A vehicle trajectory prediction method and system based on deep learning

Abstract

The application discloses a vehicle trajectory prediction method based on deep learning, which comprises the following steps: collecting driving data of all vehicles in an environment range; obtaining local map data of a position where a trajectory prediction target vehicle is located; corresponding position information of all vehicles and the local map data to each other based on timestamps; obtaining position information and lane line information of a host vehicle, the trajectory prediction target vehicle and other vehicles at the same time based on the timestamps, 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 trajectory prediction target vehicle; inputting the position information encoding vector and the lane line information encoding vector into a first neural network for training, so that a hidden vector representing driving habit decision basis of the trajectory prediction target vehicle is output; and inputting the hidden vector and a historical trajectory of the trajectory prediction target vehicle into a second neural network, so that a predicted trajectory of the trajectory prediction target vehicle is output.

Description

TECHNICAL FIELD

[0001] The present application relates to a trajectory prediction method and system, in particular to a vehicle trajectory prediction method and system. BACKGROUND

[0002] With the development of intelligent driving technology, trajectory prediction of intelligent vehicles is becoming more and more important.

[0003] In the current prior art, the vehicle trajectory prediction method currently adopted is mostly based on the historical state of the vehicle itself such as position and speed, and the surrounding environment is detected and analyzed, which detects and identifies the dynamic information around the vehicle such as vehicle position, pedestrian position, speed information, and combines the road information contained in the high-definition map such as lane line position, lane line direction and signal light in the scene information where the vehicle itself is positioned, thereby describing a state of the target vehicle to predict the vehicle trajectory.

[0004] However, the vehicle trajectory prediction in the prior art is based on the trajectory of the surrounding vehicles and the lane line points of the high-definition map, and the position quantity actually has a qualitative rather than quantitative influence on the trajectory of the target vehicle. Therefore, the prediction accuracy is not high.

[0005] Therefore, it is desirable to obtain a method for effectively improving the accuracy of vehicle trajectory prediction. SUMMARY

[0006] One of the purposes of the present application is to provide a vehicle trajectory prediction method based on deep learning, which can encode the driving data and local map data around the target vehicle for trajectory prediction based on the rule coding mode, and combine the neural network to represent the driving habit decision-making idea of the driver through the hidden variable, and finally predict the future vehicle trajectory of the target vehicle based on the hidden variable and the historical trajectory of the target vehicle.

[0007] In order to achieve the above purpose, the present application provides a vehicle trajectory prediction method based on deep learning, which comprises the following steps:

[0008] Collect driving data of all vehicles in the environment, the vehicles including the ego vehicle, the target vehicle for trajectory prediction and other vehicles, and the driving data at least including position information of the vehicles;

[0009] Obtain local map data of the position where the target vehicle for trajectory prediction is located, the local map data at least including lane line information;

[0010] Correspond the position information and the local map data of all vehicles to each other based on the time stamp;

[0011] The position information and lane line information of the ego vehicle, the trajectory prediction target vehicle and other vehicles at the same time are acquired based on a timestamp, the position information of the vehicle at the time is encoded according to a set rule, so as to obtain a position information encoding vector of the trajectory prediction target vehicle; and the lane line information at the time is encoded according to a set rule, so as to obtain a lane line information encoding vector of the trajectory prediction target vehicle.

[0012] The position information encoding vector and the lane line information encoding vector are input into a first neural network, and the first neural network is trained, so that a hidden vector representing driving habit decision basis of the trajectory prediction target vehicle is output; and the hidden vector and the historical trajectory of the trajectory prediction target vehicle are input into a second neural network, so that a predicted trajectory of the trajectory prediction target vehicle is output.

[0013] Research finds that the image encoding method can fuse all relevant information into a picture, but subsequent convolutional neural networks are needed to extract features from the picture, and similar position and speed information are relatively refined and independent features, and more packaging and extraction increase complexity. The recurrent neural network (RNN) can encode historical information into a hidden vector, but it is difficult to fuse different vehicle and pedestrian dynamic information together. These limit the more intuitive and comprehensive representation of the current environment state of the vehicle.

