Doppler shift, the angle of arrival and the driving angle of the p-th NLOS path signal, and the motion parameters of the target vehicle, and the driving angle of the p-th scatterer is determined by the driving direction of the p-th scatterer.

[0218] The Doppler shift of the p-th NLOS path includes three main parts:

[0219] ;

[0220] ;

[0221] ;

[0222] ;

[0223] wherein, represents the Doppler shift of the pth scatterer is the Doppler shift of the transmitted signal after being scattered by the pth scatterer, is the Doppler shift of the reflected signal after being scattered by the pth scatterer, is the Doppler shift of the received signal at the base station, is the carrier frequency, represents the speed of light, represents the driving speed of the target vehicle k, represents the included angle of the driving direction of the target vehicle k in the coordinate system, represents the angle of the pth scatterer relative to the target vehicle k. represents the driving speed of the pth scatterer, represents the included angle of the moving direction of the pth scatterer in the coordinate system, represents the angle of arrival of the pth NLOS path signal.

[0224] According to the above four equations related to the Doppler shift of the pth NLOS path signal, and the following parameters determined: , , , , the motion speed of the pth scatterer can be calculated.

[0225] It can be understood that the Cartesian coordinates, driving direction and driving speed of the target vehicle, and the Cartesian coordinates, driving direction and driving speed of each scatterer can be respectively calculated in the above manner.

[0226] The process of step (2) in the above embodiment will be described in detail. Figure 4

[0227] As described above, the motion parameters of multiple nodes can be obtained in each period. However, since the target vehicle is moving at all times, the multiple nodes obtained in each period may be different as the target vehicle moves. The embodiment of the present application needs to obtain the trajectory information (motion parameters in continuous multiple periods) of each node in the current period. Therefore, it is necessary to match each node in continuous multiple periods to obtain the motion parameters of each node in different periods.

[0228] The embodiment of the present application matches each node in the manner of graph matching to obtain the trajectory information of each node in the current period.

[0229] ​Figure 8 The process diagram for obtaining the current trajectory information according to the motion parameters of each node in the current period and the historical trajectory information is provided for the embodiments of the present application.

[0230] S301, generating a motion feature graph according to the encoded motion parameters of each node in the current period, wherein the nodes in the motion feature graph are vertex i;

[0231] S302, generating a trajectory feature graph according to the encoded historical trajectory information, wherein the nodes in the trajectory feature graph are vertex j;

[0232] S303, cyclically enhancing the initial vertex feature of vertex i in the motion feature graph to obtain the enhanced vertex feature of vertex i, wherein the initial vertex feature of vertex i is the motion parameter of vertex i in the current period;

[0233] S304, cyclically enhancing the initial vertex feature of vertex j in the trajectory feature graph to obtain the enhanced vertex feature of vertex j, wherein the initial vertex feature of vertex j is the historical motion parameter of vertex j;

[0234] S305, after obtaining the enhanced vertex feature of each vertex i in the motion feature graph, obtaining the edge feature in the motion feature graph according to the vertex features of any two vertices in the motion feature graph;

[0235] S306, after obtaining the enhanced vertex feature of each vertex j in the trajectory feature graph, obtaining the edge feature in the trajectory feature graph according to the vertex features of any two vertices in the trajectory feature graph;

[0236] S307, matching the motion feature graph and the trajectory feature graph according to the enhanced vertex features and edge features in the motion feature graph and the enhanced vertex features and edge features in the trajectory feature graph to obtain the current trajectory information.

[0237] Referring to Figure 9 The corresponding diagram for matching is provided for the embodiments of the present application. Figure 8

[0238] 2.1 Feature encoding.

[0239] Motion feature set: the motion parameters of the nodes (also referred to as vertices) collected in the current period are obtained. The motion feature set obtained in the current period is . .

[0240] Trajectory feature set: the historical state sequence set of the nodes (also referred to as vertices) obtained by matching in the last period is .​ Each state sequence including a series of historical motion parameters of the current jth node, the state sequence of each node j in the trajectory feature set (also vertex feature) The historical motion parameters in the state sequence have the same target index The motion parameters of a certain node in the state sequence are continuous for multiple periods, so it can also be recorded as a running sequence.

[0241] The motion features of each user equipment in the current period are encoded to obtain an encoding result The state sequence of each state sequence is encoded to obtain an encoding result These encoding results can be used as node features in the graph matching process.

[0242] In actual applications, feature encoding can also not be performed.

[0243] Referring to Figure 9 The motion feature graph (motion feature set generation) and the trajectory feature graph (trajectory feature generation) provided by the embodiments of the present application, the motion feature graph is defined as: The trajectory feature graph is defined as: .

[0244] The vertex i in the graph represents the ith motion parameter in the motion feature graph , The vertex j represents the jth state sequence in the trajectory feature graph , .

