Segmented vehicle trajectory prediction method based on training
By employing a segmented vehicle trajectory prediction method, utilizing intermediate acceleration and steering wheel angle constraints, and combining high-dimensional features of driving habits, the inconsistency problem of existing vehicle trajectory prediction methods when driving behavior intentions change is solved, achieving higher prediction accuracy and adaptability.
Patent Information
- Application Number
- CN202411332933.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-24
- Publication Date
- 2025-01-07
AI Technical Summary
Existing vehicle trajectory prediction methods struggle to accurately predict future trajectories when faced with changes in driving intent, especially within time intervals of about half a second. They suffer from significant black-box uncertainty and inconsistent predictions.
A segmented vehicle trajectory prediction method is adopted, which constrains the model mapping by predicting intermediate acceleration and steering wheel angle, and introduces high-dimensional features as a representation of driving habits during the encoding process to measure the vehicle's driving behavior, and ensures the consistency of predicted trajectories during training.
It achieves better trajectory smoothness and consistency, can better adapt to driving habits in different driving scenarios, provides constraints on intermediate acceleration and turning angle, and improves the accuracy and adaptability of prediction.
Smart Images
Figure CN119272114A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a vehicle trajectory prediction method based on training, and in particular to a segmented vehicle trajectory prediction method based on training. Background Art
[0002] The current vehicle trajectory prediction method mainly relies on the historical data of the vehicle's own position, speed, etc., while detecting and analyzing the surrounding environment. The state of the target vehicle is described by detecting the surrounding dynamic information (such as the position and speed of vehicles and pedestrians) and combining the road information contained in the high-definition map where the vehicle is located (such as lane line position, direction, traffic lights and other scene information). Then, this state is encoded and mapped to the future movement trajectory of the vehicle. The training method is to construct an encoding module and a mapping module, and train the model parameters on a large amount of training data to obtain the final module parameters, thereby building a training-based vehicle trajectory prediction model.
[0003] The encoding of dynamic information can be achieved in the form of images, where the historical positions and speeds of vehicles and pedestrians are presented in the images through shapes and colors, and then the images are feature extracted using convolutional neural networks (CNNs). Another method is to use recurrent neural networks (RNNs) to encode these dynamic information into the hidden variables of the module to extract the features of the dynamic information. Finally, the extracted features are mapped to the predicted trajectory through a fully connected layer.
[0004] At present, most trajectory prediction methods directly encode dynamic information in an end-to-end manner and map it to the final predicted future trajectory points. Directly encoding dynamic information and mapping it to the final predicted future trajectory points is relatively black-box fog, and most methods use a one-time prediction of all future trajectory points in this process. This is difficult to sustain in practical applications when faced with a common time interval of about half a second, due to the impact of changes in driving behavior intentions in the middle. Summary of the invention
[0005] The object of the present invention is to provide a segmented vehicle trajectory prediction method based on training, which constrains the mapping of the model by predicting intermediate acceleration and steering wheel angle, and introduces a high-dimensional feature as a representation of driving habits in the encoding process to measure the driving behavior of the vehicle and to constrain the trajectory position mapping at the next prediction time point, and ensures the consistency of the driving habit representation of the same vehicle in the sample during the training process.
[0006] The present invention provides a training-based segmented vehicle trajectory prediction method, comprising:
[0007] S1) data collection, wherein the data is vehicle driving data collected on a test road;
[0008] S2) data preparation, dividing the data set into a training set, a validation set and a test set, the data to be prepared include: vehicle position data based on timestamps corresponding to different vehicle IDs, map data, vehicle dynamic environment data, driving scene data, acceleration and true value (GT) values of steering wheel angle;
[0009] S3) model construction, the model includes a vehicle environment dynamic encoding module, a driving habit high-dimensional representation module, a historical trajectory alignment module, a prediction mapping module and a model optimization module, and the model is used for data training;
[0010] S4) Model optimization: during the training process, the model optimization module is used to calculate the error between the predicted trajectory position, acceleration, steering wheel angle, and driving habit characterization vector and the true value, and the model parameters are optimized by the back propagation algorithm;
[0011] S5) Use the model to make predictions; Use the trained model to make segmented predictions on the future trajectory of the target vehicle, constrain the predicted trajectory by predicting the intermediate acceleration and steering wheel angle, and introduce a high-dimensional representation vector of driving habits to ensure the consistency of the predicted trajectory.
