CN-121973808-A - Method, device, electronic equipment and medium for predicting track of automatic driving vehicle

CN121973808ACN 121973808 ACN121973808 ACN 121973808ACN-121973808-A

Abstract

The invention relates to a method, a device, electronic equipment and a medium for predicting the track of an automatic driving vehicle, belonging to the technical field of intelligent driving, wherein the method comprises the steps of obtaining historical track data of a plurality of drivers in a target scene and preference track data at the current moment; the method comprises the steps of inputting historical track data into a complete LSTM model to obtain a candidate track set, cleaning the candidate track set to obtain non-preference track data, combining the historical track data, the preference track data and the non-preference track data to obtain a driving preference data set, using the LSTM model as a strategy model, optimizing the strategy model by adopting a direct preference optimization algorithm based on the driving preference data set to obtain an optimized strategy model, inputting data to be predicted of a target vehicle into the optimized strategy model, and predicting the track of the target vehicle at the next moment. The prediction track has better smoothness of the driving path and better driving comfort, and accords with the preference of human driving.

Inventors

CHEN ZHIJUN
HAN YUNXIANG
XIONG SHENGGUANG
FENG LEI
WU KAI

Assignees

武汉理工大学

Dates

Publication Date: 20260505
Application Date: 20251230

Claims (10)

1. A method of autonomous vehicle trajectory prediction, comprising: acquiring historical track data of a plurality of drivers in a target scene and preference track data at the current moment; inputting the historical track data into a complete LSTM model to obtain a candidate track set; Obtaining displacement errors based on the candidate track sets and real track data corresponding to the candidate track sets; Cleaning the candidate track set based on the displacement error and a preset driving habit scoring function to obtain non-preference track data; combining the historical track data, the preference track data and the non-preference track data to obtain a driving preference data set; taking the LSTM model as a strategy model, optimizing the strategy model by adopting a direct preference optimization algorithm based on a driving preference data set, and obtaining an optimized strategy model; And inputting the track data of the target vehicle at the moment in the target scene and the track data of the interactive vehicles around the target vehicle at the moment in the last moment into the optimized strategy model, and predicting the track of the target vehicle at the next moment.
2. The method of autonomous vehicle trajectory prediction of claim 1, wherein the expression of the hidden state vector of the LSTM model is: In the formula, Represents the hidden state of the vehicle j adjacent to the target vehicle i in the target scene, The depth of interaction coefficient of the target vehicle j with the vehicle i is represented, Representing the hidden state of the vehicle i in the target scene, , Representing the weights of the LSTM model.
3. The method for predicting the trajectory of an autonomous vehicle according to claim 1, wherein the cleaning the candidate trajectory set based on the displacement error and a preset driving habit scoring function to obtain non-preference trajectory data comprises: screening out tracks with the critical value of the track movement direction in the candidate track set not smaller than the corresponding displacement error to obtain a transition track set; and screening out tracks with the values of the driving habit scoring functions of the transition track concentrated tracks larger than the preset values to obtain non-preference track data.
4. A method of automatically driving a vehicle trajectory prediction as claimed in claim 3, wherein the expression of the preset driving habit scoring function is: wherein: A score representing the m candidate trajectories, The degree of impact is indicated by the degree of impact, Indicating the lateral acceleration of the vehicle, The total travel distance of the track is indicated, A weight coefficient representing the degree of impact, A weight coefficient representing the lateral acceleration, A weight coefficient representing the total travel distance of the track.
5. The method of autonomous vehicle trajectory prediction according to claim 1, wherein the optimizing the strategy model based on the driving preference dataset using a direct preference optimization algorithm results in an optimized strategy model comprising: carrying out weighted average on probability density of the mixed Gaussian distribution of the preset number of each track of the non-preference track data to obtain a weighted average; obtaining a first conditional probability of the preference trajectory data and a second conditional probability of the non-preference trajectory data based on a weighted average by a chained method; obtaining response difference values of the strategy model and the reference strategy model for the preference track data and the non-preference track data based on the first conditional probability and the second conditional probability; Determining and obtaining the strategy model to respond to human preference based on the response difference value; obtaining a loss function of the strategy model based on the probability of the strategy model responding to human preference; And determining an optimized strategy model based on the loss function of the strategy model.
6. The method of autonomous vehicle trajectory prediction of claim 1, wherein the probability density for each trajectory of the non-preferential trajectory data is expressed as: In the formula, 、 For the coordinate values of the trace point, The characteristic value of the two-dimensional Gaussian distribution at the current t moment is represented, The characteristic value of the two-dimensional Gaussian distribution at the current t moment is represented, The correlation coefficient representing the gaussian distribution, The number n of vehicles is represented as an n-th vehicle, The probability density is represented by a value representing, Representing the kth gaussian distribution.
7. The method of autonomous vehicle trajectory prediction of claim 1, wherein the expression of the loss function of the optimized strategy model is: In the formula, The loss function is represented by a function of the loss, Indicating that the control deviates from the reference strategy, Representing the LSTM model to be trained, The super-parameter is represented by a parameter, The history of the track is represented and, A preference trajectory of the vehicle n is indicated, Representing a non-preferential trajectory of the vehicle n, , The coordinate axis of the vehicle n at time t is shown, 、 The coordinate values representing the locus point, Representation superparameters for characterizing control deviation reference strategies To a degree of (3), Representing a Sigmoid function.
8. An apparatus for autonomous vehicle trajectory prediction, comprising: The track data acquisition module is used for acquiring historical track data of a plurality of drivers in the target scene and preference track data at the current moment; the candidate track set acquisition module is used for inputting the historical track data into the LSTM model with complete training to obtain a candidate track set; The displacement error determining module is used for obtaining displacement errors based on the candidate track set and real track data corresponding to the candidate track set; The non-preference track data determining module is used for cleaning the candidate track set based on the displacement error and a preset driving habit scoring function to obtain non-preference track data; the driving preference data set determining module is used for combining the historical track data, the preference track data and the non-preference track data to obtain a driving preference data set; The optimized strategy model determining module is used for taking the LSTM model as a strategy model, optimizing the strategy model by adopting a direct preference optimization algorithm based on the driving preference data set, and obtaining an optimized strategy model; The vehicle track prediction module is used for inputting track data of a target vehicle at a moment in a target scene and track data of interactive vehicles around the target vehicle at a moment in the last moment into the optimized strategy model, and predicting the track of the target vehicle at a next moment.
9. An electronic device comprising a memory and a processor, wherein, The memory is used for storing programs; The processor, coupled to the memory, for executing the program stored in the memory to implement the steps in a method of autonomous vehicle trajectory prediction as claimed in any one of the preceding claims 1 to 7.
10. A computer readable storage medium storing a computer readable program or instructions which when executed by a processor is capable of carrying out the steps of a method of autonomous vehicle trajectory prediction as claimed in any one of the preceding claims 1 to 7.

