CN115523934A

CN115523934A - Vehicle track prediction method and system based on deep learning

Info

Publication number: CN115523934A
Application number: CN202211012871.6A
Authority: CN
Inventors: 张睿凡
Original assignee: SAIC Volkswagen Automotive Co Ltd
Current assignee: SAIC Volkswagen Automotive Co Ltd
Priority date: 2022-08-23
Filing date: 2022-08-23
Publication date: 2022-12-27

Abstract

The invention discloses a vehicle track prediction method based on deep learning, which comprises the following steps: collecting driving data of all vehicles in an environment range; acquiring local map data of the position of a track prediction target vehicle; corresponding the position information of all vehicles and the local map data to each other based on the time stamps; acquiring position information and lane line information of the own vehicle, the track prediction target vehicle and other vehicles at the same time based on the timestamp, and encoding the position information and the lane line information of the vehicles to obtain a position information encoding vector and a lane line information encoding vector of the track prediction target vehicle; inputting the position information coding vector and the lane line information coding vector into a first neural network for training so as to output a hidden vector representing a decision basis of a driving habit of a trajectory prediction target vehicle; and inputting the hidden vector and the historical track of the track prediction target vehicle into a second neural network, namely outputting the predicted track of the track prediction target vehicle.

Description

Vehicle track prediction method and system based on deep learning

Technical Field

The present invention relates to a method and a system for predicting a vehicle trajectory, and more particularly, to a method and a system for predicting a vehicle trajectory.

Background

With the development of intelligent driving technology, the method has more and more important significance for the track prediction of the intelligent vehicle.

In the prior art, most of the currently adopted vehicle trajectory prediction methods are based on historical states such as the position and the speed of a vehicle, and the like, detection and analysis are performed on surrounding environments, and vehicle trajectories are predicted by detecting and identifying surrounding dynamic information such as the position of the vehicle, the position of a pedestrian and speed information and combining scene information such as lane line position, lane line trend and signal lamps contained in a high-definition map where the vehicle is located by self positioning.

However, the vehicle trajectory prediction in the prior art is based on the trajectories of surrounding vehicles and lane line points of a high-definition map, and the influence of these position quantities on the trajectory of the target vehicle is rather qualitative rather than quantitative. The accuracy of the prediction is not high.

Based on this, it is desirable to obtain a method that effectively improves the accuracy of vehicle trajectory prediction.

Disclosure of Invention

The invention aims to provide a vehicle track prediction method based on deep learning, which can encode driving data and local map data around a track prediction target vehicle based on a rule encoding mode, represents a driving habit decision thought of a driver through hidden variables by combining a neural network, and finally predicts a future vehicle track of the target vehicle based on the hidden variables and the historical track of the track prediction target vehicle.

In order to achieve the above object, the present invention provides a vehicle trajectory prediction method based on deep learning, which includes the steps of:

acquiring running data of all vehicles in an environment range, wherein the vehicles comprise own vehicles, track prediction target vehicles and other vehicles, and the running data at least comprises position information of the vehicles;

obtaining local map data of the position of a track prediction target vehicle, wherein the local map data at least comprises lane line information;

corresponding the position information of all vehicles and the local map data to each other based on the time stamp;

acquiring position information and lane line information of the own vehicle, the track prediction target vehicle and other vehicles at the same moment based on the timestamp, and coding the position information of the vehicle at the moment according to a set rule to obtain a position information coding vector of the track prediction target vehicle; coding the lane line information at the moment according to a set rule to obtain a lane line information coding vector of the track prediction target vehicle;

inputting the position information coding vector and the lane line information coding vector into a first neural network, and training the first neural network so as to output a hidden vector representing a driving habit decision basis of a trajectory prediction target vehicle; and inputting the hidden vector and the historical track of the track prediction target vehicle into a second neural network, and enabling the second neural network to output the predicted track of the track prediction target vehicle.

The research finds that the image coding mode can fuse all relevant information into one image for representation, but a convolutional neural network is required to be used for extracting features subsequently, similar position and speed information is relatively refined and independent features, and one more packaging extraction increases complexity. The recurrent neural network RNN can encode the historical information into a hidden vector, but it cannot be fused well with independent dynamic information relative to different vehicles, pedestrians, etc. These all limit the more intuitive and comprehensive characterization of the current environmental state of the vehicle.