[0012] The present invention also includes a coding model for training high-dimensional representation of driving habits. The method encodes the dynamic environment, historical trajectory and actual position at the next moment into a high-dimensional representation vector of driving habits through a model of a coding and decoding structure. The position at the next moment is predicted in combination with the historical trajectory. A prediction scenario is output at the decoder end. An average feature vector output by the model after the full amount of data is used as an initialization vector in the actual reasoning in the vehicle trajectory prediction model.
[0013] The method described in step S1) detects and identifies the position trajectory of surrounding vehicles during driving through sensors carried by the vehicle, and combines the global coordinates and high-definition map provided by the vehicle's positioning and navigation system to obtain a local map where the target vehicle is located; the method simultaneously generates multiple sets of training data, and generates multiple segments of data with intervals but different driving states for a target vehicle; and aligns the timestamps of the vehicle driving data in step s1).
[0014] The ratio of the training set, validation set and test set described in step S2) is 8:1:1.
[0015] Step S2) The division of these driving scenarios is based on road traffic rules, driving behavior characteristics and road design features, and the driving scenarios are used to train the high-dimensional representation vector of the driving habits.
[0016] The vehicle environment dynamic encoding module described in step S3) is used to encode the surrounding environment information of the target vehicle, the historical positions of other vehicles, and the surrounding lane line information into a vector form. The driving habit high-dimensional representation module is used to initialize and iteratively update the driving habit representation vector of a specific vehicle. The historical trajectory alignment module is used to align the historical trajectories of the past period of time into a one-dimensional vector. The three groups of inputs are mapped and aligned separately in the prediction mapping module in the fully connected mapping network, and then spliced together to make the final prediction position mapping.
[0017] The prediction mapping module described in step S3) will also output two intermediate quantities, acceleration and steering wheel angle, during the alignment mapping process, which are used for constraints and analysis and judgment during the model training process.
[0018] Step S3) The model also includes a module for calculating the driving habit consistency loss, which evaluates the matching degree between the driving habit representation vector and the current driving scene based on predefined driving scene rules.
[0019] Step S4) The GT values of the acceleration and steering wheel angle are obtained by formula calculation through the vehicle historical trajectory and the GT trajectory.
[0020] Step S4) also includes a fully connected head for updating the high-dimensional representation vector of driving habits. BRIEF DESCRIPTION OF THE DRAWINGS
[0021] Figure 1 A flow chart of a segmented vehicle trajectory prediction method based on training preferred in an embodiment of the present invention;
[0022] Figure 2 A schematic diagram of a preferred training model for an embodiment of the present invention;
[0023] Figure 3 A schematic diagram of a coding model for training high-dimensional representation of driving habits preferred in an embodiment of the present invention;
[0024] Figure 4 A vehicle environment dynamic coding data diagram preferably used in an embodiment of the present invention;
[0025] Figure 5 A vehicle driving scene division diagram preferably used in an embodiment of the present invention;
[0026] Figure 6 A data set partition diagram preferred in an embodiment of the present invention;
[0027] Figure 7 This is a curve chart of iterative decrease in the loss round of the validation set preferred in the embodiment of the present invention.
[0028] Specific implementation method (without drawings)
[0029] In order to make the purpose, technical scheme and advantages of the implementation of this application clearer, the technical scheme in the embodiment of this application will be described in more detail below in conjunction with the drawings in the embodiment of this application. In the drawings, the same or similar reference numerals represent the same or similar elements or elements with the same or similar functions. The described embodiments are part of the embodiments of this application, not all of them. The embodiments described below with reference to the drawings are exemplary and are intended to be used to explain this application, and cannot be understood as limiting this application. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application. In addition, it should be noted that, unless otherwise clearly specified and limited, the similar words such as "installation", "connection", "connection" used in the description of this application should be understood in a broad sense. For example, the connection can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be directly connected, it can be indirectly connected through an intermediate medium, or it can be the internal connection of two components. The technicians in the field can understand its specific meaning in this application according to the specific circumstances.