[0245] The edge in the graph represents the relationship between nodes in the motion feature graph, The edge represents the relationship between nodes in the trajectory feature graph, .

[0246] Therefore, the inter-frame association problem of multiple nodes is converted into a graph matching problem between the motion feature graph and the trajectory feature graph .

[0247] 2.2 Feature enhancement.

[0248] In order to more effectively utilize the structural information and contextual information between nodes in the two feature graphs, a graph convolutional network (GCN) can be used to enhance the initial vertex features (motion parameters and state sequences) obtained by encoding.

[0249] ​The initial vertex features can be input into a graph convolutional network for the first feature enhancement, resulting in the vertex features for the first iteration; then, the vertex features from the first iteration are input into the graph convolutional network for the second feature enhancement, resulting in the vertex features for the second iteration; and so on, in this manner, iteratively... After this, the enhanced final vertex features are obtained. And the enhanced edge features are obtained based on the final vertex features.

[0250] In this embodiment of the application, motion feature map Initial vertex features of vertex i in The initial vertex features of vertex j in the trajectory feature map are: .

[0251] In the In the next propagation process, the weight coefficient of the aggregation is defined as the cosine similarity of the geometric features between vertex i and vertex j: By analyzing the features of all vertices in the trajectory feature map We obtain the aggregated neighborhood features of vertex i by performing a weighted summation. Correspondingly, the input features of vertex i are in the th... The enhanced vertex features obtained after the next propagation are as follows:

[0252] ;

[0253] MLP stands for Multilayer Perceptron.

[0254] The input features of vertex j are in the th... The enhanced vertex features obtained after the next propagation are as follows:

[0255] ;

[0256] in, This represents the enhanced vertex feature of vertex i obtained in the previous iteration. For the enhanced vertex features of vertex i obtained in this iteration, MLP represents a multilayer perceptron, and the aggregated neighborhood features of vertex i are... Cosine similarity of geometric features between vertices i and j .

[0257] Following the above method After several iterations of enhancement, the final enhanced feature of vertex i is obtained. And the enhanced features of the final vertex j Of course, the enhanced features obtained can also be obtained through... Normalization is performed.

[0258] Edges in motion feature graph edge feature of an edge in the trajectory feature graph, by connecting the enhanced features of the two vertices of the edge and after normalization is: .

[0259] Correspondingly, the edge edge feature of an edge in the trajectory feature graph, by connecting the enhanced features of the two vertices of the edge and after normalization is: .

[0260] The enhanced vertex features and the constructed edge features can be used as a core matrix for subsequent matching of the two graphs.

[0261] 2.3, graph matching.

[0262] Referring to Figure 10 , the graph matching process according to the vertex features and the edge features of the two graphs includes the following steps:

[0263] S401, calculating vertex similarity between vertex features in the motion feature graph and vertex features in the trajectory feature graph, and edge similarity between edge features in the motion feature graph and edge features in the trajectory feature graph;

[0264] S402, maximizing the vertex similarity and the edge similarity to obtain a matching score of a vertex in the motion feature graph and a vertex in the trajectory feature graph, the vertex in the motion feature graph being a motion parameter of a node, and the vertex in the trajectory feature graph being a motion sequence of each node, wherein a first vertex in the motion feature graph is matched with a motion sequence having the highest matching score with the first vertex, and the first vertex is any vertex in the motion feature graph;

[0265] S403, if the first vertex in the motion feature graph does not exist a matched state sequence, the first vertex is taken as a newly appeared node to generate a state sequence of the first vertex;

[0266] S404, if the first vertex in the motion feature graph exists a matched state sequence, and the Mahalanobis distance between the position of the first vertex and the predicted position of the matched state sequence is greater than a distance threshold, the first vertex is deleted;

[0267] S405, if the first vertex in the motion feature graph exists a matched state sequence, and the Mahalanobis distance between the position of the first vertex and the predicted position of the matched state sequence is less than or equal to the distance threshold, the motion parameter of the first vertex is added to the matched state sequence;

[0268] S406 If there is a first state sequence in which no motion parameter of any vertex is added in the trajectory feature graph for a plurality of continuous periods, the first state sequence is deleted.

[0269] S401 and S402 in the graph matching process are described below through specific embodiments.

[0270] Embodiments of the present application convert the graph matching process into a quadratic assignment problem and solve it through relaxation and optimization techniques, while maintaining the end-to-end differentiability for training through a neural network.

[0271] 2.31 Perform quadratic assignment problem (QAP) graph matching modeling:

[0272] ;

[0273] ;

[0274] wherein, is a permutation matrix, is the number of vertices of the motion feature graph, is the number of vertices of the trajectory feature graph, is the weighted adjacency matrix of the motion feature graph, is the weighted adjacency matrix of the trajectory feature graph. is the vertex similarity matrix of the two graphs, represents the consistency between the edges of the two feature graphs, represents the similarity between the vertices of the two feature graphs, is a vector of all 1s.