Therefore, after the above situation is considered and analyzed, the invention designs a new vehicle trajectory prediction method, which can encode the driving data (including the position information of the vehicle) around the trajectory prediction target vehicle and the local map data (including the lane line information) based on a regular encoding mode, represent the driving habit decision idea of the driver through hidden variables by combining with a Recurrent Neural Network (RNN), and finally predict the future vehicle trajectory of the target vehicle based on the hidden variables and the historical trajectory of the target vehicle.

In the present invention, it is necessary to encode the position information in the travel data of all vehicles within the range of the environment in which the trajectory prediction target vehicle is located, according to a set rule. The coding mode based on the set rule can be more convenient for subsequent targeted analysis on the model on the basis of keeping simplicity.

In the invention, a high-definition map is introduced, the influence of different road environments on the vehicle track decision is added, and the influence is different from the prior map attribute which emphasizes the central point position of a lane line.

In addition, the invention can fuse the current environment state with the historical track of the track prediction target vehicle by introducing the recurrent neural network so as to obtain more unique track prediction.

It should be noted that, in the technical solution designed by the present invention, all vehicles within the collection environment range need to be specifically divided into "own vehicle", "trajectory prediction target vehicle", and "other vehicles". For convenience of understanding, an example is that there are 9 cars in the collection environment range, where "own car" is one car, and for other 8 cars, the trajectory prediction needs to be performed on that car, and the car is the "trajectory prediction target car", and the other 7 cars are "other cars".

When the vehicle track prediction system based on deep learning designed by the invention is practically applied to predict the vehicle track, the actual driving data of the 'own vehicle', 'track prediction target vehicle' and 'other vehicles' are acquired on a test road through the 'own vehicle' sensor. The driving data at least comprises position information of the vehicle, and when the driving data are actually collected, the position track information of the surrounding vehicle during driving can be detected, identified and tracked through sensors such as a laser radar and a camera carried by the vehicle. Meanwhile, by combining the global coordinate given by the positioning navigation system of the 'own vehicle', a high-definition map drawn in advance can be obtained, and local map data of the position of the 'track prediction target vehicle' can be obtained. The collected local map data may include information such as "lane line position", "turn", "intersection" and "controlled".

In addition, in the vehicle trajectory prediction method designed by the present invention, the historical trajectory of the trajectory prediction target vehicle input into the second neural network is obtained from the acquired position information of the trajectory prediction target vehicle over a period of time.

Further, in the vehicle track prediction method based on deep learning of the present invention, the driving data is collected by using a vehicle end environment sensing device provided on the vehicle.

Further, in the vehicle track prediction method based on deep learning of the present invention, the vehicle-end environment sensing device at least includes a vehicle-mounted high-definition camera and/or a laser radar.

Further, in the vehicle trajectory prediction method based on deep learning according to the present invention, the position information of the vehicle is encoded by using a mesh division method to obtain a position information encoding vector having a fixed length.

Further, in the vehicle trajectory prediction method based on deep learning according to the present invention, a possible destination position or direction of the trajectory prediction target vehicle is determined according to the position of the trajectory prediction target vehicle in the local map and the lane line information in the local map; obtaining corresponding lane line information closest to the track prediction target vehicle based on the possible destination position or direction; and obtaining the lane line information coding vector based on the lane line information closest to the track prediction target vehicle.

Further, in the deep learning-based vehicle trajectory prediction method of the present invention, the first neural network is an RNN, and/or the second neural network is a full-connection mapping neural network.

Accordingly, another object of the present invention is to provide a vehicle trajectory prediction system based on deep learning, which can effectively implement the vehicle trajectory prediction method of the present invention to predict the target vehicle trajectory.

In order to achieve the above object, the present invention provides a vehicle trajectory prediction system based on deep learning, which includes:

the system comprises a running data acquisition device, a data processing device and a data processing device, wherein the running data acquisition device is used for acquiring running data of all vehicles in an environment range, the vehicles comprise own vehicles, track prediction target vehicles and other vehicles, and the running data at least comprises position information of the vehicles;

the map data acquisition module is used for acquiring local map data of the position where the track prediction target vehicle is located, wherein the local map data at least comprises lane line information;

an encoding module that corresponds the position information of all vehicles and the local map data to each other based on the time stamp; acquiring position information and lane line information of the own vehicle, the track prediction target vehicle and other vehicles at the same moment based on the timestamp, and coding the position information of the vehicle at the moment according to a set rule to obtain a position information coding vector of the track prediction target vehicle; encoding the lane line information at the moment according to a set rule to obtain a lane line information encoding vector of the track prediction target vehicle;

the first neural network module and the second neural network module are used for inputting the position information coding vector and the lane line information coding vector into the first neural network module and training the first neural network module so as to enable the first neural network module to output a hidden vector representing a driving habit decision basis of a trajectory prediction target vehicle; and inputting the hidden vector and the historical track of the track prediction target vehicle into a second neural network module, and enabling the second neural network module to output the predicted track of the track prediction target vehicle.