Description

Method, device, electronic equipment and medium for predicting track of automatic driving vehicle Technical Field The invention relates to the technical field of intelligent driving, in particular to a method, a device, electronic equipment and a medium for predicting an automatic driving vehicle track. Background With the development of intelligent traffic systems and automatic driving technologies, the road traffic environment is increasingly complex, and the action diversity of traffic participants, the complex interaction between the traffic participants and the environment and the uncertainty of perceived information are all key to realizing the safety of automatic driving decision. The human driver drives the vehicle and can usually rapidly recognize scene information, observe and capture the relationship between the vehicle and surrounding traffic participants, judge the current vehicle driving state and accurately predict the future state of the surrounding traffic participants, discover potential hazards in advance and react. Although the automatic driving technology realizes intelligent decision and control of the vehicle, the decision model cannot accurately capture the interaction relation between vehicles when predicting the state of the vehicle, and the matching degree is poor when predicting different driving preferences of surrounding vehicle movements, the decision result tends to deviate to deterministic output, and the prediction effect and the recognition precision are deteriorated to different degrees in partial scenes. Therefore, the automatic driving vehicle needs to have the judging and predicting driving ability similar to that of a human driver, can predict the future states of traffic participants and environments nearby the vehicle, and provides scientific and reliable basis for vehicle decision. The intelligent network-connected vehicle can acquire a large amount of sensing data of multiple sensors through sensing equipment of vehicle-mounted and road side sections, provides support for acquiring space-time motion states of surrounding vehicles and road environment conditions, and is also a premise for realizing vehicle track prediction. And processing the acquired driving environment information, and predicting the vehicle group track by the automatic driving vehicle. The work of predecessor based on track prediction is mainly developed from several methods of physics, classical machine learning, deep learning and reinforcement learning, and can realize the prediction of vehicle driving behavior intention, single track prediction and multi-mode track prediction. The current track prediction model is often used for extracting motion track characteristics based on the historical track of the target vehicle, so that the future path of the target vehicle can be predicted more accurately. However, the existing method mainly fits the historical motion trail of the vehicle from a time frame, fails to wholly analyze the motion rule of the vehicle, and cannot accurately capture the interaction relationship among traffic participants. Second, although the prediction model recognizes and predicts the driving intention, the driving preference is only predicted macroscopically, the key element of the specificity of the driver is ignored, the movement difference between different vehicles cannot be considered, and the human driving preference cannot be fully understood by the vehicle track. This prediction inevitably weakens the universality of the model and the reliability of the track prediction result, and also leads to the reduction of prediction accuracy. In summary, the prior art lacks a method to incorporate the interaction relationship between traffic participants and the driving selection preference of the human driver into the consideration range of the trajectory prediction, so as to fit a finer driving behavior pattern, thereby realizing the accurate prediction of the future driving trajectory of the vehicle and providing more reliable and accurate trajectory prediction data support for automatic driving. Disclosure of Invention In view of the foregoing, there is a need for an automatic driving vehicle trajectory prediction method, apparatus, electronic device, and medium, which are used for solving the problem of low trajectory prediction accuracy caused by the fact that the interaction relationship between traffic participants and the driving selection preference of human drivers are not included in the consideration range of trajectory prediction in the prior art. To solve the above problems, in a first aspect, the present invention provides a method for predicting an automatically driven vehicle trajectory, including: acquiring historical track data of a plurality of drivers in a target scene and preference track data at the current moment; inputting the historical track data into a complete LSTM model to obtain a candidate track set, and adding the influence coefficient of s