[0275] 2.32 Convert the quadratic assignment graph matching model into a quadratic programming (QP) model and solve it through relaxation and optimization techniques.

[0276] Constraint is a double random matrix X, after constraint , , the scalar edge weight is expanded into a d-dimensional edge feature vector , the weighted adjacency matrix becomes a weighted adjacency tensor and .

[0277] The quadratic programming model is obtained as follows:

[0278] ;

[0279] After vectorization, the solution of the following quadratic programming problem is obtained, that is, the solution of the quadratic programming model:

[0280] ;

[0281] ;

[0282] wherein, , , M is a symmetric quadratic similarity matrix between edge features in the motion feature graph and edge features in the trajectory feature graph; , denotes the edge feature cosine similarity matrix, ; denotes the starting vertex association matrix of the motion feature graph; denotes the starting vertex association matrix of the sequence graph; denotes the ending vertex association matrix of the motion feature graph; denotes the ending vertex association matrix of the sequence graph; the element of the vertex similarity matrix B is to enhance the feature cosine similarity between the post-vertices. denotes the motion parameters of each node in the motion feature graph, denotes the relationship between nodes in the motion feature graph, denotes the state sequence of each node in the trajectory feature graph, denotes the relationship between nodes in the trajectory feature graph.

[0283] 2.33 Differentiable solving and gradient backpropagation.

[0284] In the process of training the quadratic programming model, two parameters are involved: gradient and loss function.

[0285] The loss function is a function that measures the gap between the output of the neural network (predicted value) and the true label. The smaller the value of the loss function, the closer the prediction result of the neural network to the true value, and vice versa. Therefore, the training goal of the neural network is to minimize the loss function.

[0286] The gradient is the derivative of the loss function with respect to each parameter (such as weight and bias). It is used to represent the change direction and speed of the loss function at the current point. According to the calculated gradient, the model parameters can be updated in the opposite direction.

[0287] In the training process, first, the predicted value of the model is calculated by forward propagation, and the loss function is used to measure the gap between the predicted value and the true value; then, through backpropagation, the gradient of the loss function with respect to each model parameter is calculated; after obtaining the gradient, the gradient descent method is used to update the parameters of the model in the opposite direction of the gradient, thereby gradually reducing the value of the loss function in the subsequent iteration process. Repeat the above process until the model converges (or reach the number of iterations).

[0288] For the loss function, the weighted binary cross-entropy is adopted as the loss function:

[0289] ;

[0290] where, is the predicted matching score between the target and the sequence , is the true label value, is the balance factor, The original matching score is processed using the softmax function to obtain: .

[0291] where, is the temperature parameter. is the last matching score.

[0292] For the gradient, the solution of the QP problem and its dual variables satisfy the Karush-Kuhn-Tucker (KKT) conditions. Based on the implicit function theorem and the KKT conditions, the backpropagation gradient of the graph matching layer can be derived, so that the entire module can be trained end-to-end.

[0293] Therefore, for the QP problem, the conditions that need to be satisfied include:

[0294] ;

[0295] ;

[0296] The gradient of the parameter can be derived according to the KKT conditions.

[0297] After training converges (or reaches the number of iterations), the final predicted matching score is obtained, which can approximate the true value.

[0298] The matching score matrix X output by the QP layer is reshaped by x ( ), and the matching score matrix X output by the QP layer may be a continuous value. A greedy strategy is used to generate the final binary permutation assignment matrix from X, that is, each target is matched with the sequence with the maximum matching score value.

[0299] Of course, in actual applications, you can also use libraries such as qpth to build differentiable QP layers and use libraries such as CVXPY to forward solve QP problems.

[0300] 2.34 Filter management.

[0301] As an example of the filtering process:

[0302] If a currently detected target With all existing sequences Predicted location Mahalanobis distance between All are greater than the preset threshold If the target is not matched with any existing sequence, the match is considered invalid, and the match will be deleted even if it achieves a high matching score in the differentiable graph matching layer. ).

[0303] If the target Unable to match any existing sequence (i.e.) If it is a newly emerging target, then it will be considered as a new sequence. , The initial state is added to the sequence set. .

[0304] If the target with sequence Successful match (i.e.) And if the sequence passes the Mahalanobis distance check, then update the sequence: for .

[0305] If a sequence In continuous If no update is received within a certain time period (i.e., no new observed target matches it), it is considered that the target corresponding to the sequence has left the current environmental range or disappeared, and the sequence will be removed from the candidate sequence set. Delete it. This is called the maximum lifetime of the sequence.

[0306] The trajectory feature set at the end of the matching can be obtained in this way: .

[0307] Continuing with a single cycle as an example, let's describe... Figure 4 Step (3) is the process of obtaining the motion characteristics, interaction characteristics and map characteristics of each node based on the current trajectory information and map data.