Further, in the deep learning-based vehicle trajectory prediction system of the present invention, the driving data collection device at least includes a vehicle-mounted high-definition camera and/or a laser radar that are disposed on the vehicle.

Further, in the deep learning-based vehicle trajectory prediction system of the present invention:

the coding module codes the position information of the vehicle by adopting a mesh division method to obtain a position information coding vector with a fixed length; and/or

The encoding module determines the possible destination position or direction of the track prediction target vehicle according to the position of the track prediction target vehicle in the local map and the lane line information in the local map; obtaining corresponding lane line information closest to the track prediction target vehicle based on the possible destination position or direction; and obtaining the lane line information coding vector based on the lane line information closest to the track prediction target vehicle.

Further, in the deep learning based vehicle trajectory prediction system of the present invention, the first neural network module is an RNN, and/or the second neural network module is a fully-connected mapping neural network.

The vehicle track prediction method and the system based on deep learning have the following advantages that:

the method is mainly characterized in that the driving data (including the position information of the vehicle) around the track prediction target vehicle and the local map data (including lane line information) are coded in a regular coding mode, the driving habit decision idea of a driver is represented by hidden variables by combining a Recurrent Neural Network (RNN), and the future vehicle track of the target vehicle is predicted by combining the historical track of the track prediction target vehicle based on the hidden variables.

With respect to the currently applied vehicle trajectory prediction, the influence of these position quantities on the trajectory of the target vehicle is rather qualitative rather than quantitative depending on the trajectories of the surrounding vehicles and the lane marking points of the high-definition map. Different from the prior art, the technical scheme can more effectively extract the influence of the surrounding environment on the driving of the driver by using a rule-based coding form.

The driving habit decision-making thinking of the drivers is consistent, and the habit thinking of different drivers can be mapped into hidden vectors through a recurrent neural network, so that the representation can better predict the vehicle track.

Generally speaking, most of the track predictions of the vehicles are influenced by historical tracks, and the hidden vector representing the decision making idea finally and the historical tracks of the track prediction target vehicles are combined and then mapped onto the predicted tracks, so that the continuity of the historical tracks can be effectively ensured, and the influence of the driving habit decision making idea on the tracks can be more effectively highlighted in scenes such as turning roads and the like.

In the invention, the track prediction is carried out under different scenes based on the coding mode of the rule, the classification analysis of the result and the like are more intuitively and more carefully helped, the performance gap scene of the track prediction can be effectively positioned, and the updating or the optimization of the subsequent rule is more direct and more based.

Drawings

Fig. 1 schematically shows a method for coding a position information coding vector of a vehicle to obtain a trajectory prediction target in one embodiment of a deep learning-based vehicle trajectory prediction system according to the present invention.

Fig. 2 schematically shows a method for coding a lane line information coding vector of a vehicle with a track prediction target according to the deep learning-based vehicle track prediction system of the present invention.

Fig. 3 schematically shows a process of predicting the position information encoding vector and lane line information encoding vector of the target vehicle based on 20 frames of tracks and outputting corresponding last 30 frames of predicted tracks after neural network training processing.

Detailed Description

The vehicle trajectory prediction method and system of the present invention will be further explained and illustrated with reference to the drawings and the specific embodiments of the present disclosure, however, the explanation and illustration should not be construed as an undue limitation on the technical solutions of the present invention.

In the technical solution designed by the present invention, in order to effectively predict a vehicle trajectory, the inventor specifically designs a new vehicle trajectory prediction system based on deep learning, and the vehicle trajectory prediction system specifically includes: the device comprises a driving data acquisition device, a map data acquisition module, a coding module, a first neural network module and a second neural network module.

The specific content of the vehicle track prediction system designed by the invention for predicting the vehicle track specifically comprises the following steps (1) to (4):

(1) The method comprises the steps of collecting running data of all vehicles in an environment range by using a running data collecting device, wherein all vehicles in the environment range comprise 'own vehicles', 'track prediction target vehicles' and 'other vehicles', and the collected running data at least comprise position information of the vehicles.