[0030] Some of the drawings in the present invention use color images, because there are many parts in the drawings of the present invention that need to be distinguished in terms of status, and the use of color images can better display different statuses, which is conducive to understanding the present invention. Figure 4 and Figure 5 The “date in scenario” shown represents scenario data, “map lanes” represents map lane information, “vehicles” represents vehicles, “his_traj” represents historical trajectory, “gt_traj” represents real trajectory, “lanes” represents lanes, “candidates” represents candidates, “CHECK” represents check, “go_straight” represents straight going, “turn_left” represents left turn, “turn_right” represents right turn, “uturn” represents U-turn, “others” represents others, and “follow” represents follow. Figure 7 The “Eval avg loss” shown means evaluation average loss.
[0031] Figure 1 The flowchart of the segmented vehicle trajectory prediction method based on training is preferred in the embodiment of the present invention. Figure 2 The present invention provides a segmented vehicle trajectory prediction method based on training, comprising:
[0032] S1) Data collection, wherein the data is vehicle driving data collected on a test road.
[0033] Specifically, the method described in this step detects and identifies the position trajectory of surrounding vehicles while driving through sensors carried by the vehicle, and combines the global coordinates and high-definition map provided by the vehicle's positioning and navigation system to obtain a local map where the target vehicle is located; the method simultaneously generates multiple sets of training data, and generates multiple segments of data with intervals but different driving states for a target vehicle; and aligns the timestamps of the vehicle driving data in the step.
[0034] S2) Data preparation: dividing the data set into a training set, a validation set and a test set. The data to be prepared include: vehicle position data based on the timestamp corresponding to different vehicle IDs, map data, vehicle dynamic environment data, driving scene data, acceleration and steering wheel angle GT values.
[0035] like Figure 6 As shown, the ratio of the training set, validation set, and test set described in this step is 8:1:1.
[0036] Preferably, the division of the driving scenarios in this step is based on road traffic rules, driving behavior characteristics and road design features, and the driving scenarios are used to train the high-dimensional representation vector of the driving habits.
[0037] S3) Model construction, the model includes a vehicle environment dynamic encoding module, a driving habit high-dimensional representation module, a historical trajectory alignment module, a prediction mapping module and a model optimization module, and the model is used for data training.
[0038] Specifically, the vehicle environment dynamic encoding module described in this step is used to encode the surrounding environment information of the target vehicle, the historical positions of other vehicles, and the surrounding lane line information into a vector form. The driving habit high-dimensional representation module is used to initialize and iteratively update the vehicle's driving habit representation vector. The historical trajectory alignment module is used to align the historical trajectories of the past period of time into a one-dimensional vector. The three groups of inputs are mapped and aligned separately in the prediction mapping module in the fully connected mapping network, and then spliced together to make the final prediction position mapping.
[0039] Preferably, the prediction mapping module described in this step will also output two intermediate quantities, acceleration and steering wheel angle, during the alignment mapping process, for use in constraining and analyzing and judging during the model training process.
[0040] Furthermore, the model in this step also includes a module for calculating the driving habit consistency loss, which evaluates the degree of matching between the driving habit representation vector and the current driving scene based on predefined driving scene rules.
[0041] S4) Model optimization: During the training process, the model optimization module is used to calculate the error between the predicted trajectory position, acceleration, steering wheel angle, and driving habit characterization vector and the true value, and the model parameters are optimized by the back propagation algorithm; the true value of the acceleration and steering wheel angle is obtained by formula calculation based on the vehicle's historical trajectory and the true trajectory. Figure 7 As shown in the figure, the loss in the validation set decreases as the number of iterations increases and finally tends to a straight line.