[0308] In this embodiment, motion features, interaction features, and map features are extracted through three modules. It should be noted that the nodes have been redefined in this step; the nodes in this step are denoted by i, and the node i in this step is different from the node i in the graph matching.

[0309] 3.1 The process of extracting motion features is described below.

[0310] The motion feature can be obtained by encoding the motion feature in the trajectory feature set through a GRU (gated recurrent unit) and extracting sequence information therein. The motion feature is a long-term motion trajectory of each target node.

[0311] ;

[0312] wherein, represents a history encoder shared by all target nodes of type K, this step generates a history encoder of type K, K represents the type of node i, K=1 represents a high-speed moving type, K=2 represents a low-speed moving type, represents a dynamic feature of node i, represents a state sequence in the trajectory feature set.

[0313] Therefore, the final output motion feature is: .

[0314] 3.2 The process of extracting the interaction feature is described below.

[0315] Referring to Figure 11 , the flowchart for extracting the interaction feature provided by the embodiment of the present application is shown.

[0316] S501, a directed edge heterogeneous feature graph is constructed through the motion feature of each node and the directed edge feature, wherein the directed edge feature includes an edge weight and an edge type.

[0317] In the embodiment of the present application, first, a directed edge feature heterogeneous graph is defined: , is a set of n nodes. is a set of directed edges. Each node contains its own node feature and belongs to a specific type, and each directed edge contains an edge weight and belongs to a specific edge type. The directed edge feature heterogeneous graph can represent the interaction relationship between different nodes, so the HEAT network encoder can be used to extract the interaction feature from the constructed directed edge feature heterogeneous graph.

[0318] Then, the motion feature (long-term motion trajectory of each node) of each node obtained by the GRU encoding module is written into the constructed directed edge feature heterogeneous graph, and the interaction feature between each node is modeled:

[0319] ;

[0320] wherein, represents the interaction feature of node i at time t, contains the interaction feature between all nodes. is the motion feature of each node extracted by the GRU, For the directed edge feature set, For the directed edge feature from node j to node i, containing edge weight and edge type, where node j is in the neighborhood of node i Within the range of 30 meters, the motion of node i will be affected by node j, and the neighborhood of node i can be set to 30 meters, which is only an example.

[0321] In the above model, the edge weight contained in the directed edge feature from node j to node i represents the relative motion state of node j to node i, which is the difference between the motion state vectors of the two: .

[0322] In the above model, the edge type contained in the directed edge feature from node j to node i represents the combination of the type of node j and the type of node i, for example, the directed edge type from node j of type [1, 0, 0] to node i of type [0, 0, 1] is set to [1, 0, 0, 0, 0, 1].

[0323] S502, extracting the initial node feature of each node in the directed edge heterogeneous feature graph.

[0324] After constructing the directed edge feature heterogeneous graph, GNN can be used to process the heterogeneity of nodes, directed edges and edge features to obtain the initial node feature , wherein is the feature vector of node i, which contains a set of edge weights and edge types.

[0325] The edge weight is , wherein is the edge attribute in .

[0326] The edge type is , wherein is the edge type in .

[0327] Currently, GNN cannot directly process the above interaction representation problem, so the application extends the traditional GNN and adds a HEAT layer to aggregate node neighborhood information and update its own node features by adding edge-enhanced attention mechanism, so that the target node feature can be obtained according to the initial node feature, and finally the target node features of each node are used as the extracted node features.

[0328] S503, projecting the type of the initial node feature of node i to a feature space shared by each type to obtain the type-converted initial node feature of node i;

[0329] S504, respectively convert the edge weight and the edge type of the node i to obtain the converted edge feature of the node i, the converted edge feature of the node i including: the converted edge weight and the converted edge type.

[0330] Different types of nodes in the directed edge feature heterogeneous graph often have different feature spaces, and in actual processing, it is often necessary to project them into a feature space shared by all types. Therefore, a node type conversion matrix is used to convert the two types of nodes, and the conversion can be represented as: .

[0331] In addition, the edge weight is often a measurement value in a continuous space, for example, the distance between two nodes and the speed difference; the edge type is a discrete value indicator. Therefore, an edge weight conversion matrix and an edge type conversion matrix are also needed to convert the two features: , . For an edge from node j to node i, the edge feature , that is, the edge feature is the connection of the converted edge weight and the converted edge type.

[0332] S505, the converted edge feature of the node i and the converted initial node feature of the node i are spliced into the feature vector of the node i.

[0333] For node i, the connected feature vector becomes: . The connected feature vector represents the influence of node j on node i.

[0334] S506, according to the feature vector of the node i and the converted initial node feature, the attention coefficient of the node i is obtained.