(2) And acquiring local map data of all vehicles including positions of the track prediction target vehicles by using a map data acquisition module, wherein the local map data at least comprises lane line information.

(3) Employing an encoding module to correspond the obtained position information of all vehicles and local map data to each other based on the time stamp; acquiring the position information and lane line information of the 'own vehicle', 'track prediction target vehicle', 'other vehicles' at the same time based on the time stamp, and coding the position information of the vehicle at the time according to a set rule to obtain a position information coding vector of the track prediction target vehicle; and coding the lane line information at the moment according to a set rule to obtain a lane line information coding vector of the trajectory prediction target vehicle.

(4) Inputting the obtained position information coding vector and lane line information coding vector into a first neural network module, and training the first neural network module to enable the first neural network module to output a hidden vector representing a decision basis of a trajectory prediction target vehicle driving habit; the obtained hidden vector and the historical trajectory of the trajectory prediction target vehicle are input into a second neural network, and the predicted trajectory of the trajectory prediction target vehicle is output.

In the step (4), the first neural network module may be specifically selected as the RNN; the second neural network module can be specifically selected as a fully-connected mapping neural network; as for the historical trajectory of the "trajectory prediction target vehicle" input into the second neural network module, it may be obtained from the collected position information of the "trajectory prediction target vehicle" over a period of time.

It should be noted that, in the technical solution designed by the present invention, all vehicles within the collection environment range need to be specifically divided into "own vehicle", "trajectory prediction target vehicle", and "other vehicles". For the convenience of understanding, an example is that there are 9 automobiles in the collection environment range, where "own automobile" is one of the automobiles, and for other 8 automobiles, the trajectory prediction needs to be performed on that automobile, and the automobile is the "trajectory prediction target vehicle", and the other 7 automobiles are "other vehicles".

When the vehicle track prediction system based on deep learning designed by the invention is actually applied to predict the vehicle track:

in step (1) of the present invention, the driving data of the "own vehicle", "the trajectory prediction target vehicle", and "another vehicle" within the actual collection environment range of the test road needs to be collected by a vehicle-end environment sensing device provided on the "own vehicle". In the step (1) of the present invention, the driving data at least includes position information of the vehicle, and during actual acquisition, sensors such as a laser radar and a vehicle-mounted high definition camera carried by the "self vehicle" may be used as a vehicle-end environment sensing device to detect, identify, and track position track information of the surrounding vehicle during driving.

Meanwhile, in the step (2) of the invention, local map data of the position of the track prediction target vehicle can be acquired in a high-definition map drawn in advance by combining the global coordinate given by the positioning navigation system of the 'own vehicle'. The collected local map data may include information such as "lane line position", "turn", "intersection" and "controlled".

It should be noted that, in the present invention, different driving data need to be collected in different driving scenes, and it is ensured that the "trajectory prediction target vehicle" needs to have different numbers of "other vehicles" to drive in the process of extracting data, so as to ensure the dynamic change of the surrounding environment.

Often there are a plurality of target cars simultaneously in same collection environmental scope, and can generate multiunit training data simultaneously, often can have a target car orbit of longer period under a scene simultaneously, need to have the interval to a target car and have the data of different driving states, such as straight going turn lane change etc. to the multistage that produces as far as possible.

Accordingly, in this technical solution devised by the present invention, in the subsequent step (3), it is necessary to correspond the position information of all vehicles and the local map data to each other based on the time stamp because: the method adopted by the invention is to encode the position information of the vehicle of each frame and the local map data of the position of the track prediction target vehicle so as to ensure the accuracy of the surrounding environment at the same moment.

In step (3) of the present invention, the collected driving data may be processed specifically according to the following method to obtain training data which is subsequently input into the first neural network module RNN:

the method comprises the steps of combing the collected driving data of all vehicles in an environment range, extracting position information with time stamps corresponding to different vehicle IDs as criteria, and collecting 50 frames of data of the position information of each vehicle under a normal condition, wherein the first 20 frames are used as training data, and the last 30 frames of data are used as predicted targets. The invention can extract lane line position data and related lane description information data in a set range around the track prediction target vehicle through the vehicle positioning position specifically aiming at the acquired local map data of the position of the track prediction target vehicle.