[0014] Therefore, the present application designs a new vehicle trajectory prediction method based on the above situation analysis, which can encode the driving data (including the position information of the vehicle) and the local map data (including the lane line information) around the trajectory prediction target vehicle based on the rule encoding method, and combine the recurrent neural network (RNN) to represent the driving habit decision-making idea of the driver through a hidden variable, and finally predict the future vehicle trajectory of the trajectory prediction target vehicle based on the hidden variable and the historical trajectory of the trajectory prediction target vehicle.

[0015] In the present application, the position information in the driving data of all vehicles in the environment range of the trajectory prediction target vehicle needs to be encoded according to a set rule. This encoding method based on the set rule can facilitate subsequent targeted analysis of the model while maintaining simplicity.

[0016] In the present application, the influence of different road environments on vehicle trajectory decision is added by introducing a high-definition map, which is different from the previous map attribute focusing on the lane line center point position. The present application focuses on the influence of the curvature factor direction of the lane on the trajectory based on the lane line information encoding in the local map data.

[0017] In addition, the present application can fuse the current environment state with the historical trajectory of the trajectory prediction target vehicle by introducing a recurrent neural network to obtain a more unique trajectory prediction.

[0018] It should be noted that in the technical solution designed by the present application, all vehicles in the collected environment range need to be specifically divided into "ego vehicle", "trajectory prediction target vehicle" and "other vehicles". In order to facilitate understanding, taking an example of 9 cars in the collected environment range, one of which is the "ego vehicle", and for the other 8 cars, one needs to be trajectory predicted, which is the "trajectory prediction target vehicle", and the remaining 7 cars are "other vehicles".

[0019] In the actual application of the vehicle trajectory prediction system based on deep learning designed by the present application to predict the vehicle trajectory, the driving data of the "ego vehicle", "trajectory prediction target vehicle" and "other vehicles" need to be collected on the test road through the sensors of the "ego vehicle". Among them, the driving data at least includes the position information of the vehicle, and in the actual collection, the position trajectory information of the surrounding vehicles during driving can be detected and recognized by the sensors such as laser radar and camera carried by the "ego vehicle". At the same time, combined with the global coordinates given by the positioning and navigation system of the "ego vehicle", the local map data of the position where the "trajectory prediction target vehicle" is located can be obtained on the high-definition map drawn in advance. Among them, the collected local map data can contain corresponding "lane line position", "turning", "whether it is an intersection", "whether it is controlled" and other information.

[0020] In addition, in the vehicle trajectory prediction method designed by the present application, the historical trajectory of the trajectory prediction target vehicle input into the second neural network can be obtained according to the position information of the trajectory prediction target vehicle collected within a period of time.

[0021] Further, in the vehicle trajectory prediction method based on deep learning described in the present application, the driving data is collected by the vehicle-end environment perception device arranged on the ego vehicle.

[0022] Further, in the vehicle trajectory prediction method based on deep learning described in the present application, the vehicle-end environment perception device at least includes a vehicle-mounted high-definition camera and / or a laser radar.

[0023] Further, in the vehicle trajectory prediction method based on deep learning described in the present application, the position information of the vehicle is encoded by using a grid division method to obtain a position information encoding vector with fixed length.

[0024] Further, in the deep learning-based vehicle trajectory prediction method, the 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; the corresponding lane line information closest to the trajectory prediction target vehicle is obtained based on the possible destination position or direction; and the lane line information encoding vector is obtained based on the lane line information closest to the trajectory prediction target vehicle.

[0025] Further, in the deep learning-based vehicle trajectory prediction method, the first neural network is an RNN, and / or the second neural network is a fully connected mapping neural network.

[0026] Correspondingly, another object of the present application is to provide a deep learning-based vehicle trajectory prediction system which can effectively implement the vehicle trajectory prediction method described above to predict the trajectory of a target vehicle.