[0335] In the embodiments of the application, the connected feature vector of the node i will be sent to a shared attention mechanism, that is, a single-layer feedforward neural network , and then after being nonlinearly normalized by a LeakyReLU function, the attention coefficient :

[0336] ;

[0337] wherein, , is the angle of the control negative slope; is the feature vector of the node i, is the converted initial node feature of the node i; represents a single-layer feedforward network.

[0338] The attention coefficient fully considers the continuous multi-dimensional edge feature.

[0339] The input of the HEAT layer includes the initial node feature and the feature vector of node i.

[0340] S507, the attention coefficient of node i is taken as a weight to combine the feature vectors of each node in the neighborhood of node i, to obtain a neighborhood feature vector of node i;

[0341] S508, the neighborhood feature vector of node i is updated by using a sigmoid function, to obtain a target node feature of node i.

[0342] In this application, the obtained attention coefficient is taken as a weight to linearly combine the integrated features of all nodes in the neighborhood of node i, and the feature of node i is updated by using a sigmoid function, and finally, the target node feature of node i is output.

[0343] ;

[0344] wherein, is a sigmoid function, denotes the neighborhood feature vector of node i, which includes the converted edge weight and the converted initial node feature, and does not include the converted edge type; this is mainly because the edge type has been considered when calculating the attention coefficient. It can be understood that the output of the HEAT layer is a new set of node features of node i , i.e., the target node feature.

[0345] S509, after combining the target node features of each node, the interaction feature is obtained.

[0346] In the embodiments of the application, since the influence of node j on node i needs to be considered, the directed edge weight and the edge type between the two nodes need to be considered. The feature vector of the connected node i can jointly consider the importance of node j to node i in terms of node features and edge features. In the constructed directed edge feature heterogeneous graph, each node represents a target with a specific type and contains the motion features extracted by GRU. Compared with representing the interaction behavior as a homogeneous graph, an edge feature homogeneous graph or a heterogeneous graph without edge attributes, the represented interaction behavior between nodes is more comprehensive. For example, the heterogeneity between multiple target types is retained by the heterogeneous nodes, the spatial and motion relationships between all nodes are retained by the edge weight, and the difference in mutual influence between the two nodes with directed edges is considered.

[0347] 3.3 The process of extracting map features is described below.

[0348] First, the global map features in the high-precision map are extracted by the CNN, and then the local map features related to the position and speed of the node are selected from the global map features by the selection module. The local map features of each node are the map features extracted in this step. All nodes can share a global map, thereby saving storage space. For any node, the local map features can be selected from the global map features according to the position and speed of the node and other parameters.

[0349] ;

[0350] wherein, is the local map feature selected for node i, is a map selection module, and M is a global map feature matrix extracted from the bird's eye view of the complete road scene, is the position and speed of node i.

[0351] In the process of selecting the local map feature of node i, the application sets a selection gate in the map selection module:

[0352] ;

[0353] wherein, is a sigmoid function, is a weight matrix, is a feature of the map M extracted using the CNN, represents the combination of the position vector and the speed vector of node i at time t, is an offset. After the sigmoid function, the elements of are between 0 and 1.

[0354] The vector output by the selection gate can be used to select the local map feature of node i at time t:

[0355] ;

[0356] wherein, represents the Hadamard product, i.e., the element-wise product.

[0357] After obtaining the local map features of each node, the local map features of each node are the map features extracted in this step.

[0358] The process of step (4) in the above formula (1) is described as follows. Figure 4

[0359] ​In the embodiments of the present application, different types of heterogeneous decoding predictors are used to predict the subsequent trajectory information of the later period.

[0360] For a node i of type K, the motion characteristics of the node i are jointly considered , the interaction characteristics with other nodes , and the current local map characteristics to obtain the connected high-dimensional hidden characteristics: .

[0361] The motion parameters of the node i of type K in the future multiple periods are predicted through the connected high-dimensional hidden characteristics.

[0362] ;

[0363] wherein, the decoding predictor shared by the nodes of type K.

[0364] In the embodiments of the present application, as described above, the motion parameters of the target vehicle and the scatterer are also estimated by calculation, so after obtaining the predicted motion parameters of the subsequent multiple periods, the motion parameters estimated according to the SRS signal can also be corrected through the predicted motion parameters of the subsequent multiple periods.

[0365] It should be understood that the size of the serial number of each step in the above embodiments does not mean the order of execution, and the execution order of each process should be determined according to its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0366] The embodiments of the present application also provide a computer readable storage medium, which stores a computer program, and the computer program can implement the steps in each of the method embodiments when running on an electronic device.

[0367] The embodiments of the present application also provide a computer program product, which, when running on an electronic device or a wireless router, enables the electronic device to implement the steps in each of the method embodiments.