[0042] Furthermore, this step also includes a fully connected head for updating the high-dimensional representation vector of driving habits.
[0043] S5) Use the model to make predictions; Use the trained model to make segmented predictions on the future trajectory of the target vehicle, constrain the predicted trajectory by predicting the intermediate acceleration and steering wheel angle, and introduce a high-dimensional representation vector of driving habits to ensure the consistency of the predicted trajectory.
[0044] Figure 3 Schematic diagram of the encoding model for training high-dimensional representation of driving habits preferred in the embodiment of the present invention. Figure 3 As shown, the present invention includes a coding model for training high-dimensional representation of driving habits. The method encodes the dynamic environment, historical trajectory and actual position at the next moment into a high-dimensional representation vector of driving habits through a model of a coding and decoding structure. The position at the next moment is predicted in combination with the historical trajectory. A prediction scenario is output at the decoder end. An average feature vector output by the model after the full amount of data is used as an initialization vector in the actual reasoning in the vehicle trajectory prediction model.
[0045] The training-based segmented vehicle trajectory prediction method provided by the present invention has the following specific implementation steps: first, the data collection stage, in which different driving data under various driving scenarios are collected, and it is necessary to ensure that different numbers of vehicles are required to travel around the test vehicle during the data extraction process to ensure the dynamic changes of the surrounding environment. In an actual application scenario, a variety of sensors installed on the test vehicle, including laser radar, camera, etc., are used to detect and identify the position trajectory of surrounding vehicles during driving. At the same time, the global coordinates given by the vehicle's positioning and navigation system are combined with the high-definition map drawn in advance to accurately obtain the local map information where the target vehicle is located. Specifically, the local map information includes lane line position, turn indication, whether it is at an intersection, and traffic signal status. Data is collected in different driving scenarios during peak hours to ensure that there are different numbers of vehicles driving around to simulate a dynamically changing traffic environment. For example, near an intersection, data of multiple target vehicles in different driving states such as straight, turning or waiting for traffic lights are collected. At the same time, multiple segments of data with intervals but different driving states are generated for a target vehicle as much as possible. For example, for a specific target vehicle, we record its trajectory segments of multiple straight-moving, lane-changing, and turning movements over a period of time, and ensure that the timestamps of these data are aligned so that the local map of the environment and the target vehicle's position in each frame can be accurately encoded later.
[0046] Next is the data preparation stage. In the data preparation stage, a large amount of collected driving data needs to be sorted and classified. Because the method adopted by the present invention is to encode the local map of the environment and the position of the target vehicle in each frame, it is necessary to align the timestamps of these data in each training data preparation process to ensure the accuracy of the surrounding environment at the same time. The collected driving data is processed according to the following method to obtain training data for different target vehicles. For each data sample, 50 frames of data are collected, of which the first 20 frames are used as training data and the following 30 frames are used as predicted targets. Taking a target vehicle as an example, its position data at different time points are extracted and sorted based on the vehicle ID and the corresponding timestamp. At the same time, according to the vehicle positioning position, the lane line position data and related lane description information data within a certain range around the target vehicle are extracted, such as lane width, whether it is a solid line, etc. Driving scenes are divided according to road traffic rules, driving behavior characteristics and road design characteristics. For example, the scenes are divided into different categories such as straight driving, turning, intersection, non-intersection, overtaking, etc. According to these scene classifications, the corresponding driving scene data is annotated for each sample to train the high-dimensional representation vector of driving habits. The sorted data set is divided into training set, validation set and test set in a ratio of 8:1:1. During the division process, the data distribution of the three sets is kept consistent as much as possible, and each set contains the target vehicle's driving trajectory under different driving scenarios.