Accordingly, in the encoding module, the position information and the lane line information of the "own vehicle", the "trajectory prediction target vehicle", and the "other vehicle" at the same time are acquired based on the time stamp, and the position information of the vehicle at this time can be encoded according to a rule set by a human. For example: the position information of the vehicle is encoded into a position information encoding vector of a fixed length by adopting a grid division mode, all vehicles in a corresponding range can be horizontally and vertically divided by taking the motion direction of a track prediction target vehicle as an axis, and weighting coefficients are set according to the distance so as to encode the vehicle. In the practical application, the behavior specification of a person during driving can be introduced by modifying and applying different coding rules, so that the coding effect on the surrounding environment information is improved, and the dynamic environment information around the track prediction target vehicle is better represented.

Similarly, the encoding module in the system of the present invention further needs to encode lane line information of the "trajectory prediction target vehicle" at the current time according to a set rule to obtain a lane line information encoding vector of the trajectory prediction target vehicle. For example: by setting the five lane changes possible for the vehicle: the "straight-ahead", "left lane change", "right lane change", "left turn", and "right turn" are used to encode the corresponding lane line information in the local map data of the position where the trajectory prediction target vehicle is currently located. In practical applications, specific encoded information often needs to include: the information of the position point, the direction, the intersection and the controlled or not of the lane line can also be added with curvature information representing the curve change of the lane line, and finally the lane line information is coded into a vector with fixed length. Of course, the coding rules can also be adjusted by introducing the driver's understanding of the different lanes of the structured road, in order to better adapt to the representation of the map information.

Therefore, in the encoding module, two sets of encoding vectors corresponding to the "trajectory prediction target vehicle", that is, the encoding vector of the position information and the encoding vector of the lane line information of the "trajectory prediction target vehicle", can be correspondingly generated based on the position information and the lane line information of each frame in the time stamp.

For easy understanding, in the present invention, the following encoding method shown in fig. 1 and 2 is specifically adopted, and a position information encoding vector and a lane line information encoding vector of the "trajectory prediction target vehicle" are obtained by encoding.

Fig. 1 schematically shows a method for coding a position information coding vector of a vehicle to obtain a trajectory prediction target vehicle according to a deep learning-based vehicle trajectory prediction system in an embodiment of the invention.

As shown in fig. 1, in the present embodiment, M1 denotes a "trajectory prediction target vehicle", which is a grid center, and a single grid specification is specifically set to be 5 meters × 5 meters; m2 denotes "own vehicle", and M3 denotes "other vehicle". Within the designed grid specification, as long as there is a vehicle, the grid code is marked as "1". Thus, an 8 × 8 position information encoding vector with a length of 64 can be obtained for each frame of position information. Similarly, the vehicle position may be encoded by other rules by dividing the range and introducing the encoding rule, and a weight may be introduced, which is not described in detail herein.

Fig. 2 schematically shows a method for coding a lane line information coding vector of a vehicle to obtain a trajectory prediction target vehicle according to a deep learning-based vehicle trajectory prediction system in an embodiment of the invention.

As shown in fig. 2, in the present embodiment, "five stars" indicate the positions of the "trajectory prediction target vehicles" in the local map, and L1 to L8 indicate lane lines in the current local map, and five possible destination positions, i.e., a to E, of the "trajectory prediction target vehicles" can be marked according to the traveling direction of the "trajectory prediction target vehicles", and the specific destination positions can be calculated according to the current positions in combination with the set angles and speeds.

And further searching a lane line closest to the five destination positions A, B, C, D and E, wherein the lane line is L3-L8-L8-L2-L4. The code of each lane line can be represented by the closest point (x, y) of the lane line, the secondary curve parameters (p 0, p1, p 2) of the lane line, the direction (dl, dr) of the lane line, whether the intersection (i) is present, and whether the intersection (c) is controlled, and the corner marks of these parameters respectively represent the corresponding parameters of the corresponding lane lines "L3", "L8", "L2", and "L4".

In the present invention, five sets of lane line codes are combined to obtain a complete lane line information code vector, which is here a lane line information code vector with a length of 45 with 5 × 9. Similarly, the adjustment of the selection strategy for the possible destination position of the "trajectory prediction target vehicle" and the adjustment of the representation manner of the lane line vector may be performed by vector coding based on different rules, which are not listed in detail herein.

Correspondingly, dividing a position information coding vector data set and a lane line information coding vector data set into a training set, a verification set and a test set, wherein the ratio of the position information coding vector data set to the lane line information coding vector data set is 8. The data distribution of the three is ensured to be consistent as much as possible. Wherein each set of data corresponds to a driving track of a target vehicle.

In step (4) of the present invention, the position information code vector and the lane line information code vector of the "trajectory prediction target vehicle" obtained for each frame are combined and input as training data to the first neural network RNN to train the vehicle.