[0368] The integrated unit, if implemented in the form of a software function unit and sold or used as an independent product, can be stored in a computer readable storage medium. Based on such understanding, the computer program for instructing the related hardware to complete all or part of the processes in the above-mentioned embodiments can be stored in a computer readable storage medium, and the computer program can implement the steps of the above-mentioned method embodiments when executed by a processor. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or some intermediate forms. The computer readable medium at least includes any entity or device capable of carrying the computer program code to the first device, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunications signal, and a software distribution medium. For example, a U disk, a mobile hard disk, a magnetic disk or an optical disk, etc. In some jurisdictions, according to legislation and patent practice, the computer readable medium cannot be an electrical carrier signal and a telecommunications signal.

[0369] The embodiments of the present application also provide a chip, which includes a processor and a memory. The processor is coupled with the memory. The processor invokes a computer program stored in the memory to implement the steps of any method embodiment of the present application. The chip can be a single chip or a chip module composed of multiple chips.

[0370] In the above embodiments, the description of each embodiment has its own focus. The parts not described or recorded in detail in a certain embodiment can be referred to the relevant description of other embodiments.

[0371] Those skilled in the art can appreciate that the units and method steps of the examples described in combination with the embodiments disclosed herein can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether the functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. The skilled person can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0372] The above embodiments are only used to illustrate the technical solutions of the present application, but not limit them. Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that the technical solutions recorded in the foregoing embodiments can be modified, or some technical features can be replaced by equivalent ones. Such modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.

Claims

1. A vehicle trajectory prediction method, characterized by, The method comprises: acquiring a SRS signal periodically transmitted by a target vehicle and received by a base station; obtaining motion parameters of each node including the target vehicle and scatterers around the target vehicle according to the SRS signal, the motion parameters including position, driving speed and driving direction; obtaining current trajectory information containing the motion parameters of each node in the current period according to the motion parameters of each node in the current period and historical trajectory information, the current trajectory information serving as historical trajectory information in the next period; obtaining motion features, interaction features and map features of each node according to the current trajectory information and map data; obtaining predicted trajectory information of each node in at least one subsequent period according to the motion features, interaction features and map features of each node; wherein the obtaining of the motion parameters of each node according to the SRS signal comprises: separating a LOS path signal and NLOS path signals of the SRS signal to obtain a first angle of arrival, a first time delay and a first Doppler shift of the LOS path signal, and a second angle of arrival, a second time delay and a second Doppler shift of each NLOS path signal; obtaining the motion parameters of the target vehicle according to the first angle of arrival, the first time delay and the first Doppler shift of the LOS path signal, wherein the first Doppler shift is used to obtain a first radial speed of the target vehicle, and the first radial speed is used to obtain a first driving speed of the target vehicle; obtaining a second driving speed of the pth scatterer according to the second angle of arrival, the second time delay and the second Doppler shift of the pth NLOS path signal, and the motion parameters of the target vehicle; wherein the second Doppler shift of the pth NLOS path signal comprises: a third Doppler shift generated through the pth scatterer, the third Doppler shift being related to the first driving speed, a first driving angle of the target vehicle, and an angle between the target vehicle and the pth scatterer; a fourth Doppler shift generated when the pth scatterer reflects the signal, the fourth Doppler shift being related to the third Doppler shift, the second driving speed, a second driving angle of the pth scatterer, and the angle between the target vehicle and the pth scatterer; and a fifth Doppler shift generated by the base station receiving the SRS signal, the fifth Doppler shift being related to the fourth Doppler shift, the first driving speed of the target vehicle, the second driving angle and the second angle of arrival of the pth scatterer.

2. The method of claim 1, wherein, The obtaining of the motion parameters of each node according to the SRS signal comprises: obtaining a first coordinate of the target vehicle according to a coordinate of the base station receiving the SRS signal, a first propagation distance of the LOS path signal and the first angle of arrival, wherein the first propagation distance of the LOS path signal is determined by the first time delay; obtaining a first driving direction of the target vehicle according to the first coordinate of the target vehicle and a road direction in the map data. The first travel speed of the target vehicle is obtained according to the first radial speed of the target vehicle, the first arrival angle and the first travel angle of the target vehicle, the first radial speed being a speed of the target vehicle in a direction towards a base station, the first radial speed being obtained according to the first Doppler shift, and the first travel angle of the target vehicle being determined by the first travel direction.

3. The method of claim 2, wherein, The first travel speed of the target vehicle is obtained according to the first radial speed of the target vehicle, the first arrival angle and the first travel angle of the target vehicle, the first radial speed being a speed of the target vehicle in a direction towards a base station, the first radial speed being obtained according to the first Doppler shift, and the first travel angle of the target vehicle being determined by the first travel direction. ； wherein, is a first travel speed of the target vehicle, denotes the speed of light, is a first radial speed of the target vehicle, is a carrier frequency, is a first travel angle of the target vehicle; is a first angle of arrival of the target vehicle.