[0027] To achieve the above object, the present application provides a deep learning-based vehicle trajectory prediction system, which comprises:

[0028] A driving data acquisition device acquires driving data of all vehicles in the environment, including the ego vehicle, the trajectory prediction target vehicle and other vehicles, and the driving data at least includes position information of the vehicles;

[0029] A map data acquisition module acquires local map data of the position where the trajectory prediction target vehicle is located, and the local map data at least includes lane line information;

[0030] An encoding module corresponds the position information of all vehicles and the local map data to each other based on the time stamp; acquires the position information and the lane line information of the ego vehicle, the trajectory prediction target vehicle and other vehicles at the same time based on the time stamp; encodes the position information of the vehicles at the time according to the set rules to obtain the position information encoding vector of the trajectory prediction target vehicle; and encodes the lane line information at the time according to the set rules to obtain the lane line information encoding vector of the trajectory prediction target vehicle;

[0031] A first neural network module and a second neural network module, wherein the position information encoding vector and the lane line information encoding vector are input into the first neural network module, which is trained to output a hidden vector representing the driving habit decision basis of the trajectory prediction target vehicle; and the hidden vector and the historical trajectory of the trajectory prediction target vehicle are input into the second neural network module, which outputs the predicted trajectory of the trajectory prediction target vehicle.

[0032] Further, in the deep learning-based vehicle trajectory prediction system, the driving data acquisition device comprises at least a high-definition camera and / or a laser radar arranged on the ego vehicle.

[0033] Further, in the deep learning-based vehicle trajectory prediction system,

[0034] The encoding module encodes the position information of the vehicle by using a grid division method to obtain a position information encoding vector with a fixed length.

[0035] The encoding module determines a possible destination position or direction of the trajectory prediction target vehicle according to the position of the trajectory prediction target vehicle in the local map and the lane line information in the local map, obtains corresponding lane line information closest to the trajectory prediction target vehicle based on the possible destination position or direction, and obtains the lane line information encoding vector based on the lane line information closest to the trajectory prediction target vehicle.

[0036] Further, in the deep learning-based vehicle trajectory prediction system, the first neural network module is an RNN, and / or the second neural network module is a fully connected mapping neural network.

[0037] The deep learning-based vehicle trajectory prediction method and system have the following advantages:

[0038] The most important innovation of the present application is that 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) are encoded by using a rule-based encoding method, and the driving habit decision-making thought of the driver is represented by a hidden variable through a recurrent neural network (RNN), and finally the future vehicle trajectory of the trajectory prediction target vehicle is predicted based on the hidden variable and the historical trajectory of the trajectory prediction target vehicle.

[0039] Compared with the current vehicle trajectory prediction based on the trajectories of surrounding vehicles and the lane line points of high-definition maps, the influence of these position quantities on the trajectory of the target vehicle is actually qualitative rather than quantitative. Unlike the prior art, the present technical solution uses a rule-based encoding form to more effectively extract the influence of the surrounding environment on the driver's driving.

[0040] The driving habit decision-making thought of the driver is consistent, and the habit thought of different drivers can be mapped to a hidden vector through a recurrent neural network, so as to better predict the trajectory of the vehicle.

[0041] Generally speaking, the trajectory prediction of the vehicle is mostly affected by the historical trajectory. The combination of the hidden vector finally representing the decision-making idea and the historical trajectory of the trajectory prediction target vehicle and the re-mapping to the predicted trajectory can effectively ensure the continuity of the historical trajectory and can more effectively highlight the influence of the driving habit decision-making idea on the trajectory in scenes such as turning at the intersection.

[0042] In the present application, the rule-based encoding mode is used for trajectory prediction in different scenarios, which has more intuitive and more detailed help for the classification analysis of the results, and can effectively locate the performance gap scene of trajectory prediction. The subsequent rule updating or optimization is more direct and has more basis. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 The method for encoding the position information encoding vector of the trajectory prediction target vehicle by the deep learning-based vehicle trajectory prediction system according to the present application is schematically shown in an embodiment.