[0047] Next, it is the model building stage. The vehicle environment dynamic encoding module encodes the surrounding environment information of the target vehicle, including the historical positions of other vehicles and the surrounding lane line information. This process can be directly straightened by arranging the information, or it can be combined with rules to further encode and quantify the position quantity or use model parameters for encoding alignment. The driving habit high-dimensional representation module initially generates an average driving habit representation value. In the subsequent training process, the module will iteratively update the driving habit representation vector of a specific vehicle based on the input historical trajectory and current environmental information. The historical trajectory alignment module aligns the historical trajectories of the past five seconds into a one-dimensional vector. For example, the position changes of the target vehicle in the past five seconds are arranged in chronological order into a continuous vector to better capture the driving trend of the vehicle. It is worth noting that the historical trajectory time here is not limited to five seconds. In the prediction mapping module, three sets of inputs (vehicle environment dynamic encoding, driving habit high-dimensional representation vector, historical trajectory alignment vector) are mapped and aligned separately in the fully connected mapping network, and then spliced together to make the final predicted position mapping. In this process, two intermediate quantities, acceleration and steering wheel angle, are also output for constraints and analysis and judgment during model training. For example, when predicting the position of the target vehicle at the next moment, the corresponding acceleration and steering wheel angle are output at the same time to better constrain the predicted trajectory and make it more consistent with the actual driving situation. In addition, the model also includes a module for calculating the driving habit consistency loss, which evaluates the degree of matching between the driving habit representation vector and the current driving scene based on predefined driving scene rules. For example, if the current driving scene is a turn at an intersection, the module will determine whether the driving habit representation vector is consistent with the turning scene to ensure that the predicted trajectory conforms to the driving habits in the scene. The present invention also includes a coding model for training high-dimensional representations of driving habits. The model of the encoding and decoding structure encodes the dynamic environment, historical trajectory and the actual position at the next moment into a high-dimensional representation vector of driving habits. Combined with the historical trajectory to restore the predicted position at the next moment. Among them, a predicted scene will be output at the decoder end, and the module at this place will be used when calculating the driving habit consistency loss. The module parameters will be frozen after the model is trained and only used to calculate the loss. An average feature vector output by the model after the full amount of data will be used as an initialization vector in the actual reasoning in the aforementioned model.
[0048] Furthermore, it is the model optimization stage. During the training process, the model optimization module is used to calculate the error between the predicted trajectory position, acceleration, steering wheel angle, and driving habit characterization vector and the true value. For example, the position error is calculated by comparing the difference between the target vehicle position predicted by the model and the actual position. At the same time, the formula is calculated based on the vehicle's historical trajectory and the actual trajectory to obtain the true value of the acceleration and steering wheel angle, and compared with the model prediction value to calculate the corresponding error. The model parameters are optimized through the back propagation algorithm. For example, according to the calculated error, the back propagation adjusts the various weights and biases in the model so that the model gradually converges and improves the accuracy of the prediction. At the same time, the high-dimensional characterization vector of driving habits is updated using the fully connected header to make it better adapt to different driving scenarios and vehicle behaviors.
[0049] Finally, the model prediction stage is here. The trained model is used to predict the future trajectory of the target vehicle in segments. For example, for a target vehicle traveling on a main road, its short-term trajectory in the next few seconds is first predicted, and then the subsequent trajectory segments are gradually predicted based on the real-time updated environmental information and driving habit representation vectors. The predicted trajectory is constrained by predicting the intermediate quantities acceleration and steering wheel angle to ensure that the predicted trajectory conforms to the actual driving situation. For another example, if it is predicted that the target vehicle needs to accelerate or turn, the corresponding acceleration and steering wheel angle will constrain the predicted trajectory to make it smoother and more reasonable.
[0050] The present invention has the following beneficial effects:
[0051] 1. The present invention provides a segmented vehicle trajectory prediction method based on training. Compared with the previous unified prediction of the trajectory position at each time point, the predicted position can be better combined with the previous related position, and better trajectory smoothness can be obtained compared with the previous unified prediction.
[0052] 2. The present invention provides a segmented vehicle trajectory prediction method based on training, which introduces a high-dimensional feature as a representation of driving habits in the encoding process to measure the driving behavior of the vehicle. The driving habits of the vehicle have an inertia, and the method can better predict the trajectory of the vehicle by introducing this representation.