4. The method of claim 2 or 3, wherein, The motion parameters of each node are obtained according to the SRS signals. The second coordinate of the pth scatterer is obtained according to the coordinates of the base station receiving the SRS signals, the second propagation distance of the pth NLOS path signal and the second arrival angle of the pth NLOS path signal, wherein the second propagation distance of the pth NLOS path signal is determined by the second time delay of the pth NLOS path signal, and the maximum value of p is related to the number of NLOS path signals; The second travel direction of the pth scatterer is obtained according to the second coordinate of the pth scatterer and the road direction in the map data; The second travel speed of the pth scatterer is obtained according to the second Doppler shift, the second arrival angle and the second travel angle of the pth NLOS path signal, and the motion parameters of the target vehicle, the second travel angle being determined by the second travel direction.

5. The method of claim 4, wherein, The second travel speed of the pth scatterer is obtained according to the second Doppler shift, the second arrival angle and the second travel angle of the pth NLOS path signal, and the motion parameters of the target vehicle, the second travel angle being determined by the second travel direction. ； ； ； ； wherein, denotes a second Doppler shift of the p-th scatterer, denotes the third Doppler shift, denotes the fourth Doppler shift, denotes the fifth Doppler shift, is a carrier frequency, denotes the speed of light, denotes a first travel speed of the target vehicle calculated, denotes a first travel angle of the target vehicle, denotes an angle between the p-th scatterer and the target vehicle; denotes a second travel speed of the p-th scatterer calculated, denotes a second travel angle of the p-th scatterer, denotes a second angle of arrival of the p-th NLOS path signal.

6. The method of claim 1, wherein, The current trajectory information containing the motion parameters of each node in the current period is obtained according to the motion parameters of each node in the current period and the historical trajectory information. A motion feature graph is generated according to the motion parameters of each node in the current period, and the nodes in the motion feature graph are vertices i; A trajectory feature graph is generated according to the historical trajectory information, and the nodes in the trajectory feature graph are vertices j; The initial vertex feature of vertex i in the motion feature graph is cyclically enhanced to obtain an enhanced vertex feature of vertex i, and the initial vertex feature of vertex i is the motion parameter of vertex i in the current period; The initial vertex feature of vertex j in the trajectory feature graph is cyclically enhanced to obtain an enhanced vertex feature of vertex j, and the initial vertex feature of vertex j is the historical motion parameter of vertex j; After obtaining the enhanced vertex feature of each vertex i in the motion feature graph, an edge feature in the motion feature graph is obtained according to the vertex features of any two vertices in the motion feature graph; After obtaining the enhanced vertex feature of each vertex j in the trajectory feature graph, an edge feature in the trajectory feature graph is obtained according to the vertex features of any two vertices in the trajectory feature graph; The motion feature graph and the trajectory feature graph are matched according to the enhanced vertex features and edge features in the motion feature graph, and the enhanced vertex features and edge features in the trajectory feature graph, to obtain the current trajectory information.

7. The method of claim 6, wherein, The initial vertex feature of vertex i in the motion feature graph is cyclically enhanced to obtain an enhanced vertex feature of vertex i, including: In each cyclic enhancement process of the vertex feature of vertex i in the motion feature graph, the enhanced vertex feature of vertex i is: ； wherein, represents the enhanced vertex feature of vertex i obtained in the last loop, represents the enhanced vertex feature of vertex i obtained in the current loop, MLP represents a multi-layer perceptron, and the aggregated neighborhood feature of vertex i , the geometric feature cosine similarity between vertex i and vertex j .

8. The method of claim 6 or 7, wherein, The current trajectory information includes a motion sequence set of multiple nodes, and the motion feature graph and the trajectory feature graph are matched according to the enhanced vertex feature and edge feature in the motion feature graph and the enhanced vertex feature and edge feature in the trajectory feature graph to obtain the current trajectory information, including: The vertex similarity between the vertex feature in the motion feature graph and the vertex feature in the trajectory feature graph and the edge similarity between the edge feature in the motion feature graph and the edge feature in the trajectory feature graph are calculated; The vertex similarity and the edge similarity are maximized to obtain a matching score of the vertex in the motion feature graph and the vertex in the trajectory feature graph, the vertex in the motion feature graph being a motion parameter of a node, and the vertex in the trajectory feature graph being a motion sequence of each node, wherein a first vertex in the motion feature graph is matched with a motion sequence with the highest matching score of the first vertex, and the first vertex is any vertex in the motion feature graph; If the first vertex in the motion feature graph does not exist a matched state sequence, the first vertex is generated as a newly appearing node to generate a state sequence of the first vertex; If the first vertex in the motion feature graph exists a matched state sequence, and the Mahalanobis distance between the position of the first vertex and the predicted position of the matched state sequence is greater than a distance threshold, the first vertex is deleted; If the first vertex in the motion feature graph exists a matched state sequence, and the Mahalanobis distance between the position of the first vertex and the predicted position of the matched state sequence is less than or equal to the distance threshold, the motion parameter of the first vertex is added to the matched state sequence; If there are a plurality of first state sequences in the trajectory feature graph without adding the motion parameter of the first vertex in a plurality of periods, the first state sequence is deleted.