[0044] Figure 2 The method for encoding the lane line information encoding vector of the trajectory prediction target vehicle by the deep learning-based vehicle trajectory prediction system according to the present application is schematically shown in an embodiment.

[0045] Figure 3 The process of outputting the corresponding 30-frame predicted trajectory after the neural network training and processing of the 20-frame position information encoding vector and lane line information encoding vector of the trajectory prediction target vehicle is schematically shown. DETAILED DESCRIPTION

[0046] The vehicle trajectory prediction method and system according to the present application will be further explained and described below in combination with the drawings and specific embodiments, but the explanation and description do not constitute undue limitations on the technical solutions of the present application.

[0047] In the technical solution designed in the present application, in order to effectively predict the vehicle driving trajectory, the inventors specifically design a new deep learning-based vehicle trajectory prediction system, which specifically includes a driving data acquisition device, a map data acquisition module, an encoding module, a first neural network module, and a second neural network module.

[0048] The specific content of predicting the vehicle trajectory by the vehicle trajectory prediction system designed in the present application specifically includes the following steps (1)-(4):

[0049] (1) Use the driving data acquisition device to acquire the driving data of all vehicles in the environment, all vehicles in the environment including the "self vehicle", the "trajectory prediction target vehicle", and the "other vehicles", and the acquired driving data at least including the position information of the vehicle.

[0050] (2) Utilize the map data acquisition module to obtain the local map data of the positions of all vehicles, including the "trajectory prediction target vehicle", and the local map data at least includes lane line information.

[0051] (3) Adopt the encoding module to correspond the obtained position information and local map data of all vehicles to each other based on the timestamp; obtain the position information and lane line information of the "ego vehicle", "trajectory prediction target vehicle" and "other vehicles" at the same time based on the timestamp, encode the position information of the vehicles at the time according to the set rules to obtain the position information encoding vector of the trajectory prediction target vehicle; encode the lane line information at the time according to the set rules to obtain the lane line information encoding vector of the trajectory prediction target vehicle.

[0052] (4) Input the obtained position information encoding vector and lane line information encoding vector into the first neural network module to train it, so that it outputs the hidden vector representing the driving habit decision basis of the trajectory prediction target vehicle; input the obtained hidden vector and the historical trajectory of the trajectory prediction target vehicle into the second neural network to make it output the predicted trajectory of the trajectory prediction target vehicle.

[0053] In the above step (4), the first neural network module can be specifically selected as RNN; the second neural network module can be specifically selected as a fully connected mapping neural network; and the historical trajectory of the "trajectory prediction target vehicle" input into the second neural network module can be obtained according to the position information of the "trajectory prediction target vehicle" collected within a period of time.

[0054] It should be noted that in the technical solution designed in the present application, all vehicles within the collection environment range need to be specifically divided into "ego vehicle", "trajectory prediction target vehicle" and "other vehicles". In order to facilitate understanding, taking 9 cars within the collection environment range as an example, the "ego vehicle" is one of them, and for the other 8 cars, one of them needs to be predicted, which is the "trajectory prediction target vehicle", and the remaining 7 cars are "other vehicles".

[0055] In the actual application of the vehicle trajectory prediction system based on deep learning designed in the present application to predict the vehicle trajectory:

[0056] In step (1) of the present application, the driving data of the "ego vehicle", "trajectory prediction target vehicle" and "other vehicles" within the actual collection environment range of the test road needs to be collected by the vehicle-end environment perception device provided on the "ego vehicle". In step (1) of the present application, the driving data at least includes the position information of the vehicle, and when actually collecting, the position trajectory information of the surrounding vehicles can be detected and tracked by the laser radar, vehicle-mounted high-definition camera and other sensors carried by the "ego vehicle" as the vehicle-end environment perception device.