[0053] 3. The driving habit representation introduced by the training-based segmented vehicle trajectory prediction method of the present invention will also include pre-trained decision-making information under different driving scenarios and road conditions. Under the time characteristics of predicting multiple seconds of future trajectories, real-time updating of driving habits will have better prediction adaptability.
[0054] 4. The present invention provides a training-based segmented vehicle trajectory prediction method, and the output intermediate quantities of acceleration and turning angle can provide intuitive help in predicting failure or outlier sample analysis.
[0055] Those skilled in the art will appreciate that information, signals, and data may be represented using any of a variety of different technologies and techniques. For example, data, instructions, commands, information, signals, bits, symbols, and chips cited throughout the above description may be represented by voltage, current, electromagnetic waves, magnetic fields or magnetic particles, optical fields or optical particles, or any combination thereof.
[0056] Those skilled in the art will further appreciate that the various illustrative logic blocks, modules, circuits, and algorithm steps described in conjunction with the embodiments disclosed herein can be implemented as electronic hardware, computer software, or a combination of the two. To clearly illustrate this interchangeability of hardware and software, various illustrative components, boxes, modules, circuits, and steps are generally described above in the form of their functionality. Whether such functionality is implemented as hardware or software depends on the specific application and the design constraints imposed on the overall system. The technician can implement the described functionality in different ways for each specific application, but such implementation decisions should not be interpreted as resulting in a departure from the scope of the present application.
[0057] The various illustrative logic modules and circuits described in conjunction with the embodiments disclosed herein may be implemented or executed with a general purpose processor, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic, discrete hardware components, or any combination thereof designed to perform the functions described herein. A general purpose processor may be a microprocessor, but in the alternative, the processor may be any conventional processor, controller, microcontroller, or state machine. The processor may also be implemented as a combination of computing devices, such as a combination of a DSP and a microprocessor, a plurality of microprocessors, one or more microprocessors in cooperation with a DSP core, or any other such configuration.
[0058] The steps of the method or algorithm described in conjunction with the embodiments disclosed herein may be embodied directly in hardware, in a software module executed by a processor, or in a combination of the two. The software module may reside in a RAM memory, a flash memory, a ROM memory, an EPROM memory, an EEPROM memory, a register, a hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art. An exemplary storage medium is coupled to a processor so that the processor can read and write information from / to the storage medium. In an alternative, a storage medium may be integrated into a processor. The processor and the storage medium may reside in an ASIC. The ASIC may reside in a user terminal. In an alternative, the processor and the storage medium may reside in a user terminal as discrete components.
[0059] In one or more exemplary embodiments, the functions described may be implemented in hardware, software, firmware, or any combination thereof. If implemented as a computer program product in software, each function may be stored on or transmitted by a computer-readable medium as one or more instructions or codes. Computer-readable media include both computer storage media and communication media, including any medium that facilitates the transfer of a computer program from one place to another. Storage media may be any available medium that can be accessed by a computer. As an example and not limitation, such a computer-readable medium may include RAM, ROM, EEPROM, CD-ROM or other optical disk storage, disk storage or other magnetic storage device, or any other medium that can be used to carry or store the desired program code in the form of an instruction or data structure and can be accessed by a computer. Any connection is also properly referred to as a computer-readable medium. For example, if the software is transmitted from a website, a server, or other remote source using a coaxial cable, a fiber optic cable, a twisted pair, a digital subscriber line (DSL), or wireless technologies such as infrared, radio, and microwaves, the coaxial cable, fiber optic cable, twisted pair, DSL, or wireless technologies such as infrared, radio, and microwaves are included in the definition of the medium. Disk and disc as used herein include compact disc (CD), laser disc, optical disc, digital versatile disc (DVD), floppy disk and Blu-ray disc, wherein disk often reproduces data magnetically, while disc reproduces data optically with lasers. Combinations of the above should also be included within the scope of computer-readable media.