9. The method of claim 8, wherein, The maximization of the vertex similarity and the edge similarity to obtain the matching score of the vertex in the motion feature graph and the vertex in the trajectory feature graph includes: A quadratic programming model related to the vertex similarity and the edge similarity is trained to obtain a trained quadratic programming model; After the training is completed, the matching score of the vertex in the motion feature graph and the vertex in the trajectory feature graph is obtained; The quadratic programming model is: ； where d is the dimension of the feature vector of edge weight, , is the weighted adjacency tensor of the motion feature map, , is the weighted adjacency tensor of the motion feature map, , B is used to represent vertex similarity, B is a vertex similarity matrix, the element , represents the similarity between the vertex of the motion feature map and the vertex of the trajectory feature map, is a permutation matrix, is the number of vertices of the motion feature map, is the number of vertices of the trajectory feature map, c is a natural number from 1 to d, represents the enhanced vertex feature of vertex i, represents the enhanced vertex feature of vertex j; A loss function used for training the quadratic programming model is: ； wherein, ; is a predicted matching score between a vertex in the motion feature map and a vertex in the trajectory feature map, is a true label value, is a balancing factor, is a temperature parameter, is a last matching score.

10. The method of claim 1, wherein, According to the current trajectory information and the map data, the motion feature, the interaction feature and the map feature of each node are obtained, including: The motion feature is: ​​ wherein, represents a history encoder shared by all target nodes of type K, K represents the type of node i, K=1 represents the type of high-speed mobile, K=2 represents the type of low-speed mobile, represents the dynamic characteristics of node i, represents the state sequence in the current trajectory information.

11. The method of claim 10, wherein, According to the current trajectory information and the map data, the motion feature, the interaction feature and the map feature of each node are obtained, including: A directed edge heterogeneous feature graph is constructed through the motion feature of each node and the directed edge feature, and the directed edge feature includes an edge weight and an edge type; extract initial node features of each node in the directed edge heterogeneous feature graph; project a type of the initial node features of node i into a feature space shared by each type to obtain converted initial node features of node i; convert the edge weight and the edge type of node i respectively to obtain converted edge features of node i, the converted edge features of node i including converted edge weight and converted edge type; concatenate the converted edge features of node i and the converted initial node features of node i into a feature vector of node i; obtain an attention coefficient of node i according to the feature vector of node i and the converted initial node features of node i; combine feature vectors of each node in the neighborhood of node i using the attention coefficient of node i as a weight to obtain a neighborhood feature vector of node i; update the neighborhood feature vector of node i using a sigmoid function to obtain target node features of node i; combine target node features of each node to obtain the interaction features.

12. The method of claim 11, wherein, The obtaining of the attention coefficient of node i according to the feature vector of node i and the converted initial node features of node i includes: ； wherein, , is the angle of the negative slope; is the eigenvector of node i influenced by node j, is the transformed initial node feature of node i; denotes a single-layer feedforward network, is the number of vertices of the trajectory feature map, is the eigenvector of node i influenced by node k.

13. The method of claim 12, wherein, The updating of the neighborhood feature vector of node i using the sigmoid function to obtain the target node features of node i includes: ； wherein, is a sigmoid function, denotes the neighborhood feature vector of node i, containing the transformed edge weight and the transformed initial node feature, but not the transformed edge type, denotes the total number of j in the neighborhood range of node i, is the attention coefficient of node i, is a weight matrix, denotes the transformed edge weight of node i, denotes the transformed initial node feature of node j.

14. The method of claim 1, wherein, The obtaining of the motion features, the interaction features and the map features of each node according to the current trajectory information and the map data includes: ； wherein, represents a map feature of node i, represents a selection gate, with values between 0 and 1, represents a Hadamard product, is a feature of the map data extracted using a CNN, , is a sigmoid function, is a weight matrix, represents a combination of the position vector and velocity vector of node i at time t, is an offset.

15. An electronic device, comprising: The electronic device includes one or more processors and one or more memories; wherein the one or more memories are coupled to the one or more processors, and the one or more memories are configured to store a computer program, and when the one or more processors execute the computer program, the electronic device is caused to execute the method according to any one of claims 1-14.

16. A chip system applied to an electronic device, the chip system comprising one or more processors, characterized in that, The processor is configured to invoke computer instructions to cause the electronic device to execute the method according to any one of claims 1-14.

17. A computer readable storage medium comprising a computer program, characterized in that, When the computer program runs on the electronic device, the electronic device is caused to execute the method according to any one of claims 1-14.

18. A computer program product, characterised in that, When the computer program product runs on the electronic device, the electronic device is caused to execute the method according to any one of claims 1-14.

Images