[0057] Meanwhile, in step (2) of the present application, combined with the global coordinates given by the positioning and navigation system of the "ego vehicle", the local map data of the position where the "trajectory prediction target vehicle" is located can be obtained in the high-definition map drawn in advance. The collected local map data can include the corresponding "lane line position", "turning", "whether it is an intersection", "whether it is controlled" and other information.

[0058] It should be noted that in the present application, different driving data needs to be collected in different driving scenarios, and it is ensured that the "trajectory prediction target vehicle" needs to have different numbers of "other vehicles" driving in the data extraction process to ensure the dynamic change of the surrounding environment.

[0059] In the same collection environment range, there are often multiple target vehicles, and multiple sets of training data can be generated at the same time, and there is often a long period of target vehicle trajectory in one scene, and it is necessary to generate as much data as possible for one target vehicle, such as straight turning and lane changing, which has intervals but different driving states.

[0060] Correspondingly, in the technical solution designed in the present application, in subsequent step (3), the position information of all vehicles and the local map data need to be corresponded to each other based on the time stamp, because the method adopted in the present application is to encode the position information of each frame of vehicle and the local map data of the position where the "trajectory prediction target vehicle" is located, to ensure the accuracy of the surrounding environment at the same time.

[0061] In step (3) of the present application, the collected driving data can be processed according to the following method to obtain the training data input into the first neural network module RNN:

[0062] The driving data of all vehicles in the environment range are combed, and the position information corresponding to the time stamp of different vehicle IDs is extracted. In general, 50 frames of data are collected for the position information of each vehicle, of which the first 20 frames are used as training data, and the last 30 frames are used as prediction targets. For the local map data of the position of the "trajectory prediction target vehicle" obtained, the vehicle positioning position is used to extract the lane line position data and related lane description information data within a set range around the "trajectory prediction target vehicle".

[0063] Correspondingly, in the encoding module, the position information and lane line information of the "ego vehicle", "trajectory prediction target vehicle" and "other vehicles" at the same time are obtained based on the time stamp, and the position information of the vehicles at this time can be encoded according to the rules set by humans. For example, the position information of the vehicle is encoded into a fixed-length position information encoding vector in the form of grid division, or the motion direction of the "trajectory prediction target vehicle" is used as the axis to horizontally and vertically divide all vehicles within the corresponding range, and the vehicles are encoded according to the distance to set a weighting coefficient. In actual application, the behavior norms of human driving can be introduced by modifying different encoding rules, so as to improve the encoding effect of the surrounding environment information, thereby better representing the dynamic environment information around the "trajectory prediction target vehicle".

[0064] Similarly, the encoding module in the system also needs to encode the lane line information of the "trajectory prediction target vehicle" at the current time according to the set rules to obtain the lane line information encoding vector of the "trajectory prediction target vehicle". For example, the corresponding lane line information in the local map data of the position of the "trajectory prediction target vehicle" is encoded by setting five possible lane changing actions of the vehicle: "straight", "left lane change", "right lane change", "left turn" and "right turn". In actual application, the specific encoding information often needs to include: the position point of the lane line, the direction of the lane line, the intersection and whether it is controlled, and of course, the curvature information representing the curve change of the lane line can also be added. Finally, the lane line information is encoded into a fixed-length vector. Of course, the encoding rules can also be adjusted by introducing the understanding of the driver for the different lanes of the structured road, so as to better adapt to the representation of the map information.

[0065] Therefore, in the encoding module, based on the position information and lane line information of each frame in the time stamp, two groups of encoding vectors corresponding to the "trajectory prediction target vehicle" can be generated, i.e. the position information encoding vector and the lane line information encoding vector of the "trajectory prediction target vehicle".

[0066] In order to facilitate understanding, the following Figure 1 andFigure 2 The encoding method shown encodes the position information encoding vector and lane line information encoding vector of the "trajectory prediction target vehicle".

[0067] Figure 1 The figure schematically shows a method for encoding a position information encoding vector of a target vehicle for trajectory prediction in a deep learning-based vehicle trajectory prediction system according to the present invention in one embodiment.