It should be noted that, when the vehicle collects data of 50 frames, the first 20 frames are used as training data, and the last 30 frames are used as prediction targets. The first 20 frames of the "trajectory prediction target vehicle" are used as training data, so that a hidden vector representing a driving habit decision basis of the "trajectory prediction target vehicle" can be correspondingly output and updated, as shown in fig. 3.

Fig. 3 schematically shows a process of predicting the position information code vector and the lane line information code vector of the target vehicle based on the 20 frames of trajectories and outputting the corresponding next 30 frames of predicted trajectories after the neural network training process.

As shown in fig. 3, "vehicle track code" of each frame in "track prediction target vehicle" shown in fig. 3 may be understood as the aforementioned "position information code vector"; the "map data encoding" shown in fig. 3 may be understood as a corresponding "lane line information encoding vector".

It should be noted that, in fig. 3, the recurrent neural network RNN input to the fully-connected mapping network is a hidden vector after 20 iterations.

Inputting the hidden vector after 20 iterations and the historical track of the track prediction target vehicle into a fully-connected mapping network of a second neural network module, and outputting the predicted track of thirty frames in the future of the track prediction target vehicle.

It should be noted that the combination of the features in the present application is not limited to the combination described in the claims of the present application or the combination described in the specific examples, and all the features described in the present application may be freely combined or combined in any manner unless contradicted by each other.

It should also be noted that the above-listed embodiments are only specific embodiments of the present invention. It is apparent that the present invention is not limited to the above embodiments and similar changes or modifications can be easily made by those skilled in the art from the disclosure of the present invention and shall fall within the scope of the present invention.

Claims

1. A vehicle track prediction method based on deep learning is characterized by comprising the following steps:

inputting the position information coding vector and the lane line information coding vector into a first neural network, and training the first neural network so as to output a hidden vector representing a decision basis of a trajectory prediction target vehicle driving habit; and inputting the hidden vector and the historical track of the track prediction target vehicle into a second neural network, and enabling the second neural network to output the predicted track of the track prediction target vehicle.

2. The deep learning-based vehicle trajectory prediction method according to claim 1, characterized in that the driving data is collected by using a vehicle-end environment sensing device provided on a host vehicle.

3. The deep learning-based vehicle trajectory prediction method according to claim 2, characterized in that the vehicle-end environment sensing device at least comprises a vehicle-mounted high-definition camera and/or a laser radar.

4. The deep learning-based vehicle track prediction method according to claim 1, characterized in that the position information of the vehicle is coded by using a mesh division method to obtain a position information coding vector with a fixed length.

5. The deep learning-based vehicle trajectory prediction method according to claim 1, characterized in that a possible destination position or direction of the trajectory prediction target vehicle is determined based on a position of the trajectory prediction target vehicle in the local map and lane line information in the local map; obtaining corresponding lane line information closest to the track prediction target vehicle based on the possible destination position or direction; and obtaining the lane line information coding vector based on the lane line information closest to the track prediction target vehicle.

6. The deep learning-based vehicle trajectory prediction method according to claim 1, characterized in that the first neural network is an RNN and/or the second neural network is a fully-connected mapped neural network.

7. A vehicle trajectory prediction system based on deep learning, characterized in that it comprises:

an encoding module that corresponds the position information of all vehicles and the local map data to each other based on the time stamp; acquiring position information and lane line information of the own vehicle, the track prediction target vehicle and other vehicles at the same moment based on the timestamp, and coding the position information of the vehicle at the moment according to a set rule to obtain a position information coding vector of the track prediction target vehicle; coding the lane line information at the moment according to a set rule to obtain a lane line information coding vector of the track prediction target vehicle;

8. The deep learning-based vehicle trajectory prediction system of claim 7, wherein the travel data acquisition device comprises at least an on-board high-definition camera and/or a lidar disposed on the host vehicle.

9. The deep learning-based vehicle trajectory prediction system of claim 7, wherein:

The encoding module determines the possible destination position or direction of the track prediction target vehicle according to the position of the track prediction target vehicle in the local map and lane line information in the local map; obtaining corresponding lane line information closest to the track prediction target vehicle based on the possible destination position or direction; and obtaining the lane line information coding vector based on the lane line information closest to the track prediction target vehicle.

10. The deep learning-based vehicle trajectory prediction system of claim 7, wherein the first neural network module is an RNN and/or the second neural network module is a fully-connected mapped neural network.