[0060] The above-described embodiments are only further explanations of the present invention, and are not intended to limit the present invention in other forms. The present invention may also have other various embodiments. Without departing from the spirit and essence of the present invention, those skilled in the art may make various corresponding modifications and changes according to the present invention, but these corresponding modifications and changes should all fall within the protection scope of the present invention.
Claims
1. A training-based segmented vehicle trajectory prediction method, characterized in that: include: S1) data collection, wherein the data is vehicle driving data collected on a test road; S2) Data preparation, dividing the data set into a training set, a validation set and a test set, the data to be prepared include: vehicle location data based on timestamps corresponding to different vehicle IDs, map data, vehicle dynamic environment data, driving scene data, acceleration and true value of steering wheel angle (GT value); S3) model construction, the model includes a vehicle environment dynamic encoding module, a driving habit high-dimensional representation module, a historical trajectory alignment module, a prediction mapping module and a model optimization module, and the model is used for data training; S4) Model optimization: during the training process, the model optimization module is used to calculate the error between the predicted trajectory position, acceleration, steering wheel angle, and driving habit characterization vector and the true value, and the model parameters are optimized by the back propagation algorithm; S5) Use the model to make predictions; Use the trained model to make segmented predictions on the future trajectory of the target vehicle, constrain the predicted trajectory by predicting the intermediate acceleration and steering wheel angle, and introduce a high-dimensional representation vector of driving habits to ensure the consistency of the predicted trajectory.
2. The method according to claim 1, characterized in that It also includes a coding model for training high-dimensional representation of driving habits. The method encodes the dynamic environment, historical trajectory and actual position at the next moment into a high-dimensional representation vector of driving habits through a model of a coding and decoding structure. The position at the next moment is predicted by combining the historical trajectory. A prediction scenario is output at the decoder end. An average feature vector output by the model after full data is used as an initialization vector in actual reasoning in the vehicle trajectory prediction model.
3. The method according to claim 1, characterized in that The method described in step S1) detects and identifies the position trajectory of surrounding vehicles during driving through sensors carried by the vehicle, and combines the global coordinates and high-definition map provided by the vehicle's positioning and navigation system to obtain a local map where the target vehicle is located; the method simultaneously generates multiple sets of training data, and generates multiple segments of data with intervals but different driving states for a target vehicle; and aligns the timestamps of the vehicle driving data in step S1).
4. The method according to claim 1, characterized in that: The ratio of the training set, validation set and test set described in step S2) is 8:1:
1.
5. The method according to claim 1, characterized in that: Step S2) The division of these driving scenarios is based on road traffic rules, driving behavior characteristics and road design features, and the driving scenarios are used to train the high-dimensional representation vector of the driving habits.
6. The method according to claim 1, characterized in that The vehicle environment dynamic encoding module described in step S3) is used to encode the surrounding environment information of the target vehicle, the historical positions of other vehicles, and the surrounding lane line information into a vector form. The driving habit high-dimensional representation module is used to initialize and iteratively update the driving habit representation vector of a specific vehicle. The historical trajectory alignment module is used to align the historical trajectories of the past period of time into a one-dimensional vector. The three groups of inputs are mapped and aligned separately in the prediction mapping module in the fully connected mapping network, and then spliced together to make the final prediction position mapping.
7. The method according to claim 1, characterized in that The prediction mapping module described in step S3) will also output two intermediate quantities, acceleration and steering wheel angle, during the alignment mapping process, which are used for constraints and analysis and judgment during the model training process.
8. The method according to claim 1, characterized in that Step S3) The model also includes a module for calculating the driving habit consistency loss, which evaluates the matching degree between the driving habit representation vector and the current driving scene based on predefined driving scene rules.
9. The method according to claim 1, characterized in that: Step S4) The GT values of the acceleration and steering wheel angle are obtained by formula calculation through the vehicle historical trajectory and the GT trajectory.
10. The method according to claim 1, characterized in that Step S4) also includes a fully connected head for updating the high-dimensional representation vector of driving habits.