[0068] like Figure 1 As shown, in this embodiment, M1 represents the "target vehicle for trajectory prediction," with it as the center of the grid, and the grid size is set to 5 meters x 5 meters. M2 represents the "own vehicle," and M3 represents "other vehicles." Within the designed grid size range, any grid containing a vehicle is marked with a code of "1." Thus, for each frame of position information, an 8x8 position information encoding vector of length 64 is generated. Similarly, by dividing the range and introducing encoding rules, vehicle positions can be encoded using other rules, and weights can also be introduced, which are not detailed here.

[0069] Figure 2 The figure schematically shows a method for encoding a lane line information encoding vector of a trajectory prediction target vehicle in a deep learning-based vehicle trajectory prediction system according to the present invention in one embodiment.

[0070] like Figure 2 As shown, in this embodiment, the "five-pointed star" represents the position of the "trajectory prediction target vehicle" in the local map, L1-L8 represent the lane lines in the current local map, and according to the driving direction of the "trajectory prediction target vehicle", five possible destination positions of the "trajectory prediction target vehicle" can be marked, namely AE, and its specific destination position can be calculated based on the current position combined with the set angle and speed.

[0071] For the five destinations A, B, C, D, and E, the nearest lanes can be further searched: L3, L8, L8, L2, and L4, respectively. Each lane can be coded using its nearest point (x, y), lane quadratic curve parameters (p0, p1, p2), lane direction (dl, dr), whether it is at an intersection (i), and whether it is controlled (c). The subscripts of these parameters are "3," "8," "8," "2," and "4," representing the corresponding parameters for lanes "L3," "L8," "L8," "L2," and "L4," respectively.

[0072] In the present application, five sets of lane line encoding groups are combined to obtain complete lane line information encoding vectors, which are 5*9 lane line information encoding vectors with a length of 45. Similarly, the selection strategy adjustment of the "trajectory prediction target vehicle" possible destination position and the adjustment of the lane line vector representation method can be based on different rules for vector encoding, which will not be listed in detail here.

[0073] Accordingly, the position information encoding vector and the lane line information encoding vector dataset are divided into a training set, a validation set and a test set, and the ratio of the three is 8:1:1. The data distribution of the three is as consistent as possible. Each set of data should correspond to the driving trajectory of a target vehicle.

[0074] In step (4) of the present application, based on the obtained position information encoding vector and lane line information encoding vector of the "trajectory prediction target vehicle" of each frame, the two are combined to serve as training data input into the first neural network RNN for training.

[0075] It should be noted that when the vehicle collects 50 frames of data, the first 20 frames are used as training data, and the last 30 frames of data are used as the prediction target. The first 20 frames of the "trajectory prediction target vehicle" are used as training data, which can correspond to the output update of the hidden vector representing the driving habit decision basis of the "trajectory prediction target vehicle", as shown in Figure 3 .

[0076] Figure 3 The process of outputting the corresponding 30 frames of predicted trajectory after the 20 frames of trajectory prediction target vehicle position information encoding vector and lane line information encoding vector are processed by the neural network is schematically shown.

[0077] As shown in Figure 3 , the "vehicle trajectory encoding" of each frame of the "trajectory prediction target vehicle" shown in Figure 3 can be understood as the aforementioned "position information encoding vector"; Figure 3 The "map data encoding" shown in can be understood as the corresponding "lane line information encoding vector".

[0078] Figure 3 It should be noted that in , the recurrent neural network RNN input into the fully connected mapping network is the hidden vector after 20 iterations.

[0079] The hidden vector after 20 iterations and the historical trajectory of the trajectory prediction target vehicle are input into the fully connected mapping network of the second neural network module, which can output the predicted trajectory of the "trajectory prediction target vehicle" for the next thirty frames.

[0080] It should be noted that the combination of the technical features in the case is not limited to the combination of the claims in the case or the combination of the embodiments in the case. All the technical features disclosed in the case can be freely combined or combined in any way, unless contradictory.

[0081] It should also be noted that the above-mentioned embodiments are only specific embodiments of the present application. Obviously, the present application is not limited to the above-mentioned embodiments, and similar changes or modifications made by those skilled in the art from the disclosure of the present application are directly derived or easily conceived, and should belong to the protection scope of the present application.

Claims

1. A vehicle trajectory prediction method based on deep learning, characterized in that: Including steps: Collecting driving data of all vehicles within the environment, including the ego vehicle, trajectory prediction target vehicles, and other vehicles, wherein the driving data includes at least the location information of the vehicles; Acquire local map data of a location of a trajectory prediction target vehicle, wherein the local map data includes at least lane line information; The location information of all vehicles and the local map data are matched to each other based on the timestamp; The position information and lane line information of the ego vehicle, the trajectory prediction target vehicle, and other vehicles at the same time are obtained based on the timestamp. The position information of the vehicle at that time is encoded according to the set rules to obtain the position information encoding vector of the trajectory prediction target vehicle. The lane line information at that time is encoded according to the set rules to obtain the lane line information encoding vector of the trajectory prediction target vehicle. The position information encoding vector and the lane line information encoding vector are input into a first neural network, which is trained so that it outputs a hidden vector representing the basis for decision-making on the driving habits of the trajectory prediction target vehicle; the hidden vector and the historical trajectory of the trajectory prediction target vehicle are input into a second neural network, which outputs the predicted trajectory of the trajectory prediction target vehicle.

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

3. The vehicle trajectory prediction method based on deep learning according to claim 2, characterized in that: The vehicle-side environment perception device includes at least a vehicle-mounted high-definition camera and / or a laser radar.

4. The vehicle trajectory prediction method based on deep learning according to claim 1, characterized in that: The position information of the vehicle is encoded using a grid division method to obtain a position information encoding vector with a fixed length.

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

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

7. A vehicle trajectory prediction system based on deep learning, characterized in that: It includes: A driving data collection device collects driving data of all vehicles within the environment, including the ego vehicle, trajectory prediction target vehicles, and other vehicles, and the driving data includes at least vehicle location information; A map data acquisition module, which acquires local map data of the location of the trajectory prediction target vehicle, wherein the local map data at least includes lane line information; an encoding module that associates the position information of all vehicles and the local map data with each other based on a timestamp; The position information and lane line information of the ego vehicle, the trajectory prediction target vehicle, and other vehicles at the same time are obtained based on the timestamp. The position information of the vehicle at that time is encoded according to the set rules to obtain the position information encoding vector of the trajectory prediction target vehicle. The lane line information at that time is encoded according to the set rules to obtain the lane line information encoding vector of the trajectory prediction target vehicle. A first neural network module and a second neural network module, wherein the position information encoding vector and the lane line information encoding vector are input into the first neural network module and trained so that it outputs a hidden vector representing the basis for decision-making on the driving habits of the trajectory prediction target vehicle; the hidden vector and the historical trajectory of the trajectory prediction target vehicle are input into the second neural network module so that it outputs the predicted trajectory of the trajectory prediction target vehicle.

8. The vehicle trajectory prediction system based on deep learning according to claim 7, characterized in that: The driving data acquisition device at least includes a vehicle-mounted high-definition camera and / or a laser radar installed on the vehicle.

9. The vehicle trajectory prediction system based on deep learning according to claim 7, characterized in that: The encoding module encodes the vehicle's position information using a grid division method to obtain a position information encoding vector with a fixed length; and / or The encoding module determines the possible destination position or direction of the trajectory prediction target vehicle based on the position of the trajectory prediction target vehicle in the local map and the lane line information in the local map; obtains the corresponding lane line information closest to the trajectory prediction target vehicle based on the possible destination position or direction; and obtains the lane line information encoding vector based on the lane line information closest to the trajectory prediction target vehicle.

10. The vehicle trajectory prediction system based on deep learning according to claim 7, characterized in that: The first neural network module is an RNN, and / or the second neural network module is a fully connected mapping neural network.

Images